From 5f2bca4ab8385e8703149e12164951248725ce77 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Tue, 18 Jun 2024 09:21:00 +0800
Subject: [PATCH 01/29] [Refactor] fix internlm2 dispatch (#779)

* fix internlm2 dispatch

* add detailed RuntimeError
---
 xtuner/model/modules/dispatch/__init__.py  |   2 -
 xtuner/model/modules/dispatch/internlm2.py | 277 ++++++++++-----------
 2 files changed, 134 insertions(+), 145 deletions(-)

diff --git a/xtuner/model/modules/dispatch/__init__.py b/xtuner/model/modules/dispatch/__init__.py
index c2033c3c8..7cb159515 100644
--- a/xtuner/model/modules/dispatch/__init__.py
+++ b/xtuner/model/modules/dispatch/__init__.py
@@ -125,8 +125,6 @@
 )
 
 ROTE_DISPATCH_MAPPING = dict(
-    InternLM2RotaryEmbedding=LazyObject(
-        'xtuner.model.modules.dispatch.internlm2', 'InternLM2RotaryEmbedding'),
     InternLMRotaryEmbedding=LazyObject(
         'xtuner.model.modules.dispatch.internlm', 'InternLMRotaryEmbedding'),
     MistralRotaryEmbedding=LazyObject('xtuner.model.modules.dispatch.mistral',
diff --git a/xtuner/model/modules/dispatch/internlm2.py b/xtuner/model/modules/dispatch/internlm2.py
index 6fbb2d0d6..5b855d4ab 100644
--- a/xtuner/model/modules/dispatch/internlm2.py
+++ b/xtuner/model/modules/dispatch/internlm2.py
@@ -1,62 +1,16 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-import warnings
 from typing import Optional, Tuple
 
 import torch
 import torch.distributed as dist
-import torch.nn.functional as F
 from einops import rearrange
 from mmengine import MessageHub
+from transformers.cache_utils import Cache, StaticCache
 
-from .attention import (SUPPORT_FLASH2, flash_attn_w_mask, flash_attn_wo_mask,
-                        varlen_flash_attn)
-from .triton_kernels import apply_rotary_emb
-
-
-class InternLM2RotaryEmbedding(torch.nn.Module):
-
-    def __init__(self,
-                 dim,
-                 max_position_embeddings=2048,
-                 base=1000000,
-                 device=None):
-        super().__init__()
-        self.dim = dim
-        self.max_position_embeddings = max_position_embeddings
-        self.base = base
-        self.inv_freq = 1.0 / (
-            base**(torch.arange(0, dim, 2).float().to(device) / dim))
-
-        # Build here to make `torch.jit.trace` work.
-        self.max_seq_len_cached = max_position_embeddings
-        t = torch.arange(
-            self.max_seq_len_cached,
-            device=self.inv_freq.device,
-            dtype=self.inv_freq.dtype)
-        freqs = torch.einsum('i,j->ij', t, self.inv_freq)
-        emb = torch.cat((freqs, freqs), dim=-1)
-        self.cos_cached = emb.cos()
-        self.sin_cached = emb.sin()
-
-    def forward(self, x, seq_len):
-        # x: [bs, num_attention_heads, seq_len, head_size]
-        if (seq_len > self.max_seq_len_cached
-                or self.cos_cached.device != x.device
-                or self.cos_cached.dtype != x.dtype):
-            self.max_seq_len_cached = seq_len
-            assert self.inv_freq.dtype == torch.float32
-            t = torch.arange(
-                self.max_seq_len_cached,
-                device=x.device,
-                dtype=self.inv_freq.dtype)
-            freqs = torch.einsum('i,j->ij', t, self.inv_freq.to(t.device))
-            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
-            self.cos_cached = emb.cos().to(x.dtype)
-            self.sin_cached = emb.sin().to(x.dtype)
-        return (
-            self.cos_cached[:seq_len, ...],
-            self.sin_cached[:seq_len, ...],
-        )
+from xtuner.parallel.sequence import (get_sequence_parallel_world_size,
+                                      post_process_for_sequence_parallel_attn,
+                                      pre_process_for_sequence_parallel_attn)
+from .attention import SUPPORT_FLASH2, flash_attn_wo_mask, varlen_flash_attn
 
 
 def rotate_half(x):
@@ -66,9 +20,9 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
-    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
-    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
     q_embed = (q * cos) + (rotate_half(q) * sin)
     k_embed = (k * cos) + (rotate_half(k) * sin)
     return q_embed, k_embed
@@ -111,18 +65,17 @@ def internlm2_attn_forward(
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.LongTensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    past_key_value: Optional[Cache] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
-    **kwargs,
+    cache_position: Optional[torch.LongTensor] = None,
 ):
-    if 'padding_mask' in kwargs:
-        warnings.warn(
-            'Passing `padding_mask` is deprecated and will be removed in v4.37'
-            'Please make sure use `attention_mask` instead.`')
-
-        # overwrite attention_mask with padding_mask
-        attention_mask = kwargs.pop('padding_mask')
+    if isinstance(past_key_value, StaticCache):
+        raise ValueError(
+            '`static` cache implementation is not compatible with '
+            '`attn_implementation==flash_attention_2` make sure to use `sdpa` '
+            'in the mean time, and open an issue at '
+            'https://github.com/huggingface/transformers')
 
     output_attentions = False
 
@@ -146,64 +99,68 @@ def internlm2_attn_forward(
     key_states = key_states.transpose(1, 2)
     value_states = value_states.transpose(1, 2)
 
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-
-    # This modification is necessary for sequential parallel
-    assert position_ids is not None and (position_ids.max() + 1) >= kv_seq_len
-    cos, sin = self.rotary_emb(value_states, seq_len=position_ids.max() + 1)
+    cos, sin = self.rotary_emb(value_states, position_ids)
     query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
-                                                    cos, sin, position_ids)
+                                                    cos, sin)
 
     if past_key_value is not None:
-        # reuse k, v, self_attention
-        key_states = torch.cat([past_key_value[0], key_states], dim=2)
-        value_states = torch.cat([past_key_value[1], value_states], dim=2)
-
-    past_key_value = (key_states, value_states) if use_cache else None
+        # sin and cos are specific to RoPE models;
+        # cache_position needed for the static cache
+        cache_kwargs = {
+            'sin': sin,
+            'cos': cos,
+            'cache_position': cache_position
+        }
+        key_states, value_states = past_key_value.update(
+            key_states, value_states, self.layer_idx, cache_kwargs)
 
-    # repeat kv for sequence parallel
     key_states = repeat_kv(key_states, self.num_key_value_groups)
     value_states = repeat_kv(value_states, self.num_key_value_groups)
 
-    if SUPPORT_FLASH2:
-        # the shape of attention_mask used by flash_attn and
-        # F.scaled_dot_product_attention are different
-        assert attention_mask is None or attention_mask.ndim == 2, \
-            ('When using flash_attn, attention_mask.ndim should equal to 2.'
-             f'But got attention_mask.shape = {attention_mask.shape}.'
-             'We can pass the `attn_implementation="flash_attention_2"` flag '
-             'to `.from_pretrained` method when instantiating a Internlm2 '
-             'model.')
-        # flash attn 2 need (bs, seq_len, nhead, h_dim)
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-
-        causal = self.is_causal and q_len != 1
-
-        if attention_mask is not None:
-            attn_output = flash_attn_w_mask(
-                query_states,
-                key_states,
-                value_states,
-                attention_mask,
-                causal=causal,
-                training=self.training)
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    # In PEFT, usually we cast the layer norms in float32 for training
+    # stability reasons therefore the input hidden states gets silently
+    # casted in float32. Hence, we need cast them back in the correct dtype
+    # just to be sure everything works as expected.
+    # This might slowdown training & inference so it is recommended to not
+    # cast the LayerNorms in fp32. (InternLM2RMSNorm handles it correctly)
+
+    input_dtype = query_states.dtype
+    if input_dtype == torch.float32:
+        if torch.is_autocast_enabled():
+            target_dtype = torch.get_autocast_gpu_dtype()
+        # Handle the case where the model is quantized
+        elif hasattr(self.config, '_pre_quantization_dtype'):
+            target_dtype = self.config._pre_quantization_dtype
         else:
-            attn_output = flash_attn_wo_mask(
-                query_states,
-                key_states,
-                value_states,
-                causal=causal,
-                training=self.training)
-    else:
-        # use flash attention implemented by pytorch
-        # do not support sequence parallel
-        attn_output = F.scaled_dot_product_attention(
-            query_states, key_states, value_states, attn_mask=attention_mask)
-        attn_output = attn_output.transpose(1, 2)
+            target_dtype = self.wqkv.weight.dtype
+
+        query_states = query_states.to(target_dtype)
+        key_states = key_states.to(target_dtype)
+        value_states = value_states.to(target_dtype)
+
+    enable_sequence_parallel = (
+        dist.is_initialized() and get_sequence_parallel_world_size() > 1
+        and self.training)
+    if enable_sequence_parallel:
+        query_states, key_states, value_states = \
+            pre_process_for_sequence_parallel_attn(
+                query_states, key_states, value_states)
+
+    dropout_rate = 0.0
+    attn_output = self._flash_attention_forward(
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_states.shape[1],
+        dropout=dropout_rate)
+
+    if enable_sequence_parallel:
+        attn_output = post_process_for_sequence_parallel_attn(attn_output)
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
     attn_output = self.wo(attn_output)
@@ -217,14 +174,21 @@ def internlm2_attn_forward(
 def internlm2_varlen_attn_forward(
     self,
     hidden_states: torch.Tensor,
-    attention_mask: Optional[torch.Tensor] = None,
+    attention_mask: Optional[torch.LongTensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    past_key_value: Optional[Cache] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    cache_position: Optional[torch.LongTensor] = None,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
            Optional[Tuple[torch.Tensor]]]:
-    # Modified from https://huggingface.co/internlm/internlm-7b/blob/939a68c0dc1bd5f35b63c87d44af05ce33379061/modeling_internlm.py#L161  # noqa:E501
+
+    if isinstance(past_key_value, StaticCache):
+        raise ValueError(
+            '`static` cache implementation is not compatible with '
+            '`attn_implementation==flash_attention_2` make sure to use `sdpa` '
+            'in the mean time, and open an issue at '
+            'https://github.com/huggingface/transformers')
 
     message_hub = MessageHub.get_instance('varlen_attn_args')
     rank = dist.get_rank()
@@ -238,6 +202,7 @@ def internlm2_varlen_attn_forward(
                       f' set to 1, but got {bsz}')
 
     qkv_states = self.wqkv(hidden_states)
+
     qkv_states = rearrange(
         qkv_states,
         'b q (h gs d) -> b q h gs d',
@@ -250,40 +215,63 @@ def internlm2_varlen_attn_forward(
     key_states = qkv_states[..., -2, :]
     value_states = qkv_states[..., -1, :]
 
-    kv_seq_len = key_states.shape[-3]
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    try:
+        cos, sin = self.rotary_emb(value_states, position_ids)
+    except RuntimeError:
+        raise RuntimeError(
+            'You are using the old version of InternLM2 model. The '
+            '`modeling_internlm2.py` is outdated. Please update the InternLM2 '
+            'model.')
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
+                                                    cos, sin)
+
     if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
+        # sin and cos are specific to RoPE models;
+        # cache_position needed for the static cache
+        cache_kwargs = {
+            'sin': sin,
+            'cos': cos,
+            'cache_position': cache_position
+        }
+        key_states, value_states = past_key_value.update(
+            key_states, value_states, self.layer_idx, cache_kwargs)
 
-    if use_varlen_atten:
-        cos, sin = self.rotary_emb(value_states, max_seqlen)
-        query_states = apply_rotary_emb(query_states,
-                                        cos[position_ids].squeeze(0),
-                                        sin[position_ids].squeeze(0))
-        key_states = apply_rotary_emb(key_states, cos[position_ids].squeeze(0),
-                                      sin[position_ids].squeeze(0))
-    else:
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-        cos, sin = self.rotary_emb(value_states, kv_seq_len)
-        query_states, key_states = apply_rotary_pos_emb(
-            query_states, key_states, cos, sin, position_ids)
-
-        if past_key_value is not None:
-            # reuse k, v, self_attention
-            key_states = torch.cat([past_key_value[0], key_states], dim=2)
-            value_states = torch.cat([past_key_value[1], value_states], dim=2)
-
-        past_key_value = (key_states, value_states) if use_cache else None
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    # In PEFT, usually we cast the layer norms in float32 for training
+    # stability reasons therefore the input hidden states gets silently
+    # casted in float32. Hence, we need cast them back in the correct dtype
+    # just to be sure everything works as expected.
+    # This might slowdown training & inference so it is recommended to not
+    # cast the LayerNorms in fp32. (InternLM2RMSNorm handles it correctly)
+
+    input_dtype = query_states.dtype
+    if input_dtype == torch.float32:
+        if torch.is_autocast_enabled():
+            target_dtype = torch.get_autocast_gpu_dtype()
+        # Handle the case where the model is quantized
+        elif hasattr(self.config, '_pre_quantization_dtype'):
+            target_dtype = self.config._pre_quantization_dtype
+        else:
+            target_dtype = self.wqkv.weight.dtype
+
+        query_states = query_states.to(target_dtype)
+        key_states = key_states.to(target_dtype)
+        value_states = value_states.to(target_dtype)
 
     # repeat kv for sequence parallel
     key_states = repeat_kv_bshd(key_states, self.num_key_value_groups)
     value_states = repeat_kv_bshd(value_states, self.num_key_value_groups)
 
     assert SUPPORT_FLASH2
+
+    dropout_rate = 0.0
     if use_varlen_atten:
         attn_output = varlen_flash_attn(
             query_states,
@@ -291,6 +279,8 @@ def internlm2_varlen_attn_forward(
             value_states,
             cumulative_len,
             max_seqlen,
+            causal=True,
+            dropout_p=dropout_rate,
             training=self.training)
     else:
         attn_output = flash_attn_wo_mask(
@@ -298,7 +288,8 @@ def internlm2_varlen_attn_forward(
             key_states,
             value_states,
             causal=True,
-            training=False)
+            dropout_p=dropout_rate,
+            training=self.training)
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
 

From 72b645f328b889c92fbf5f00a1714ef8d685ffd0 Mon Sep 17 00:00:00 2001
From: Wang Xinjiang <swanxinjiang@gmail.com>
Date: Thu, 20 Jun 2024 17:26:21 +0800
Subject: [PATCH 02/29] Fix zero3 compatibility issue for DPO (#781)

* fix zero3

* reformat

* reformat

* reformat

* reformat
---
 xtuner/model/dpo.py | 38 ++++++++++++++++++++------------------
 1 file changed, 20 insertions(+), 18 deletions(-)

diff --git a/xtuner/model/dpo.py b/xtuner/model/dpo.py
index 782f4d20f..b46ea1c50 100644
--- a/xtuner/model/dpo.py
+++ b/xtuner/model/dpo.py
@@ -16,23 +16,6 @@
 from .sft import SupervisedFinetune
 
 
-def create_reference_model(model):
-    if is_deepspeed_zero3_enabled():
-        raise ValueError('DeepSpeed ZeRO-3 is enabled and is not compatible '
-                         'with `create_reference_model()`. Please instantiate '
-                         'your reference model directly with '
-                         '`AutoCausalLM.from_pretrained()`.')
-
-    parameter_names = [n for n, _ in model.named_parameters()]
-    ref_model = deepcopy(model)
-
-    # if no layers are shared, return copy of model
-    for param_name in parameter_names:
-        param = ref_model.get_parameter(param_name)
-        param.requires_grad = False
-    return ref_model.eval()
-
-
 class DPO(SupervisedFinetune):
     """A general class of DPO and its variants."""
 
@@ -50,7 +33,26 @@ def __init__(self,
         self.beta = beta
 
         if not self.use_lora:
-            self.ref_llm = create_reference_model(self.llm)
+            self.ref_llm = self.create_reference_model(ref_llm, **kwargs)
+
+    def create_reference_model(self, ref_llm=None, **kwargs):
+        ref_model = None
+        if ref_llm is None:
+            if is_deepspeed_zero3_enabled():
+                raise ValueError(
+                    'DeepSpeed ZeRO-3 is enabled and is not compatible '
+                    'with `deepcopy(self.llm)`. Please instantiate '
+                    'your reference model by modifying key `model.ref_llm` '
+                    'in your config with `AutoCausalLM.from_pretrained()`.')
+            ref_model = deepcopy(self.llm)
+        else:
+            ref_model = SupervisedFinetune(ref_llm, **kwargs).llm
+        # freeze parameters
+        parameter_names = [n for n, _ in ref_model.named_parameters()]
+        for param_name in parameter_names:
+            param = ref_model.get_parameter(param_name)
+            param.requires_grad = False
+        return ref_model.eval()
 
     def _gather_masked_logits(self, logits, labels, mask):
         logits = torch.gather(

From 937e9937f5f60c21bc7bc4d510936dfc2ba671eb Mon Sep 17 00:00:00 2001
From: fanqiNO1 <75657629+fanqiNO1@users.noreply.github.com>
Date: Mon, 24 Jun 2024 11:17:15 +0800
Subject: [PATCH 03/29] [Fix] Fix map_fn in custom_dataset/sft (#785)

---
 .../sft/baichuan/baichuan2_13b_chat_qlora_custom_sft_e1.py    | 4 ++--
 .../sft/baichuan/baichuan2_7b_chat_qlora_custom_sft_e1.py     | 4 ++--
 .../sft/baichuan/baichuan_13b_chat_qlora_custom_sft_e1.py     | 4 ++--
 .../sft/baichuan/baichuan_7b_qlora_custom_sft_e1.py           | 4 ++--
 .../sft/chatglm/chatglm2_6b_qlora_custom_sft_e1.py            | 4 ++--
 .../sft/chatglm/chatglm3_6b_qlora_custom_sft_e1.py            | 4 ++--
 .../sft/deepseek/deepseek_moe_16b_chat_qlora_custom_sft_e1.py | 4 ++--
 .../deepseekcoder_6_7b_instruct_qlora_custom_sft_e1.py        | 4 ++--
 .../sft/gemma/gemma_2b_it_qlora_custom_sft_e1.py              | 4 ++--
 .../custom_dataset/sft/gemma/gemma_2b_qlora_custom_sft_e1.py  | 4 ++--
 .../sft/gemma/gemma_7b_it_qlora_custom_sft_e1.py              | 4 ++--
 .../custom_dataset/sft/gemma/gemma_7b_qlora_custom_sft_e1.py  | 4 ++--
 .../sft/internlm/internlm2_chat_1_8b_qlora_custom_sft_e1.py   | 4 ++--
 .../sft/internlm/internlm2_chat_20b_qlora_custom_sft_e1.py    | 4 ++--
 .../sft/internlm/internlm2_chat_7b_qlora_custom_sft_e1.py     | 4 ++--
 .../sft/llama/llama2_70b_qlora_custom_sft_e1.py               | 4 ++--
 .../sft/llama/llama2_7b_chat_qlora_custom_sft_e1.py           | 4 ++--
 .../sft/mistral/mistral_7b_full_finetune_custom_sft_e1.py     | 4 ++--
 .../sft/mixtral/mixtral_8x7b_instruct_qlora_custom_sft_e1.py  | 4 ++--
 .../sft/qwen/qwen1_5_0_5b_chat_qlora_custom_sft_e1.py         | 4 ++--
 .../sft/qwen/qwen1_5_14b_chat_qlora_custom_sft_e1.py          | 4 ++--
 .../sft/qwen/qwen1_5_1_8b_chat_qlora_custom_sft_e1.py         | 4 ++--
 .../sft/qwen/qwen1_5_4b_chat_qlora_custom_sft_e1.py           | 4 ++--
 .../sft/qwen/qwen1_5_72b_chat_qlora_custom_sft_e1.py          | 4 ++--
 .../sft/qwen/qwen1_5_7b_chat_qlora_custom_sft_e1.py           | 4 ++--
 .../sft/qwen/qwen_1_8b_chat_qlora_custom_sft_e1.py            | 4 ++--
 .../custom_dataset/sft/qwen/qwen_72b_qlora_custom_sft_e1.py   | 4 ++--
 .../sft/qwen/qwen_7b_chat_qlora_custom_sft_e1.py              | 4 ++--
 .../sft/starcoder/starcoder_qlora_custom_sft_e1.py            | 4 ++--
 .../custom_dataset/sft/yi/yi_34b_qlora_custom_sft_e1.py       | 4 ++--
 .../custom_dataset/sft/yi/yi_6b_qlora_custom_sft_e1.py        | 4 ++--
 .../sft/zephyr/zephyr_7b_beta_qlora_custom_sft_e1.py          | 4 ++--
 32 files changed, 64 insertions(+), 64 deletions(-)

diff --git a/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_13b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_13b_chat_qlora_custom_sft_e1.py
index c1f79073d..558887c04 100644
--- a/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_13b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_13b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_7b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_7b_chat_qlora_custom_sft_e1.py
index a2f4d445c..8df388a67 100644
--- a/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_7b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/baichuan/baichuan2_7b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/baichuan/baichuan_13b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/baichuan/baichuan_13b_chat_qlora_custom_sft_e1.py
index 178cee847..3dc38eb4f 100644
--- a/xtuner/configs/custom_dataset/sft/baichuan/baichuan_13b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/baichuan/baichuan_13b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/baichuan/baichuan_7b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/baichuan/baichuan_7b_qlora_custom_sft_e1.py
index eecce4cce..dc15b6289 100644
--- a/xtuner/configs/custom_dataset/sft/baichuan/baichuan_7b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/baichuan/baichuan_7b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/chatglm/chatglm2_6b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/chatglm/chatglm2_6b_qlora_custom_sft_e1.py
index 312044a88..09b354929 100644
--- a/xtuner/configs/custom_dataset/sft/chatglm/chatglm2_6b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/chatglm/chatglm2_6b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/chatglm/chatglm3_6b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/chatglm/chatglm3_6b_qlora_custom_sft_e1.py
index c0bd6b2a7..7e3abba71 100644
--- a/xtuner/configs/custom_dataset/sft/chatglm/chatglm3_6b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/chatglm/chatglm3_6b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -112,7 +112,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/deepseek/deepseek_moe_16b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/deepseek/deepseek_moe_16b_chat_qlora_custom_sft_e1.py
index d4cf2d9d9..f7621bc6c 100644
--- a/xtuner/configs/custom_dataset/sft/deepseek/deepseek_moe_16b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/deepseek/deepseek_moe_16b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/deepseek/deepseekcoder_6_7b_instruct_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/deepseek/deepseekcoder_6_7b_instruct_qlora_custom_sft_e1.py
index b0142dc76..629012f5b 100644
--- a/xtuner/configs/custom_dataset/sft/deepseek/deepseekcoder_6_7b_instruct_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/deepseek/deepseekcoder_6_7b_instruct_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -115,7 +115,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_it_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_it_qlora_custom_sft_e1.py
index cc3d90a8b..122ddf023 100644
--- a/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_it_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_it_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_qlora_custom_sft_e1.py
index a993ecbab..9a3d36b30 100644
--- a/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/gemma/gemma_2b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_it_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_it_qlora_custom_sft_e1.py
index 5d5dab1c0..c677c9d09 100644
--- a/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_it_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_it_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_qlora_custom_sft_e1.py
index bc906731d..443a1e663 100644
--- a/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/gemma/gemma_7b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_1_8b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_1_8b_qlora_custom_sft_e1.py
index af01dc2c7..2aaa6f24d 100644
--- a/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_1_8b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_1_8b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_20b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_20b_qlora_custom_sft_e1.py
index f0c74ce81..dfb423839 100644
--- a/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_20b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_20b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_7b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_7b_qlora_custom_sft_e1.py
index 645f9cbf6..313103992 100644
--- a/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_7b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/internlm/internlm2_chat_7b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/llama/llama2_70b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/llama/llama2_70b_qlora_custom_sft_e1.py
index 53b380719..2b0f889b4 100644
--- a/xtuner/configs/custom_dataset/sft/llama/llama2_70b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/llama/llama2_70b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -112,7 +112,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/llama/llama2_7b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/llama/llama2_7b_chat_qlora_custom_sft_e1.py
index 59d14cfb5..9aa9b6362 100644
--- a/xtuner/configs/custom_dataset/sft/llama/llama2_7b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/llama/llama2_7b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/mistral/mistral_7b_full_finetune_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/mistral/mistral_7b_full_finetune_custom_sft_e1.py
index f6aa0f574..0af78f79f 100644
--- a/xtuner/configs/custom_dataset/sft/mistral/mistral_7b_full_finetune_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/mistral/mistral_7b_full_finetune_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.dataset.samplers import InternRepoSampler
 from xtuner.engine import (DatasetInfoHook, EvaluateChatHook, ThroughputHook,
                            VarlenAttnArgsToMessageHubHook)
@@ -114,7 +114,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/mixtral/mixtral_8x7b_instruct_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/mixtral/mixtral_8x7b_instruct_qlora_custom_sft_e1.py
index 3eea2a9f5..91cda57ec 100644
--- a/xtuner/configs/custom_dataset/sft/mixtral/mixtral_8x7b_instruct_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/mixtral/mixtral_8x7b_instruct_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -114,7 +114,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_0_5b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_0_5b_chat_qlora_custom_sft_e1.py
index 8f26fe5ad..3066f0be9 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_0_5b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_0_5b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_14b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_14b_chat_qlora_custom_sft_e1.py
index 70aa8d7f6..642592f0c 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_14b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_14b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_1_8b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_1_8b_chat_qlora_custom_sft_e1.py
index d7196c53a..3790006d7 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_1_8b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_1_8b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_4b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_4b_chat_qlora_custom_sft_e1.py
index 0c2808485..36d3e6cd0 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_4b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_4b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_72b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_72b_chat_qlora_custom_sft_e1.py
index 88bc7a3ed..d152c207d 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_72b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_72b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_7b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_7b_chat_qlora_custom_sft_e1.py
index 67b02d626..1098c5ca8 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_7b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen1_5_7b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen_1_8b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen_1_8b_chat_qlora_custom_sft_e1.py
index 327c3c7c3..2d517e897 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen_1_8b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen_1_8b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -112,7 +112,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen_72b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen_72b_qlora_custom_sft_e1.py
index eb5fed1f6..e1156a1aa 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen_72b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen_72b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -112,7 +112,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/qwen/qwen_7b_chat_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/qwen/qwen_7b_chat_qlora_custom_sft_e1.py
index 22acb42c8..b6fcaacba 100644
--- a/xtuner/configs/custom_dataset/sft/qwen/qwen_7b_chat_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/qwen/qwen_7b_chat_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -112,7 +112,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/starcoder/starcoder_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/starcoder/starcoder_qlora_custom_sft_e1.py
index d15b023da..d79484dcf 100644
--- a/xtuner/configs/custom_dataset/sft/starcoder/starcoder_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/starcoder/starcoder_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -114,7 +114,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     max_dataset_length=max_dataset_length,
diff --git a/xtuner/configs/custom_dataset/sft/yi/yi_34b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/yi/yi_34b_qlora_custom_sft_e1.py
index 2cbde95a7..4906ab5f7 100644
--- a/xtuner/configs/custom_dataset/sft/yi/yi_34b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/yi/yi_34b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/yi/yi_6b_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/yi/yi_6b_qlora_custom_sft_e1.py
index 5b1281964..96a684a22 100644
--- a/xtuner/configs/custom_dataset/sft/yi/yi_6b_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/yi/yi_6b_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,
diff --git a/xtuner/configs/custom_dataset/sft/zephyr/zephyr_7b_beta_qlora_custom_sft_e1.py b/xtuner/configs/custom_dataset/sft/zephyr/zephyr_7b_beta_qlora_custom_sft_e1.py
index 7cb010a99..b2349c2da 100644
--- a/xtuner/configs/custom_dataset/sft/zephyr/zephyr_7b_beta_qlora_custom_sft_e1.py
+++ b/xtuner/configs/custom_dataset/sft/zephyr/zephyr_7b_beta_qlora_custom_sft_e1.py
@@ -26,7 +26,7 @@
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
-from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
 from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
@@ -111,7 +111,7 @@
     dataset=dict(type=load_dataset, path='json', data_files=data_files),
     tokenizer=tokenizer,
     max_length=max_length,
-    dataset_map_fn=None,
+    dataset_map_fn=openai_map_fn,
     template_map_fn=dict(
         type=template_map_fn_factory, template=prompt_template),
     remove_unused_columns=True,

From 8c3b44d0c8d9936031f94700342d3768a19b8519 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Mon, 24 Jun 2024 19:27:35 +0800
Subject: [PATCH 04/29] [Fix] fix configs (#783)

fix configs
---
 .../llama3_70b_instruct_qlora_alpaca_e3_2k_gpu8.py           | 5 ++---
 .../qwen1_5_moe_a2_7_b_chat_full_alpaca_e3.py                | 3 +--
 2 files changed, 3 insertions(+), 5 deletions(-)

diff --git a/xtuner/configs/llama/llama3_70b_instruct/llama3_70b_instruct_qlora_alpaca_e3_2k_gpu8.py b/xtuner/configs/llama/llama3_70b_instruct/llama3_70b_instruct_qlora_alpaca_e3_2k_gpu8.py
index 13e1d3888..89feac44e 100644
--- a/xtuner/configs/llama/llama3_70b_instruct/llama3_70b_instruct_qlora_alpaca_e3_2k_gpu8.py
+++ b/xtuner/configs/llama/llama3_70b_instruct/llama3_70b_instruct_qlora_alpaca_e3_2k_gpu8.py
@@ -117,7 +117,7 @@
     batch_size=batch_size,
     num_workers=dataloader_num_workers,
     dataset=alpaca_en,
-    sampler=dict(type=sampler, shuffle=True, seed=1024),
+    sampler=dict(type=sampler, shuffle=True),
     collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
 
 #######################################################################
@@ -185,8 +185,7 @@
     checkpoint=dict(
         type=CheckpointHook,
         by_epoch=False,
-        interval=-1,
-        save_last=False,
+        interval=save_steps,
         max_keep_ckpts=save_total_limit),
     # set sampler seed in distributed evrionment.
     sampler_seed=dict(type=DistSamplerSeedHook),
diff --git a/xtuner/configs/qwen_moe/qwen1_5/qwen1_5_moe_a2_7_b_chat/qwen1_5_moe_a2_7_b_chat_full_alpaca_e3.py b/xtuner/configs/qwen_moe/qwen1_5/qwen1_5_moe_a2_7_b_chat/qwen1_5_moe_a2_7_b_chat_full_alpaca_e3.py
index 5053f10c0..6e8c2fb00 100644
--- a/xtuner/configs/qwen_moe/qwen1_5/qwen1_5_moe_a2_7_b_chat/qwen1_5_moe_a2_7_b_chat_full_alpaca_e3.py
+++ b/xtuner/configs/qwen_moe/qwen1_5/qwen1_5_moe_a2_7_b_chat/qwen1_5_moe_a2_7_b_chat_full_alpaca_e3.py
@@ -184,8 +184,7 @@
     checkpoint=dict(
         type=CheckpointHook,
         by_epoch=False,
-        interval=-1,
-        save_last=False,
+        interval=save_steps,
         max_keep_ckpts=save_total_limit),
     # set sampler seed in distributed evrionment.
     sampler_seed=dict(type=DistSamplerSeedHook),

From b98d41367ae6adb34a8cde1a592ceac688d2a9e2 Mon Sep 17 00:00:00 2001
From: RangiLyu <lyuchqi@gmail.com>
Date: Mon, 24 Jun 2024 19:27:52 +0800
Subject: [PATCH 05/29] [Docs] DPO and Reward Model documents (#751)

* rebase main

* refine link

* refine link

* resolve comments

* resolve comments
---
 docs/zh_cn/dpo/modify_settings.md             |  83 +++++++++++++
 docs/zh_cn/dpo/overview.md                    |  25 ++++
 docs/zh_cn/dpo/quick_start.md                 |  71 +++++++++++
 docs/zh_cn/index.rst                          |  17 +++
 .../reward_model/images/preference_data.png   | Bin 0 -> 27640 bytes
 .../reward_model/images/var_len_atten.png     | Bin 0 -> 41318 bytes
 docs/zh_cn/reward_model/modify_settings.md    | 100 ++++++++++++++++
 docs/zh_cn/reward_model/overview.md           |  29 +++++
 docs/zh_cn/reward_model/preference_data.md    | 110 ++++++++++++++++++
 docs/zh_cn/reward_model/quick_start.md        |  86 ++++++++++++++
 10 files changed, 521 insertions(+)
 create mode 100644 docs/zh_cn/dpo/modify_settings.md
 create mode 100644 docs/zh_cn/dpo/overview.md
 create mode 100644 docs/zh_cn/dpo/quick_start.md
 create mode 100644 docs/zh_cn/reward_model/images/preference_data.png
 create mode 100644 docs/zh_cn/reward_model/images/var_len_atten.png
 create mode 100644 docs/zh_cn/reward_model/modify_settings.md
 create mode 100644 docs/zh_cn/reward_model/overview.md
 create mode 100644 docs/zh_cn/reward_model/preference_data.md
 create mode 100644 docs/zh_cn/reward_model/quick_start.md

diff --git a/docs/zh_cn/dpo/modify_settings.md b/docs/zh_cn/dpo/modify_settings.md
new file mode 100644
index 000000000..7b4672792
--- /dev/null
+++ b/docs/zh_cn/dpo/modify_settings.md
@@ -0,0 +1,83 @@
+## 修改 DPO 训练配置
+
+本章节仅介绍与 DPO（Direct Preference Optimization）训练相关的配置参数，更多 XTuner 配置文件的细节，请参考[修改训练配置](https://xtuner.readthedocs.io/zh-cn/latest/training/modify_settings.html)
+
+### 损失函数
+
+在 DPO 训练中，你可以根据需求选择不同的损失函数类型。XTuner 提供了多种损失函数选项，如 `sigmoid`、`hinge`、`ipo` 等。可以通过设置 `dpo_loss_type` 参数来选择使用的损失函数类型。
+
+此外，你还可以通过调整 `loss_beta` 参数来控制损失函数中的温度系数。同时，`label_smoothing` 参数可以用于平滑标签。
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+dpo_loss_type = 'sigmoid'  # One of ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'sppo_hard', 'nca_pair', 'robust']
+loss_beta = 0.1
+label_smoothing = 0.0
+```
+
+### 修改模型
+
+用户可以修改 `pretrained_model_name_or_path` 对预训练模型进行修改。
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'
+```
+
+### 训练数据
+
+在 Reward Model 训练中，你可以通过 `max_length` 来指定单个样本序列的最大 token 数，XTuner 会自动对数据进行截断或是填充。
+
+```python
+# Data
+max_length = 2048
+```
+
+在配置文件中，我们通过 `train_dataset` 字段来指定训练数据集，你可以通过 `dataset` 字段指定数据集的加载方式，通过 `dataset_map_fn` 字段指定数据集的映射函数。
+
+```python
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(type=load_dataset, path='mlabonne/orpo-dpo-mix-40k'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=True,
+    is_reward=False,
+    reward_token_id=-1,
+    num_proc=32,
+    use_varlen_attn=use_varlen_attn,
+    max_packed_length=max_packed_length,
+    shuffle_before_pack=True,
+)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(
+        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))
+```
+
+上述配置中，我们使用了 `load_dataset` 来加载 huggingface 上的 `mlabonne/orpo-dpo-mix-40k` 数据集，使用 `orpo_dpo_mix_40k_map_fn` 作为数据集映射函数。
+
+关于如何处理数据集以及如何编写数据集映射函数，请参考[偏好数据集章节](../reward_model/preference_data.md)。
+
+### 加速训练
+
+在使用偏好数据训练时，我们推荐您开启[变长注意力机制](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/varlen_flash_attn.html)， 以避免单个偏好内的 chosen 和 rejected 的样本长度差异造成的显存浪费。你可以通过 `use_varlen_attn=True` 来开启变长注意力机制。
+
+XTuner 中还支持了大量的训练加速方法，关于它们的使用方法，请参考[加速策略章节](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/hyper_parameters.html)。
diff --git a/docs/zh_cn/dpo/overview.md b/docs/zh_cn/dpo/overview.md
new file mode 100644
index 000000000..d1bfc4379
--- /dev/null
+++ b/docs/zh_cn/dpo/overview.md
@@ -0,0 +1,25 @@
+## DPO 介绍
+
+### 简介
+
+DPO（Direct Preference Optimization，直接偏好优化）是一种在大语言模型训练中用于直接优化模型偏好的方法。与传统的强化学习方法不同，DPO 直接使用人类偏好数据进行模型优化，从而提高生成内容的质量，使其更符合人类偏好。DPO 利用人类偏好数据，直接对模型进行优化，省略了训练 Reward Model 的训练过程，与 PPO 相比进一步省去了 Critic Model，不但避免了复杂的强化学习算法，减少了训练开销，同时还提高了训练效率。
+
+DPO 拥有大量的衍生算法，它们对 DPO 的损失函数进行了一定程度上的改进，我们在 XTuner 中除了 DPO 还实现了[Identity Preference Optimisation (IPO)](https://huggingface.co/papers/2310.12036)，[Kahneman-Tversky Optimisation (KTO)](https://github.com/ContextualAI/HALOs)等论文中的损失函数，如需使用这些算法，请参考[修改 DPO 配置](./modify_settings.md)章节。我们也提供了一些[示例配置](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/dpo)用于参考。
+
+除了 DPO 之外，还出现了如 [ORPO](https://arxiv.org/abs/2403.07691) 等无需参考模型的对齐算法。ORPO 采用了对数比值（odds ratio）的概念来优化模型，通过在模型训练过程中惩罚那些被拒绝的样本，从而更有效地适应被选择的样本。ORPO 消除了对参考模型的依赖，使得训练过程更加简化且高效。XTuner 中 ORPO 的训练方式与 DPO 非常类似，我们提供了一些  ORPO 的[示例配置](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/orpo)，用户可以参考 DPO 的教程对配置进行修改。
+
+### XTuner 中 DPO 训练的优势
+
+XTuner 中的 DPO 训练具备以下显著优势：
+
+1. **支持最新的算法**：XTuner除了支持标准的 DPO 之外，还支持了大量的衍生算法，同时也支持ORPO等不依赖参考模型的高效算法。
+
+2. **减少显存浪费**：由于偏好数据中的 chosen 和 rejected 数据通常存在长度上的差异，因此在训练数据的拼接时会存在填充（padding token）,造成显存浪费。在 XTuner 中，基于 Flash Attention2 中的[变长注意力](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/varlen_flash_attn.html)功能，我们在训练过程中通过将偏好数据打包到同一个序列中，显著减少了由于 padding token 带来的显存浪费。这不仅提高了显存的利用效率，还使得在相同硬件条件下可以训练更大的模型或处理更多的数据。
+
+![img](../reward_model/images/var_len_atten.png)
+
+3. **高效训练**：借助 XTuner 的 QLoRA 训练功能，参考模型能够被转化为移除LoRA适配器的语言模型，从而省去了参考模型权重的显存占用，大幅降低了 DPO 的训练开销。
+
+### 开始训练
+
+请参阅[快速上手](./quick_start.md)来了解最基本的概念，若希望了解更多训练参数配置相关的内容，请参考[修改DPO配置](./modify_settings.md)章节。
diff --git a/docs/zh_cn/dpo/quick_start.md b/docs/zh_cn/dpo/quick_start.md
new file mode 100644
index 000000000..a92152b0f
--- /dev/null
+++ b/docs/zh_cn/dpo/quick_start.md
@@ -0,0 +1,71 @@
+## DPO 快速上手
+
+在本章节中，我们将介绍如何使用 XTuner 训练 1.8B 的 DPO（Direct Preference Optimization）模型，以帮助您快速上手。
+
+### 准备预训练模型权重
+
+我们使用经过 SFT 的语言模型[InternLM2-chat-1.8b-sft](https://huggingface.co/internlm/internlm2-chat-1_8b-sft)作为 DPO 模型的初始化模型来进行偏好对齐。
+
+在训练配置文件中设置`pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'`，则会在启动训练时自动下载模型文件。若您需要手动下载模型权重，那么请参考[准备预训练模型权重](https://xtuner.readthedocs.io/zh-cn/latest/preparation/pretrained_model.html)章节，其中详细说明了如何从 Huggingface 或者是 Modelscope 下载模型权重的方法。这里我们附上模型的 HuggingFace 链接与 ModelScope 链接：
+
+- HuggingFace 链接位于：https://huggingface.co/internlm/internlm2-chat-1_8b-sft
+- ModelScope 链接位于：https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-chat-1_8b-sft/summary
+
+### 准备训练数据
+
+在本教程中使用 Huggingface 上的[mlabonne/orpo-dpo-mix-40k](https://huggingface.co/datasets/mlabonne/orpo-dpo-mix-40k)数据集作为演示，
+
+```python
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='mlabonne/orpo-dpo-mix-40k'),
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=True,
+    is_reward=False,
+)
+```
+
+在配置文件中使用以上配置，即可自动下载并处理该数据集。如果您希望使用其他 Huggingface 上的开源数据集或是使用自定义的数据集，请参阅[偏好数据集](../reward_model/preference_data.md)章节。
+
+### 准备配置文件
+
+XTuner 提供了多个开箱即用的配置文件，可以通过 `xtuner list-cfg` 查看。我们执行如下指令，以复制一个配置文件到当前目录。
+
+```bash
+xtuner copy-cfg internlm2_chat_1_8b_dpo_full .
+```
+
+打开复制后的配置文件，如果您选择自动下载模型和数据集，则无需修改配置。若您希望填入您预先下载的模型路径和数据集路径，请修改配置中的`pretrained_model_name_or_path`以及`train_dataset`中`dataset`的`path`参数。
+
+更多的训练参数配置，请参阅[修改DPO训练配置](./modify_settings.md)章节。
+
+### 启动训练
+
+在完成上述操作后，便可以使用下面的指令启动训练任务了。
+
+```bash
+# 单机单卡
+xtuner train ./internlm2_chat_1_8b_dpo_full_copy.py
+# 单机多卡
+NPROC_PER_NODE=${GPU_NUM} xtuner train ./internlm2_chat_1_8b_dpo_full_copy.py
+# slurm 集群
+srun ${SRUN_ARGS} xtuner train ./internlm2_chat_1_8b_dpo_full_copy.py --launcher slurm
+```
+
+### 模型转换
+
+XTuner 已经集成好了将模型转换为 HuggingFace 格式的工具，我们只需要执行
+
+```bash
+# 创建存放 hf 格式参数的目录
+mkdir work_dirs/internlm2_chat_1_8b_dpo_full_copy/iter_15230_hf
+
+# 转换格式
+xtuner convert pth_to_hf internlm2_chat_1_8b_dpo_full_copy.py \
+                            work_dirs/internlm2_chat_1_8b_dpo_full_copy.py/iter_15230.pth \
+                            work_dirs/internlm2_chat_1_8b_dpo_full_copy.py/iter_15230_hf
+```
+
+便能够将 XTuner 的 ckpt 转换为 Huggingface 格式的模型。
diff --git a/docs/zh_cn/index.rst b/docs/zh_cn/index.rst
index afe6e1c76..4acf0e882 100644
--- a/docs/zh_cn/index.rst
+++ b/docs/zh_cn/index.rst
@@ -55,6 +55,23 @@
    training/modify_settings.rst
    training/visualization.rst
 
+.. toctree::
+   :maxdepth: 2
+   :caption: DPO
+
+   dpo/overview.md
+   dpo/quick_start.md
+   dpo/modify_settings.md
+
+.. toctree::
+   :maxdepth: 2
+   :caption: Reward Model
+
+   reward_model/overview.md
+   reward_model/quick_start.md
+   reward_model/modify_settings.md
+   reward_model/preference_data.md
+
 .. toctree::
    :maxdepth: 2
    :caption: 加速训练
diff --git a/docs/zh_cn/reward_model/images/preference_data.png b/docs/zh_cn/reward_model/images/preference_data.png
new file mode 100644
index 0000000000000000000000000000000000000000..a18ea64497b35cc5838febf3005c3bca9ec9ce7a
GIT binary patch
literal 27640
zcmdSBWmH|kmM*%1;O?#ol0eYl8X%Am9D)Q75D4yW!QDcFLxAA!?jAh2yL)hdwUcx1
z>vOyN_8oWJ9^<|J16jMOR;`*+v*!2B6(lb!iH<^o0)ar#-@F!Agg{{HArP1rBm{6J
z-Qqhr1VROQBmPp^DQR!sRZV%kP59{W$6>93_ty5B-^)jM@F?OU3Q7$4FDzf7D1J-m
zMvci`f<N)jj9$8X66h@RYM5aFHJw0P9xrOCgH~A{#_}-A>ieLp$ELx_jefNe(Z|Qt
z#9SDQLnlZ54u_SK+OAl)bsq~W@l7h<qtie?xTm3W=v2_Z*a`f#{`^))s}~0!gEAtA
zpuoVuoyTn0@(jMk!}7HnVwJj6w9rOhI!oR{8_|X6#ez10(gf|y*x(b-^mh=k*%BBc
zAx-`$nKVdJlGpa==YHwC-?X%C2HcE{^S>2wm$=NZ!h_Ze3cS1^X&Apo-ya}gLHmbB
z<LI_MT55y8)32><Q*~=7NpFv&p3#3dWNr88*|Umu7~KJqWzH9PYV|5pD$igzIMSp=
zVddYW{Kz6XTu4#en}*JhJqDk#B5iz<u*L;nKL7K|2i+?hJ<nJ0OPksIK?D6`Ddldw
zkzXFiU?PwBtB-t!o2d&4@hj8u$Q5KBH=S!^+rt5KAr}2eS7M)EQm733=8ac={`t_`
zSQHonEZD=r<hJvH#2G%O=asXj6Od<L<s)9B^{A-KxO;erewURQ*N%Acck$4c=Xm$e
zFOPp2!opR~e)`mAT5U1oZo2`Y^whM)EL42jU@`wJ^g+<QZh7CF`X7DLDwON@x1VZd
znGBg4W;!fZd}?!e1z%!DJe(ad>`F?K$A8QupjqE&&&X=WpliV_Ex(F|@aC`7Y(m2D
zyB4l*P*?3>sMm_L7z=i&64+l6Aqto$RK-@Oi^hH5*dJ(N{QO65Jy?0+)%XDn%BuEL
zFK4Ujx^HiNzmh0FTN-<+_YOfM?)$g6oadW6&wN)`K2TqRQEwQZdDYZ(Ed~VX{ZRyV
z-<LiWvnzSX<H<Z}^XJF8wMEoo75Ur(LHYSuk+~9oC*?11zi0j?i1k*4OodA5Mfwg&
zZMhNZY+72I7)?QeUH7JnmQDX0*;FYU5s$o6r1&_+O5W@W7SgXJF<Pj0p}iv0wp#H$
z&*#fjL@J_@lQT!EDuYEp!`i{}IrM@Nm9u;kRPr8<N?JOfuO9}>`(UlNtF}0s$n}o`
zqdk3!tk$gUbR_mJJiLvso{)J<NN`|CE-q<fm+{dADxt3~<2S<8;4Ed>r%&66v@#WM
zd^s}}r((ZkBv(j~pj5su{8o<*)pqf&mX!a{MtThXs!UYxhyTz%UNbk}(K(_@Q~a?K
z1tr%b5vM)Ne3dpXV)TDpx;d8IHfQJgV?OCwRkqouT*PGm`0lM->;j#S{-2>uq}bR*
zJcNKM9DPk?*?XKJpEGV!ITr@CVH521bg!!uzkN7F-^t@uzsa80x(fo%Glp${G%Lrs
zXAog93Fc?tx<%0Pd^5|O#>>w~n>h?rFO;XW%Oyx{_V-~}Sib8^I?sHPCLm~Q<mKny
zHYA3w;%jYdYtgBCSJi`WL^|2BqBZDfWAl`R^uCL(7i*tXm3HRTk#5Ep;U53hi>m^y
z{QYHv;;iBN<Y-s{m$4W7CQ@=c^ISB49x%GQ)rYKj2In1RwSYyE0!#NFZ}Qv97OqGe
zK0V`eJjc2PI%j4a5f#I0wJ@5;?K!c37B)Sr+LVyhnSD5)Ffx|j(eyP221Z+%DZeA^
z7dqm7c1<IwGfG&vI%&%f2x!=+x(Zx$JwEh~Vuc#C!YD5NQ<-=gOa9kFU+7VbR~F$P
zFS8;**Ue=}7*oKFp?}zO86J2!8>b@?HRtvog^EfAQ+m)K8i$Ojfqkn=lcfP7;9+*!
zSMujh!0mPaaTWhw?tMA9V?Tc~mVDHJwLuAv)KNpjz@$sBlV`x&`jpY}R?r-qeVme<
zssBZH$0DOJ-dsn7U-gx%-`3MKc?Mm}QGAxOJ%9P2)ZNP2l_=MJ1RKk_<)KPsB?dgJ
zid|TGSN?u?7Lm=v&m%>j3Xo0S7_sPYUy~Ef3=ucFI{1YdN>*5Oh8xi2%@;>_iYaWn
z8*D7LZNGhJ`91xGt*xFidQiJ7+Mb{7yQT1H$u7Jh;oAz6ajAn190hU#k6R*IM=1p>
zK}^C?menu3U)bjISwIJg%WHF@&B6|>Nu%Yd!sqwdUquf#qN8pvN=a(D8LjzF2+5?M
zzPe1hV0E(+&d?rZ5Ue*`$5*TKK<q+uXlG#I*4Ok7lvdkS8Tu|cIk+`a$uV=gjj2|p
zNAF^%^r5#d`G{l-{V*1>Z(f`xFI%PXed|mfV?klv3Z|zM!P|*LaYxg)zQv2-+mHG6
zewvc6ThMTRe)yE>6_QM>xz|73G@$QmIfW4sOBOoQZW@y^Uy?I)JnMNM)`AU(uEmL6
zDRL8@O_RUGRX5A4+fjmT!yOQawlvkd%D~j$9=YJzCg90fn<&Pv{>6%P{h_ksr!z;G
z1A0a%W6x6t@5GESwmiCYf}Z&iga@W#Uokn!+C))tgq5jLY+=D?-NQorYZfTI#;+++
z*R&a5>~MzSIGmt{=;`-HJPzz(HDlsWPcOsy$^6ZR+b<AJe7X}yIQJeie17KVJYIB^
zes$I5^0iimcDpf07Y}h5aCf)vZN0|gSz@=&8wM5M3ZFJTNdJKPpTB!@ifE84xBm7X
zvJJP-R_+>U4RX{-uMM{t0r(|vmr_(c#BKFO`Z>GM|GS}&(V+kS-FBg_4!Q&;#Cz!p
zM#UyCtQ5%@{3LS7?NrEB&dJG%tS1&9>sQYEd2E}R*QqrdlbN+B(h@opmEa+EkR@QR
zIBMAfMt;-ah*tl63JhjtZ4VFGn`tD?kcjXP$!FBm)GB>O1Dol1u$8>JvJ-q*UUoFS
zF+jAop2@2dlaLS$k4k8=L%g>BPG9R~Yb#GwQs~&2@}s7^s=>;4&+wI>eI2xy2%`uC
z7kgbC_7}oTCyTZ<UB|XYK+^(MhmvY;i0<#0rwL`^12QWM1DCUjDMZ`T$d3j2_0FP+
zYdklrv}>i_@up`GnEbrkmyt8jyj3Y1Oq{gECa=VTX?NKfV;oP6i%kyaRdTc6k>RV5
zES{@X3b6ZGsLc+?UA(#Tl@ttjaX2UOD@~B^0w47Hn)jQpndNJqja%uuQZnC$^l>NX
zar%?)U_LY|CHcv2u7?>AMcmRZ1QI8#ADVTLw;nr~^4+<)^MX}ZP43CTe1Rk8TO+v~
zYiav4{YtIG9hUne{HZgY&^P6dt*%~zF80NDcX-{$!JE(c<R+0_K3J*PrC_|;+_s`@
zLg{WE7$IY+yTb||9F0m=%~P$`dg#2|Qo3u-D0Lldb-#R;><+}9u*=9NCkbRlyNfq_
zF_IlH+%!qnXkAlTS<Tg0=h*hdQ%*Xe5!K<2@G-b~%-09S&67l2f2}s7MDEpt`TD@1
zpqDBi>B1>^VZTcGwNx*7OJy#OM#jYW1_z^%@wySI$Hyn!(MNwNH9r1OQ(wPoCWTVQ
z&VgNRnVMAumD2dv=xFb)=NeeGE}Y<tQrm8fq}Q@t8(YS{&)@rolk>04#dGi^M#IX<
z$u$U@elxIYa=l;{6%}JMS$l0@_>qQ@1GlxcRSh5evsS&M>Cab|3!%yU4iJdL!BpDQ
zw_{1r<^SQ~_{jE~S;t6)>(u!=Eu<wH6VC4EJC&HnChAy8n%B$5GlF}<Q30_5p)Z|!
zW?f*;H_kMD;cu2}tM*IujLNbalJ7#GLXuT$@Z97QO1Z%B5A#5k96f;c>{l<tuFp{2
z8N2EHCI?j*OPS<sB^|YrRD-4YMq1d18W)&Uv(RpE_rN(DqD@(H(Xq1ce2q}$U`pnI
z;r8h7N)H5KR-=#}0#HF3J`+8y#Sem%mQxWu&jVl}Q5qPgJ4F`TNyu9h^Vq4B((Pa<
zHIg%U#b$oRO;_Xo%I)4@3t@2sJNB821RpKrPim#QXcn59pSN)l(|lef@{L>P|3EXX
zPHOADD^F`84f#yr0P$K~+*n#KRXC#1jR>gS{SGci5{n(5-^^lrUZOx+vUYWrHs<hS
ziwL&I;*(W_Q-6B4_Le)lOy2N4MO~FR8@c6W=Ofm*NReErB9>nad!;ucg|~1|-*7h}
zg)mT4JgQK1O<1I(33E2DHsFzaUZf*L(Kq4dbEt^AbE<7uV(jC^c2Hw&4bSC{;|1ot
zwcyD{lkr8t?nW<p$C#PO0hyMTRt=8;p^sS$0{Ny@r>mew)IZQ)Vsxn8`1=!xiLGsI
z2bzBIJC?9Aq2;R-v?O|XB<MpMo@J-D^^RbZa#@V!GhVH|5iim%taY#?Bn+kG#z4fN
z@b&S5TR1&|gTy5!>O)8E@3*=>{nY4!K_(DbX*!|1B(ncF&)waPb=L>8L~m81cUX3e
zY`)o+6w$%5V^Qdq2xE`};uWq{N5k*}_Tpgf7jo3(NjcjZrzaW-9|2@|Vh;AOUoB8p
ztH63=O_lGg2qiHX+furXpMxM>Kej00w}&O7zRl5xWQvJY3Za&vug}mZ`4w!Ne9945
z<nD7meVH1}RM2TDe8}fJJ3H9*gVVXyI$W?=GtzIR*LX4MXmu_!+#na_d=+X?Y;W&c
zRfIt0+io#mDK|>`!j9>ig!|`JoqevIxXJ!LBUZCh4I&{2nwbO*2-bm~%G7%UZ6PK7
z2tHy}2koccZ7aIFSJ)O`W2kA_9?`FSq}B~}ZQKo>b|*mA$Z7>jV0S+B9w%ZvWN{dI
z5JrCvyL8L`(eu#J7*eh&n5Tqx?Q5C&ui~x_bPWs^Cs2cWjck@zGw<qewM<}Cn26wv
z5SL;2xU?R;eV>oY1BGBVS{O4u1bP%0jx>gPqPS1Kr5Au}O-Xa-dgo%ovGR^SDKOD5
zrUPs4*;AjwO-Uq_5LXj_9uSC0gpuV|56#Vy2?+_kmYvW^=<M#6v|QXC%}%SR7(-5c
z2QM4v74qqmOImqzb~fEsFajD81VY4Rm!mBx8y?;>G&DJTxU}R&A?UKTPswM11fkmh
zMceCu&B;xu7@n>~XeyI~SIJr30*s4~8q<qzOJ%(NP&fGq>i1n8E%c@swj;&^!caPZ
zxbW6?JboVkScV5(S8A&^T=pM%sRl%}UyTNwo!^AJYgp=F1}s2)y|fa6B)WESKfb20
z1yZ}v9_)F2Q#9d1qW``6{;<zto_wS+mg5&o68lc&8y$HyGGIP!@26q>J~3jGW9J|T
zqc3mFy<^$R0$b{Gq34dT;z;7-o9~)uEHYc9OBUIaKK0L;@!{t7mxO1~vj;y#kUwxC
z5^N0ZRH#1t5whR=bofDe)UzUSL8!FRvS~u{Pz^DJ3yDbM;o(tW5;+sNM<6hM-NHSu
zkwXg!c{R;_`k$eCu}3Q*EK&!K)(#SHa_@Q4Hd9<>Ep=R6_+q3C!zN1(CeR`Bc1jnX
zF%%Tn5Awl3+880qDk?ksyNBCP)hB?AKzv=}`QQ;YNo^u)fp_aA`)$Jh+3w8bTbl>e
zWIIoS6HhFmkNB?2`KG28pP^<_za<Ir`c=nYcVW{tcOgNwJxtL{jm(YWG5lzJC&m1+
zz<s3)i)k_8*+W!euFQCl{oQ(TvMJMOV=Su0ZJbv2HA%WFi;+x(g?*!7!_FAL%zy?W
z>YeAJ4XUr?w3MH5>E!7L32I0hVui`&bH)dF;_2#<$x6}&#HroWYwa4F424!)NdLfn
zLA9LO+p}TjG8G5oG0s$p!gUKsZ}E<mZ{Eyn-x>eB1y@sJMHtPoBLADmCgC1YpCmlS
zz2<?vQx-2qjKyT|Tnz}BX$w*&*Qx1W&*3Lfqc$t&MLJ(emnXbF6Q$2ofHTtI*#3?R
zF;6iyt=Lr@NPi5==DtGJalnhHF%-diGT#~}WG$cJ;+U`W?w%5|B?HM1iimy8{1NZN
z_W*L)<UXGVV{WMTO!FaQ)o&FtyLX>@*~U!_Tz=?at>JpJuGZeuB!R?$(7la|)a>+G
z%+im2Z#wpbLgEiGy146oI9C%AIsCQ2MjegTfJrfSU;07ojm$HSAE}=AwO#ta>B{ZB
zq%)D7so;#IJz8LFgg4^eXZ=yNK>+n4{t?N6AX1X02oeK3C6Z1(p~aoft8Qkk<(WxT
zuwyN*nHTo@xY9}#HkYMu=#h6dag9sg8VlNWvAQ|$J%eg@h9h^s%OQTK(|GR4nTGu*
zOb8!tOUl~WGV>+aSx$W1tQel6F2L=<7}J+%4K@U&IFKti?WPj(^}Mvz9@!7DI?+Oo
zSeP#nuzPV`=P-IoHpyvB`qtzcy<p0UY<qFBqYcwdA+D;dD`j;fu$}0&rSa#jt}s$7
zE2~HWmE?c`L@6ngdqdCN3)87vp`_N~-@i|dYGFkqS)M-kOD-%_D0*vP>~PMqqP_jx
zXW}UaWlrv7x#Byh3rNeCk?UPu>EM`b(!S){rh@1t<)zJDJz1Gr3V3j>?qUa<fbQg|
z`pKqIDo1|l9*giUW<cJ+*=<6=YvpE##41kD^wS^1(&$E`XDo4)Yd)#JC-tM7=|4UX
z;_iu7Unw)R*ZTZjhm!P3Dz^VU5Mnm2aR9nQ%gE@1h|7db4X@*PyxRGCX+Ff)y8f}Y
zeclNQ&iI*Qx%&rK%dauxxhZA8b?m-bC9WL??Gmc$dPEK>;oHIZYr$esQWjUsWhzqh
z@~YcJ$Hh>K`G5aWp=_z8p)hxUM<i^a874+YvisqE!rHHo*|Mqr@$u?mtI<ITq~J+w
z95=r$I86enOSm{hlL>2(A{t@Z?R*F;Uit7OB2@q<02nZ*7#=iqu|E9RTV_5m0|RlX
z$BZ_khk!_$AV2oC+!NDKY|Dla!P0!%h;_s4RTqz{!6ywx^Lw1#d3nlTS1(oTNyZRz
zEB!6ChdR%0&1#&R(x;7#9%zcz-HW+<<;>`bV`~g1iqOA6wc~1%v?(UB+=tezFm=**
zBo9PCzBfnj`1EF`TXBH*qXCgg)x^;5qxR#Bd8^-M6oXoIMdfR8Jrmc?jqYb!$GX6l
znEF${npvt}u>=a+Wy+qd=`;s`Fl4Gy4+k1Lf#3ukW61_2PZ+*RfwPJR$Jo+LrD+^7
zS!mVAk529d_tKV6L9nqj8i-U79M}@HW-yS4y5d7&V`Ga=Hg;ssV9Mb!$742Cqi`~e
z41!}8T-?$R>r#z|zd9Dy+uB5B)YCW8H5?@y1PsuG7HgDoe;q8=h6!q@W&lUlI8L-<
zqR_#W#BYxT4sE6`PKV&jj>g1N+Y^+2^>i(g;oeM243A^S<jjh}WIkW8(JyiCp+nj_
z%8m~`0wXpXGgK{PoLB`AuefrO<=CxfU<Hgm7M&~w135xFfEGTzts(Yw4MB`A+?uR&
z2L)L`F#32Q?uTkKaKw<M^8r|y`FFMFXQDIlt2~Z7ksE4VotD@HdjUX_`Q_}LUW=;c
zQ|0!RS!UV}d=s9(sU;eYEgFku&^d<NF^)CvC$_9J6NQjlyzzCY9fo*iN@Qf3VIGJz
zmJdD@62iiD)#<ke7qgm`9Asw<Vth;Bbj&Y47+xG80KOT!elzK#&h0XzTIq7lniQ5C
z&fTZ)2CnXK+#i+^?Qf>>zCBAkQx5GdL<fnKg@-$2gU^<Ye&Eaz7a!j`*^Sux^5xl1
z(An8>D=JHa*X5-X6B;TUgyH$~<?ZdD#{rn6g{p)QKF5Pl?z>xPfCr8i1Q0#mvLP<>
zUm6$#b90)w{i+3;$eEe6;ch2kBQp}pI;cX^-cSkLf5Egq9DFV?IYH**AKux$ODZH(
z1q5j}6(0&rg4MMQ)1yckoE<i{<|k1-A66<B>MEOW6@%Po)3GRMm<T7pj^8$UFpg{o
zf<z?K*%ZEY%^Mf$7jETQu~U&9LALY}9eVGkl;`Sg^CwK1FP{+9SBvZJB;LKaIGz}D
zcB+!mAe(AbD89d<?55%~N&Bd{u&I-^9TQ(IKmuYxiM3zx5HDbpYzN7hOBW6$vb7*m
z{qSjw@X~EnabR>rn}Qu7gr@3E5u;|T=@&PNh8^FNfQ{vH@~3g7x|9&T_%iE^Yv4WZ
zZIXPh?IrsU%86J?U#PXHIYM0$caQ}zp78xE2#Dr0YP<R!5%m{0E{;@=&Rr_Gbwm04
zQ*O#-wJZj;N<FSFUB>mF;Q`dZd`oJ-VNrCYPF_s)EyB6G3IUy5{O>X0mrM*iL=+Sh
zai>KdmdNqp+#BM0quIM_76_H{K=B{*Kdw}l>EVDcCm4LS?6FOI8rsWkg{w_$%tEZJ
zk>0aq;1>Jd3)i~bptfn>mxDn-zv0Y<H}o?!`z!rzr0oDCvN%-6#mCwLo8>5kJ(z6Y
z2KLBzq;5V4aa|(zpW=NAK5SNKvtZGn4VJp<ct*n>n*QsC)%AckW^TkYof2CxS?VAr
zy>U7<(zX1)uO(``r%@d_`=HC6CrqJhk2(07EaKtB@=-XEnyVg7`ujb82MYDcx5M*;
zFpvE)V(WRu4Iq+VBGPggFvYYULWhr4Kmy62jP9Q9r{}v5>jJBVQwAXE6q$Q}c_4)m
zp47`KaV9?IO~(7>a|~l`ptcZcEC=)B^;`AYL~-%37@K>hA_MK#jdfdV)^!*Mvej&^
z=Of|znLJMC)ElzyvpJKW@WEnQY)PfwolvIjovL>12?wV>oIM<63s?9%&vH3!5c{RE
zyb`C|n(S^~mgp5ss)_XbfvSG#bbxniZ|!B+y;*wFGI==!QeYE>4~XWerXfZt%K{qd
z*{!SH5ek~`y(>;bUN_poEK|ygHZQadoy`btrd`Lv4&5t+VQhGcFmAwqIa};Dxs_KA
z63o6Qbd_C@ybSnqq*#0)(U_&+&lU*sn^=|h!Zt<o;vRaXrOaaDbUE2!II`#Ste)?w
z_|%LkZ{DkDf9@H=JrZk`!NcR4Uo{EJIo>lbST?W>gCB)G^(dM&pADEZ^}Ael8Hn(u
z8(kV}<VbRv(mfcfw!fz!3Q_nAD@7$;UQzyAe2M*k0;&G11pL1PsQ$A7#w9<{e3wVk
zNlUY-_BUz$U!z?A9K^p$502HX#(^eL6o!WvW;7cvKK^?Bt(0%Cg^EhudL~_q`4kly
zZlAh33=E7Y`H$?%znB|4F)eM2FLH|?>UFn`tzYq_lisTUfS+{drL1ltDJvWI!L=lC
zwwW3;lsFVZaj1_UMdN)ReXqcKEyKCcD401_4n^S%7#LNI#qx_;tj?Z&SApy7Z27(3
zb{~w=G9)*Pn;4%fXmn*mMK8%fiIa|Ao(_jtVNXbiY6v`;t7hJ0jo)8ff!()XT6b4P
zPb2qHbTmX&bw4$h86TF3>d`_nzu$f$ccs1Bdz0$P0vP3IX%taJk%a{nHvlaXNlwUX
zr1<nF!H!MbR%<)6f4rL6_XxqHdi0m?q{V#2%|<_r(0$DY#rsS}r}?^fL=UdlcI~sX
zqW|b4`YG38c2ucOdOT(Fr__Gar%~ViU=-bH_V03(X%}2ClpR;pooXChcU5$^<Xh%6
z<Gy!MqecIP%GL--8ABV-E(p7}VMTki>Q&UnU%oI5X2o!td120C`O1&=<-ly>XI0og
zY7|ASOm4}Dp^#IwQ{q!)ZXa=Vh0MsAqui3K)Ru?l#rpC&n}WvR_&cQT&Q8%IFzUIg
zGQ`?CtQixx{f{CTAo(DCs_uM;o5yc0TED-lWUD3SZ>(zRgrllTO{Em}PyB9YYBm~2
z@A5NGfry*2_&BGH(GWqEn3!LHxSINQ_)uZpT3jt}i9RB$)%PQ+pgh5jch#NLSU*|=
zq@WlmaWGHAQU~(Bs=ZVumgcLJlo>2DLwGbIsxfMWMHo2Wlm#$&B|zumRP5v6s2<Jt
zehIVHkz+gmi&`ow;*@wa9xVmrk&*k^JFsXDySpT$rn2%5OsXOk5HcgRsxYw#;mhPc
zQdIP1tCPw+M>dtpEMcLS<o^`*0_x+RJ3fZ`-N4Y#g8za9|L+=El%~gx%mG21z(2sO
z^QF&6Q;RBvzUJ&C|Gcys@oUz=>MuMfW4<Z-Z?La#Da?&B2!A1=v6v4Zyl&lqDl9V!
z>r6>Xq{-)xo79rN1ZqD<BmN5JWD|835d*1swHviKDypTy`FBO1?;kKu;UqE{J&6}C
zCbn-Lr@h`&r>5E)c6mraLBu$xr14`I*V@c1oVK<|clhs$(Kse1UsqcPi(h^0&{d3T
zSXuEcal9y+d(PG*o9t^vN|aMC%q)x4d~wh_T9h?bTaI;}-;5~H_7}KPuUDCE@70z`
zbL6LmcRJ1aG&3HPGFERc{^tRoMED@`YFE<c30XML@nmR+|G{3@mNyVI8&Sk1;W&(b
z9xS;C|C0+aPx1%jlaDMRc_rbm!15Prm$95=*8OepU6qB3MeblCl8lVZ*KpF|;1Qh+
z`;1}DAQXP*Y&_ZYQ^mw|o_+ZnyZHhNaQ%Vf(zEDg6#)vo;1KCXCr6;&*^DaDriTVy
z!_yQ`3a3l?xU`2hRN(Q}j+Q=`{PvH^FgDGYkOd8%(DJA`84jD*YCFL=-A--A>J9w^
z=+8Lc`}S2;ccALg8Xg{Xa|u!ZxVSPQiXW_?d5lwT_H)s!>kJ>i_zg96XZQ(=w<3dQ
zG;3O$hxqB)exhk2Ec}cO#+E*j7e73TR{tx6D`oyid2FfO9R}L|X6S-oeOp)A<M-7o
zZan`>gsTjl|9A#V_0+an40Bf2c)srI+yHGl-{sk4ca4e@W^B=1OX;X;soOao8s!`z
zz|bDM-EoDUsC)?AEw|d6Ab1#+qIR_HeSLqU7B-jol_&WM6(cK##%<7*NT?=Nq4*Qn
z@XLOqFljZz=_AzZ#l$AEF~1b2E3nTu8qP2^I^CKfHorebWST9(6|8$&YSu1$*?bUm
ztS<ntE9`Nl-HE~_T}xy{*S7+b=MhjKtj@7g{PKD~PN)nwtKp+}1QMMfm;($UKS#|}
ze42Y=H_~DWIRy(ZQDwj+fY<aCQdg@VZb?5sMj)v(p)vV<PA>OqqH^}guDX3JsJLgI
zjE3FI+_kxVopJIwuP>EtXc1t1k<t@S<|M8>pO@OC9H)FYim`_0g@0vbTfrs)!YkNX
z{+Xq0N&KzD5disMMSVVXJp7#cvyZJpjvqi?>?FUsVQVK*!d%cZeCKc5K1`|oUPYo|
zM<eYEXHdNhrKm~<Ko~ahYp~!9N_V9H#}JpaNPm}k+}1_P+yHA}kjG(sf1~`#e}Q!2
zvYxc(PJOA-I{1i%&B>2e|8GIBt=w`;Be#$p`>!V+8Z)jk229}ll^o?~%F4<;^JwYW
zzmk8s+Hj}6>2+>?zTe2@KXyur6nD%EFt48Zm~2^)mbgT)U=O+o*Q<uV5n@DTE-Hd}
z0jNSBpi=+`^8@gg$}92jr-S5;wg3k^29dGQYy+`RL&J;B5Je@WQr%@y8A0%`NdY4W
zl>w9z1pu*@>cxrLDydA6#rah`m@|?AE<RHqDw?mi-k;%2HjN+=Bmqs&eqx-TBS556
z*k37qd(FI}HrtRy$@Xg&Q)Pb-yWBc)!|AvX#|;lBoUm|SZ+wt~<$@y5w!uYkeEiG}
z($jZ^-Kw<Fl9$|=wbzt<4Dn)&Eo#e478KH`hc{}L8IQ@;p5zYdesKZq3R+UJY)ba-
zj%V+EadZ4#AX~3f;m|Efz0A<`E?re6Z$Pno$EF$F*n0ifJ7|MJcj;4COjTyB@m!69
z;%kpweS7!f+>gfr#Vth-AE2-kNCJ)IyYBZ91g|=E+Ttf??>-V_;{k|OK>CSY4W^DC
zfG~{13FkeK0P58<$T*%q3wpYpZc{hj{dc2WzZg^hKZ9~TslK9g^EJc{k-yD>a0n-%
zh#=(^J2%2WcpNqu@r0kX)WykJ@5$o`p&%Dm4qhkexaB{(e@x%jJB>a4Dpmn>>QiNr
zm`XCdq&p^>yE%J#oB@mGVNZq7%!kz(fnqm<0*Oj3fXXtClnWJHiPW+qmK_W~qU2Rz
z=}=+O03!$$qqm+0QGUP8mbCX+Y<eT$kD2Y4{?O;nxi17}A@EFfnX{Oi1TO_<Y}*5d
zYJH5iAY{t7r@44+&i3JD%FzjW3|f_c1K0L509pcM(s<Dn!-4jPUZ*!36eP-$BlE=|
zFyB$`E{tX`n72WpzZmkrwqQbVnSz0YdnWqCc1u7(QHqTsA8=XpBU<O|p`|3qAV46A
zMV_A%ZB;ev-v3ZC0+@%#vj=$kp^a^=!`&ZDp`_!IC!63PX*Tal5j6Wz#GRV}B;OUB
zZc>lVH5iFstTxF?*6!%kcv!=aIqGxXWX{xkTsbjmoS-+oeV#MwI?|es4%=0;+SQ3a
zb@~PZ=A7pG&>0viD*<wq_IVk#fch2_69YxBnCk6dAZiZV-mmoZSh+dcQyN{*er9K8
zB6klAaNbBA&b79cuok>8g|)oC`V_7X8jh!WpwLT#)IyCyanafrdhA-4%8i*uElpOC
zTO#9e!UBnQr_muyh^~c2ICNAFj&^{EobB*h9Z^jbsr!6}tBlAw2YHYJOraM4Fj87E
zfHPU2&nip1dZGYPk4#hET<lwIpWvtzp$B}e#PP!NrFm$=ajiNm=Ij~YbSvHK$INa^
zZ#)vuSM^8lx}oNXbd!ECDym!2puIHPC)f*I^afubK-4NM1L7%VJHM+Ud^o9_qW=c(
zHR4gMr`4Z**?nQZ>;A0io%h9JVn~?A<WgZTOoD@kI%BDGw@b~bGctBx`i`=y82Qx>
z&-oLJ03Rhe$c!t+Sr-x$rpCLzX#=Eb=S+YEK>?wfcZ&W19)hAoAVqgp#Q5<-{fSvT
z0ub`Zw6N(L0TmVxF?t@vP;1GK^hxXrR#_#BDP6=Ln#1T{0G6q@QSTEK<*vc<gf6w$
zFCsQ$V`wpxKo@A4l|#uRsz+P@AHu0lDQWrbdaxiOGnsFr_i0R+wa+QOS6dF9-yM<X
zKG-fTjp4GKIkZh%o&}t}IK35|z0lWOB6Ts`WQq3~;GdYBxz`W)qPH+8%~AsArPHED
zcC}L@J6$wi@T9NqV%G=!VbZ8Q?vjVFkw+)zi3nVJp{ks<dzg#9FP(wvIw32;!=;d<
z4Ff!}?0+{fmGy8k{>P3c9)w|Nj1b?7OFmlBJK$4e<kvl|PsGL}P`wFa2Pe*k%axGm
z8?B^d1;spJ*{#~ml+oGPloX6iMfFwAf_i&62;ThI9c5-ffMnR_Yd=*;PC>!g?BwF8
zQdvQ5EiZ_Etsd_0NQ%YA$doaqKNOorXwRE6c^+%N4?BZuyID3fx@W(utOK*=b;A?q
zg2MPCez9A%hn3nFYO`Zx#;JA%c{4kuyLxbt8tcLm*)vXG4V`VahsOCN7zpOUL-9zF
zjQ?b-=)+?59PF|}l_K@~J<i5B>RWQs<Ge4j|M#$`L60kSqSVbzQZ@bf)?z8oXvM{o
z&ayCtw7g;IEEDFAG}5jo2v<8lTwHsP{u?o-S%i8-v8TXgn^q->gG?b533e;zSK}zE
z<ptsbfl_bF%~C6zkCvE;Vm$TT?a2@)1>@7F&L1$@6>LvMyPIJlIC33tJmxHWPG$vE
zt9$w7J{T5uEU6-L<s)jm=*05=iL$*YNY*rL*AX!xW`x!r+g;Q)Yx!Kj#j(BMLc29k
zHJL=K`CaTe<U0q<-WzSo1MlK5Wvs$Ihyl@MOM)VonBRFm@7rw1oL1}xwrM<c48O!k
zIDsMtiHA~7rw9J)X&C?tOw8J>DSA1p1_C$TPXE={5BGEg9!nu@6<;~e$+#uSx<p*W
zP{p||=Dd7-eAY^L1Ox=+6%|idE;O2T+BYti(fq%~W()V>V|ly1`Sf(p%sk9Si$=e+
zOr@<MnfPG-(-c<HbrOqgk;ztN2cHE#{$#Eh*!JX@Omkc;G;u$6&Ylh5-zHXbd|{d$
zwr+&Kn%uS9r@sqG&uWoA)Bi8wqF#6^X{vmZ{|GH&(~BDls!=C8&QnTOjm#-(PTa?x
zb<O`#AbSS*Jn7}Wvw-BI*<1q)khhe)@2DWX0!^4Vx<C7h(%ah`{iWjv8ZrAL+RRLs
z-I?H(^TX~->s1pn_7OG0(O*wux%lm`UjPDy30bW4SW?)zD>~w8Q?vkG<u_CPK}z7l
zd&@xVAbAwSTUm8xcnDMDUanRqv%*)C-dc=NP`dZ^t*Z6S$N9DcQ++q9n~!WDa-ef|
zGuk7JyIgpbObyTMt{qLbOV+&E$EQ_ZSEGd-{gRjM9t(sf^0wy3!g$27&9ML+`t~p2
z(7hK1@F7$0g^Q6h-r<Az@M1aNuB`$9Mga7A1$Z)0+z8o!RkD70ZCW6y-cHM?mo2LB
z?8N0}r+%eX1e0GaPMC3e0pMv^o?BAMJH5>wnDscnN=)0I%+=Ju0v-e@S<8A?$y@nn
zU#BWvt3T$nR+>(-a_|tD5A+{hPnFu|wU@mmbR)t3{0<u%%0g4Of9W<XCTqd$?c*Kl
zfVtWo)(*B~PRwNaT;)r4Hf%<fqA?9RnLRCLCZ@3%T9fcB&m)SmqI<DT805}sDJ%g)
zjGvw^KSuSZr%rq}WxJ;>m7ozkU4y&hZ>Dpz2p=7<^NW@cX4#r|$^}m!H@#P9Joe~8
zlSYpFd?=eT9z!_Udgg2%_%Ri1uF`In-UidP9E2-(k;L{co&^P8FU@ZFpa%g<oQO{|
zG(u#LcXx!bu1Cz0Z9DKuIsyzKZTij;<WkXOa9X2x!4i#ZupwfxrgT?9n8!R9TF=_P
z4UIm%0rwLxju`j7{hdmIjCpGnb)VB%Wva9ff;4K6zC{wjBD+y~7`2iWMgVtbBUj~a
zb1w<&GtOxhoQdiJc!5dn7~da#;?8btx>d1!!{Th5{JMUc=J5^Jhl`}qy%1qHct$R=
zSWV=WH|fIYF=oY<OA<`CE!DF2NN>LmxIF7xoA3wR1rIA$2-Q(x^%9^G=8k5y0xA$&
zBKKrI{RBXUNV&~Kq@+*~TwU8c$MLDt(pb5<yQA6z!Y1-jA<6uXU3-|q;s6u{XR}US
zOG~%2v&l=dQMyYNjb{KBB1@GJ*BxDaz#vprZ4k6hADMYIMx-(!MaKX5w5n<24546V
zdbMye`+Kq<FNkD{?QxXd2OX_;f^c&KqP`@f8YMQBm>5OaV<(na?~){EHr^8m5KnNa
zB(QHPF&l>u=&&O)PWtO>b<hncHDh$S`+;alw_?$h&s?LHi|nOsmj1LienMGqx2*o+
zB>(V4fqC*`1~kl)yd`)-$zS+s6dHqrn-3bM{Bn)ZCgvi--i_7o%i2%KE?i%~Ra;eC
zrJrYkIAM}MYW6CP6FP=dt1RypiKTW@E)s1U5lmQ^%d?kpW5H5_{RmEef)dSq7~fRB
z#Wjo9Z8iW|juIc|;SN$J1jL}iUxut|GVsZDCmABLS}0SBOK!=*l8?!$!5cPo{L)=P
zK^vzZK|d{NFBPQJ)zHk_$Tiwz*6MB=xR#SBkZB=ul+k)O3yq}*SOdO^r>W$R`xlK3
z-aLO&gvx1GX`s7kWT*#7n{s!!9p2tP%gf83-yAkEIx6)9F2v#baRiupny9_SFSjOT
zPI+spg4-dT{Eq(yL5}Ll=gBxLyn_pq-@Dfc7Hy1h`)tCmJ?AgKF#uE;1ba*3b4=<*
zmi)_TC__Y+joN?npHaQ+@2Bq0jxD?L5(3{79bfboHw%5Mr5#YZl8eBqW=wlC@s*Yx
z?|Ad7IRgD2hC}dG!pe<hOv!AYN>I9@P-$==T1L*P5L5r`(tEvjBjv=rxUktJvr$6G
z(Zm`gUPDaTaT4V)t?>#VNlFj7abq@X3kANThkf<LA^(CT<y>tGiHqA;9!@zMEHX$4
zJMv4r-{1p|u{grfRf=U)iZz^=%L8%R>h5`<F}tAYie>t(Dy_;O$3n1-#P&3L{7%GS
z(b`f^Jet0;ocQyeC=TM+wb3C1%eIkr*REtd1<exAum!a+$Yk4tEX$A7FEz=eoxRAr
zrPeKQdUd-?70Co{uxzuN%*S>sP%tsMFHbj&9&RNW`0w!9ZX#ATJ@1480lQFdp$a|4
zSwe5?Ep`<0HCVSA1shCyW&`SNQ39b9et35+GTrIz;!Cqh?AzOuhxXL_DH4MHnWcCm
zo}q7BpN@8mBaT<k>UrE~-~z&RUIK*3FLgJ3NPj=&ME>sWXQPK43B(iO@AxO)z1N5O
zoE{8$9Vto=A%gUWk8j)GIOVE3LdoLoMB;!g9yn7+$vbI5(_CU9*FjAK>y^SK#WU^L
ze*5Y3mTsDRhbc6bXci(lc$U~x@1W5|fTQ5fYgg_4r4e0p5q0>cR=e7q(xw=HC1BUl
z(2SZBKKmefd8{1)tU}?*MT=S}$)AFdz_pH{2WW!(e>b81Kgyc^3Hbc8HtPS*08t}X
z*Vm$H@>}lH)6>)p3`n`Txi;JiTv1U`pkBb9IW-|6&-o!ZIGDsvTd8sJtdU!PQcV20
zl4&iei1>4#!oufSGKpE`P+kZf9ReF0+f=isFf1&rfyw;nb{|V~$L-b0t}+)h^I%UQ
zHH}FMfGKj`OifR}K`5`a-E5njB*dkY_5y?&Gcz*-<N49<i3tLYYKuapsppK0V$#xR
z9|x0y;^N}yjLESTPFg|Y_(?tnIiK!ql{v&WMJg}R`D7!lqy+2u__*>L{^DfRV@Z;F
zb#d_l3;?_I4o*+2AJ7TdlaZ4zZ*2H$)jN#8=X&-Gg~Mo2Tvk>Vw<0t!#~GTYp8opv
z>$npWCnr8Y$&4491o2SdXR0kDzJI4m?p5!Ope(j-Q&CXB2A4nR>(lN!Iyy?q%SXy;
zczb*68X0BB53e5_g!lBk?mfXrfnQl&RZv#`5g)j@yPH4h4+a?!WWYmt^ym>4KmSe8
z;ahW2QPIexq}z@l6R_S4Oibdsy1IpDfQ`5-Lx+Zh)N;JqZ7PtRcA*k`$cP<94dSV&
zy(-H^49M1EbMxRmh^-oKfA>Kr+H|!+nk4K=;d*h<`IS+1X=5V`S^2R<C<KzkXD?NJ
zpp-4!3f5G|^R9BCG(DZhz`$V3t>JKYHv|k)pjHaIIh5Lcb+XB3F~`HCR`R{8MMG0F
z6im6{_fJM1&wBxtV(miSZ^_B%+}zwr0?s%iBO?iX=H}+hD=XfIO*cg;hu<cK)5P@j
z^wOLkespwpLT?9$h~f4@>*dRrajbf<K(sH+1swK>peqXM0A)6XAB8nm*Lmm!1)_<{
zh#MFHI;IIFC1vBCy1M#KJcm&pOpeiDQm$=5Nr}8U7+AZ}rE<X;lTt{zzP5IIZ2R!A
z`26Bxl5rP2ZEaV30H%HT{$jI^*4d;kXgxDAq2j(b#n$Kolqa{Uq-1y&Q)C}BT_`gQ
z=}1PsurM{vcQgmJRYDGXaU@=KLe0TLslvqw--SHyZ*gZcRZtIcAu_YjC$}wnJ=Y9a
zh`laVX45pz4b&GZi-)G;=K%0{<mn2MOjnne#@=#g#R=XKL?u7GW8N2{=Ygl-ThGU7
zPa!JF3#d$hEyH7Lc!wDb8LO(QirzFZG!$eL&l?vdvXq{3|HYZt9Kh+Pg^G&W_+X}N
z1g7R-YioMiIDT>SzH(k%9386kV^7S?%}ZU*Oa=%cmp3=7*{L4UnnJs~#m)YSiKKjX
zTU6?)fM)_(8cgPg_(q(<Y=yGm2e3{}O;J%%K>{$yf8{==riRh)4lh)?832*Cz{bV~
zU9B)LFEM0kd0An#cC{-kfAq|BWp7V&Rh-lM!+V)SYAUo>3@j|&-(UKqIFWmLdIH@s
zzMmU*%|E>O8kkf1$Cd+k_kwjo2%ZCtZo{R2VrXFSTF6-@O@3)TRtEtxfC9(G#Z^*S
z+3AN$_+pci&kpJS{vP6MSdD{&1L}|DH3bWR=}~iW5i~V5@qH=k*nMMdZ|}a+Nb0(;
zn;T>TypnxFoQk?SvIr@==?4o7=GtZ;gH&J^3i`)sIO0!2fFGd&z6U59kM)YL0Wr|i
z0-fgjpfkIJIrZ}W!@~$)-$w<SwK#@`hOb}07Ew_lfPA-HY=Zbk@_*<+(yXz9v9Pe<
z(|gX(Pr9(M0Pz9}_+g?zJ%M5#aCrLT*fN&6&BWi)=M{o6AXucN8reagY;yjr0UKC@
z*97Ezb|Q9mFX*J>pr@W*N%+{<*h$6$bRt$rzDl7e@EgBINA-YY)!VDZ0x*s)F1klD
zB-y~tpFDZ8CD_D6Ljw!3y1PE3_c+zjdG+d*#QyyLa2g!&qgc4O8OYbbyCvK>M^~B8
z!a}6pzD-q=%~X7IVA%sz5-IU9B5<-4FGEH`@`{MS1Rm*hV#fXMY*Il@O|95owUf37
z+}`?Nju(xTOM3His#M?l_u3mge0-?m1g6K=jf;Z=w1Aq}&J@*0vEzDlbkvl|_Gqcq
zWs;Hh#S6Uk_4UAri0)jaY<X?%aPCF7@`;nA<YcJHbEnkSf#n^D#QMH`F%kJkEKqM%
z0E{Q>_iq~}^)hjz=F5@W^j9HIhHxpp<OQGH00o|_wFN^?x(XZs-vK5XxIbGBShT@F
z&1osE6jW6FJ3|O3Zx*1pF*OYV|Gb?b1q?)v?R+-iPnK3!eZWfU?ax$#)hxC>JZ<ws
zjf#wvU(E>6*9UxjwZMx-Po47Ow2O&h=%mD~iQuq%e@;&m16^&)dzX(zM3i&oG~3lB
z!O_1Aq`|jpw7RkamfPp<=3*C~&(hM;*Ut}T`1~{&m=^>n6;xJd*RU-WpGHe`TVWr;
z!+)Z4`1<YJhvSt_s7S$}JMNxAK|x37=l;dTEWjdO;I>9bM?(jFu2Dr~#&q=K%@<a^
zwnq@alIYuYTL7D*_3p%T={Qyx;oBoPe@sdY2q=bX0rSBIoyPN3chLondm1@Up<>px
zDR!3-&nnIekP4&_a?1gpEmE`i91LOtj5E2?`R?xSY$6yRpMXG0=!fW&qLXK5OGuUP
z)i&R0QdaF$<`r8q!cx`BTAgtPVS2V4nxX<p<d>I~y*N|RQayiTViJ&(b)isL<eaQE
za({PYKOs&qtFCm_HK*a;R17U}BoTv#*fI%4(BkaC1f>700`Who%>M5QVF392?`t6c
z|2_j{%Ag=GQEE3o|DW?DG<ooUxk&#kasN*il34K#^3id(*OxDU=s;*>40|k3qlD;R
z;FXm9`sQd8odT)W^Rbv$Q4taGvTGFgov`8df`+0=Lg>=*G~@y-y?u{Q|H%c=#A?DI
zg8)KKrlPA^+oNS72oE1DKE8-W(GfhvbONb3Ev;pyvaL!HIW8XA5C_NWmm(GCxYLEi
zY0L*GP#=}npHhfkM@=na%?7AJs>QJP_igucn$&Z_>>QOZ9sdx6cvzl@lYV>!rBO6B
ziN+_9E%mY3@6yna|0$+t_eCeIXdPSBnI%Pq6cvfMH4NW8J3mShLPsNpoomQ_ZuBX4
z9#jukV~CLc_$Mn!Mm+itH^~4gmibe4ci0!^la0ig{SZRGe^7$3yUr%Y#27K$r|R(I
zgtU<Q#ZUi|CH&>z=gps=d|_6yPqKLBceJyEBAu8uU;S^Y-|xBqUG+Q7-}3kWx%wRy
zjk4%DHy{MvI|4$GRZ$KWU5>LW#}Uq(KkqSM_sv^1$HmR;Q%*Lc^M4-&STaSbnXt0S
zI4Xu%L%Hu$K@od+NQgas)4P8VfoT3RfeeuDu0K^{Qp7OZ{{G?My4I7hOTsR$7N5qo
z(!ouWhYi(~C~C0u`(ZM#hsiAy<Da5QV#N}jTo8Icql&S$!RgvkAy3t)B%Fely^A@q
z0!{=!Tj7qw*}?}JZ_&aH7g|8Rum|P$TJNZzh&wu$^h+-#JKFh}*ux<}ZZ1r_=Tr*i
z1hh2^(3{{8kT63So??rZ@{|_tup){e)$#JyeOU?&{6pVCSlxLhKyH3zV@`yRP6R2?
zVR<x}_j&AJ>fd*_7<@k+=Ne1VMn=<|qA-4)p)Yv)wM_^?i{{PWH!6TPKKvh*zQ30J
zQ~4e!P4Z_=$S4G`DDGzSzLv;{OZ;KiJqe)nj9H+g@%dBW9o;^d{A*U+K>iPL#^yD&
z@ckcOnV9@*;X93w(*LM@H(~pa3*TQEq(eDbDFoBgE%6*8&b4#T)BI~ZbuRDm@sK>j
zgJDI*{PsCbE2t+ID2{%^@;St0Cy()f%J+*QaA^59d<gBGP+!ACjW=IU@A$R{i_adX
z&0P2Ufrm$`%-hYuHc#U`&(z#rDC3CJXz!SV+<yPKZxBII(MvKSAxq0ZO>U=^UL8tx
z<5We&v|*wj8O^(%+Ou9=GnYdJc}MXv&DvTRV3#8!E&cI;ffKmHNQZ!gf}=9m(-T?G
z`5gLU4F%hZe<^zp%t-_6pUm|??4M}R2xk(!_QE85FqKeSP+-j*5t}pol7s19TImav
z+-DNr(feE}t7jI2#w4lti=XPTQ9=S)02^RQNdBYneb3`h;k!8c8Q}kTqpW^2d*3$?
z0fp~ae+u7m1n2+of7A|CDvbY!La_Z$>@e&8qR%n^LJ1-x7ZtY$745^L5<#&pGc*Qy
z!hC9RpTdKhRXrGLf+$ffPcBeKl)%}yWg&}EduJLwZQP?McN>fZ^CY=3G;`Nc*I3x-
z^k5pF>l^t<9X|g+P>w^h&^`LA%!{aaPt<IdJGee0xh|wbkJo&z#BV3|JG`3ihTDq>
zW3-q?^4@`(?ziIhN0-O?F>_B}wVnxRN;ikfZAf=n6*t*%=XZXXGXDWMJeRU50c<|=
ztUvo?4mPkxl=eW@g-qzn=R49^nx%HRCC*{e1s2S2;7~h;^mKD*McM=<+VNFcfGQnY
zeHb3W8L!99P%}2qhwpPT)HJvyhfdGRP%JX)li?)LNCAf^um=*JED&Dh=@L4A@<8XP
z^!|y*#KP;pyqhRcb!|X1M{a&J^jnJO)Gns}ski}ExFGp)dj$8og+xea$?WSZzs{1<
zJqAaPz#j1_H3#umhP2=T*P&dSz_JiplwgbA`A~uoy-CQw6p;G}5B>fi^%%Ut;2&>n
z8F&xMrq%ddz2^p~(zEf}neQIX4+NQhnjDf|PX2a3_fRD2aXeJUpw40VKt&Uro}L!D
zDUQeXe54!B;NubuqnHx*a7?gI7DKKVPS)~g1-ut{n1RC*ZyqUQqz_f1Bbw>BUVLow
zl(ubW<%VT<YVW_W)zX&U)xLP&?AqY(QKXItolA13LG0`9Ff04c$~n#4cB*u)E>|<k
zxt{t0NB1xHdS!2aw1Y!_UVeTihEL#39LXzuib;mkBFkTL<e3;=Td!)qC5F^CHT4Yi
zwx-nZJfWx0C}<#qu+I0P+HUlJ&J7Bh5j@Ahm?9?)IoTXq)m6j+Lq#Qid8|3q-!J}i
zBqK6a02#vi;)QQOeZ8@U$NhcWi%&Rwjyt|NaYS=9)?!<;ivWr*aa<HiqvTLhD$p#2
zWoBk3+2`Tq)ypa8aOmtpaZij-Xj@#Qq&PS*v1c(bgn?{aU!$zsr+&_P{(6BhXsrE4
zUOjCt;Zx15K|?(Ql9r6NYQAblPTK-`8jlA924r}k0~S^IJ!W^BrAdYJ&4hgtPXi&S
zDJH<lU$*$dW}U!dwwE-vv(Qr?YDqJ(r1COl_+ZTO_Gltz+T;ureuN(4e(-f^pQqEm
zPNLl!dFw8>hCDc5a#9~s<FHPKu^)nyVmdeG$mZHAMbs0Tc-l0^kay(X=5QP5Aa0zY
zs>9zPXp1co8lQelaS*N<k)i&TWPhx=YNI{Ar#1$#8%mAK<mK!Iopql4F5@V5GsinW
zK*C>WA4Fhx8c_Ex#dRM`dC8gTGGFy5tZ7%rw7@cbq9sb}7o(1(#L7aWYhZY|goP|d
zc-9XQVBAI%ObmDlkJ0I6lSJ(8ITgdn6yD+}t=kmbui>JRyg)J<9R>37xxBmat3p9V
z<(!d{lao}IMr#ulgI=F6Eq(p951Sa@@B4WENTsQ2QeRKY^r;;WKmY0-rjUs9D?bVT
zi-UPF5nU=eP(3=AM7FlE)&j&sDkfS&&n%lmh))~@MuB*NAEH|6Rco_xGCcxcybV(&
zmW)+s9XuSdAxGPamX$7?fdBi@k6CptX7wT;2RWCUb<UjK8*X9blJEAsa2M(Kg}4#i
zW3$=kggX}%2NC!obgOf@?V|HH_>QF(S7f2dFHj}RpIL*xeL`U(nqvr+4$WcHzBxzl
z;y68P-@CrNPpNU=t?RPw?DSpA&fh$C7N-gS_>P8Y0|_eCqoLb`Ke8(MI4zBLmlQBg
z&F?q0kW!1Ju$+&D-;|eYS?X^;i-SdA`v?Mlh#S{8arS*omn`<Uq<adzbt&DghC#}k
zPro@Tepu3&e<*wFl37rImha>38>Om*i(7z!;@v-t<$AF{nv?qVYi0xU7i5ro#aF|}
z^9|db)KAFHwi+1gc$AjN)mo(^Y!(tYZ2Z>a_Y)uszb3?~S^c_LLJT_E)~+suQ2{|1
z2zxwqY+YmHz>10&VGiIcy}peBlZ+-{8kf4gIA8!QnVF!Obg)9!tjw?v<#x%eGc)6B
z_LuPf-^*6Q-{UyvQ(;=H8X-MSKw#JJSFb>smu8Bz{vG+gp44YUe{Sr8)oA3Ip52tl
zN3<cwTeisuTyei*(XezB$Wfm#<&t%ns5#zb*C28by(hs3#m-u=0(yAr;|pG`MLd`L
zsw6`y%8>zIQ@fGwW~X>tEwAX-G=H|-1=oi)al1PD!phSH*h_{>hQ@nyE-$WJrq!q<
zpCs5mh0~`WXwy*oh?-Hu<sf@WWsGXWY)Ey}wym<8_<A;OFIqdEN6%X#UYgFQzMg;0
zrnZaof^QU7`PZ*sx*tvQ*g8czjW(n;dw%|U%@@GOZz^p`#AQYV(cOz<BiWxi-wp8g
zh7GeJB_%yQnAgZ~3r!@2Q0=a~8mS4KNjTY-92w{k)qnr_b3YLfDrXkr*uxG51ObgA
z_nQFJDf|rPC*oId%P#|_cvK_tIdO8ld0n2i8NG9QS|kgb9@232jnG4yKV9K~ZJj)<
zY=H~ZP2ui!kqP4kR2wnlG=9|I^woCXk4(1D8gY2vW*D<^*hkTLod#TkrjTdeyBoS+
zbaeGURIe`%S>9V%8qxPJ7v|e;4w4wgvgkky3w_+xPPhJ_)~+%v%C1`vl1eBbAt4G1
zND0y*AP(h7cS+YENF$vBN(o4bf(+dt3?(qMlr#*|A)V6A*~9yNf6jHTbIy-%e$Dml
z{nU=N*S^<%uLldlT3zTvQWG1yc*&~K5(9`Dxw-wUw5#hJx?0S!m{D@LHT|Z90a1~$
zyeQH4YX~9lLXY?3Lp*9k;k-<HJBl6hjb~b`Rnj8dj+wGbZ-*Ds(2bz-*he<Y?zYXg
z%+?{)$UfLzJeN~n^URFJf>?!rKiH6vOSgN@Gn@@yyS%9L40Hk2W&3;G72sR$d#g6+
znTa_p#Ucfougq}vD#?hw7n3UlE-qy9@Kmkyemz@qSc675&0BFq#j@6W%|B>`5@NG9
z?Mi;Eq!n3*osjcTB7?Nj?F}>yF6eQi04G8DIf`9f!7@z+Az~uhS5S0r*|Yq8*boh7
z5U8-;MqJ+t#E<YT991_7>nvXZz%lVFc6N5pMS>?PdIKGj%<@foIzITWk;uW1L}qPo
zZ8~4q@|z35Z)*r8kwwk;x!tNYXP^tIe_zC0+-zKuzifxC>aMIalDC0Y7ZfW@M}oc(
zP&B<)r|Jv$e*vH@fWCYl3XY8>@9d-j64<3)agNI`J>P8eY`a!Ljv!X{$FfT7`I;*U
za^oJlhqC6!^o13<pT+z7axB5DiKB4at}$omU(0)ENfKz_Z~`TREQo2?{ygT$t8TnJ
zoAx9gsn+-ymg2o0bm@~>W*q3^>s=sBpU0E8G5mSJjU_M`rQ)GE@Ny5q{rh2EOk!e=
zN|rL~y@W1ls`u|fHizdr?!v0ll;v2GUQ*YxJl*+)%n_MfX6EnhY+czniGzcKiF;to
za2q<j9|&Lo+Pu(UiDwpaBmXKfPE9My;>=sxkalF%zs90AhnWV62w?ecbeu9uuhCG`
zV1ZdcbDsuiQVoS7`jlF!VvsFWxmzY%2g?06f;(LEV67ZutVMfj6QWo1RMcmW(w_EG
zG{0~5<nOebu8Gvcb+>oE%lT;^nC~W9G#}-0?%abInpm)SSy)(%l2fs~<Okt^7874z
zYHFIfjYpr^>sk{(CmN~A5eNha-P7Qray7tI=+wyn=zCscuMD+G++28POG?__-Hr3@
z<p>I<*mgop%U2<<7QsaY#5OMP$3B3$Wel7FFY;0uxq(fsiB{kVEDxg<O$CFqbm_m7
zgSB5~r(VuD=+LLtVAlSfdFqHAv6YPkj{LMwW<}fx$T|}Zj8c(l_BCZLj^v9?)dD?n
z+4QMJBVVDY?YBijspba(am%k+pD$;J1!bUQRNgsdlbBBK2q&&xz8?z^93~;2h8C!J
zE`s&K1Q=DmF~PEBTu7DYX;z=B0T>9;aSn7C52VN(H0(WJrSRmFqOqRa%@-3$<E=>L
zw9SaxuF;cYRdC#{3^OEqK=1F=qa+SmS{uO<udby<n4NNbddzS0V5E7HzPRpyh771>
z*;onuc5d@pT3Tk%2@k;2Yao4fo^CI9_gX4K32HAJ@gbkaqrJ};5@^QLI?SHj=P^VX
zV|Tw)h~8<wSx8B%;#7Dhh6j4TbbmK9r=Thj_#olbytr!>;+JQmm}VWf6DDbuoo}w*
zY+5nI#``gs|7MWypb)^XH0#rMB4;2AuEupIWB1>7i6evTdJq~VC9CP0y~ZW%*8-WO
zmF#UE4L4QCKKeUK&1Wd|Y(vMp=*&xW&P0mackI@zu*EOKNG*{&Ge-JBGN<e#(oAe#
zaV4x<$y?utR&x}U3H8dT{dUZRi-ln`$B`)AIxvY%6^<&8m`#W+JXo^>iQSUJ!746&
zNMZM48ag&K3y7UcnIwp(Yr=5K7!<)L5h6Hraz0?3ci5qEt1e$ySE|Nz^}D@S^$H<y
zj-O_Va8+?~wh!Vm6r2#W-ubH<l9nJ8aV3p(ri?Y<JUZ`SPz{lUib(CB{^^KiA(NGY
zjy3g`l_#!OC>yf#e)_Zv1d#)gF*~Q+>l1g(DL%8DkgA-jYpO&}#Vd$Diz<(Da@DhP
zITC!kGme_&?aBPx2}FvczU#7orr4errwC9BE^_8J-)H0$Vw1oNM2SBe?Zj0LK_m!$
zAD6<~wVdgW^U5$YgVEn0MLIpY^IHz(NUvfe(UQ3ile_#|Fm#_(g)sB-Ev1%Yk<r4n
zH@VOI^&aDcAB-OAdC=*5iEA-=5xE?c{`Rc+a;ErlnDj}Y;@g!>64ekH=tjeq(uJgo
z&O_DSrMZ@4{eshWs=y6fqp|%rqu$$|X9we`{?hqN|7VuiwR>bgSwhDGS(H+~mDdvK
zBs{*iJLit`*!$*zfY_ue+n2U)4{zq@k=8R|9lxtoQPyX35rPgy?v?Td7P{Ol(4Gn8
zYD5047ca<@85P9ej!Ld{C_8GU*MdPgEp4r_^lS(V^k)hWs<Wt~I<7P@D;z#7G61H5
ziiG|90Q1$BIHN4_N{bcU1H7;QWCazc2Q1k1BCMA<crWOQN%;f*U-yj`h>OEo8^Nlm
zKob5IafayYe_Aj%pF70CySBK9+_$xDrE0Jp&zh>J7M|JQoZ=!)OCQ1lO0<>ma=r$%
zme^l0-&e;=GtYo1t)aW%V5~c-(+1L4%%E&QtdA->!-gaavisuUUDK;zHB3&z3J&fR
z!6$(%EJ#{dtUaYw5p9jeBu%tE-fkX)3f=?iZQK-bT#)F`qFXuM5LNWjPQ%d=PCkjW
z@jvrg@W#SIb`*3r26m?2z1t~reB9tRom(jlg$rStGJ|-$VwgFL*J3EL?V4X8_Q%8=
zI&Y1L?W{}?KWf+*NAAB=urYiyRWOC8oGh1|jE;BrU?Yb!lKtx$s@|3!AQKYaezhKi
z8yZ%9JNun3Yy11Vd8d0FaYSEUmiYQ87KeV$%f$D~NmqSn{oLJs!$wKzuZYXiil!F3
zuZR$Yka8&A```{x?>X)Xa4UOXys@|6Sfdy9<1vLKy{wjuP@@io@?a8?#*7_obXHgL
zTCgsyT<d(`)<ro}SP5S{K4DwB<kD777ShsUr^<<-X$?w{!->6Wr5w>|0b!k5#mQES
zybR6OK{erGtu%<@SJlnVZ#CI=o+Kolf2zvWy=V1Y_-`>x?=0OL;CYFmHIy5-e;h;0
z%g?zerWANTHI{0&t*(CLW?>Z$pQR<`55)|XXaTPy;|;5If9%y=TPMd)gQV(u;lC2o
zUaDi6Kh@J)pqs4Oby|zO)@3I7x?soX0}1^QRo1Y#fp2+YXy5FL+NPEi(2SXNbRd7Q
z;o#)Fxfvge-JO)4k~1Rsni@oYM3*Y;JoDBvGm-DNN{l<389O^a$(O>7^(L)J{GRmx
z>eVvDk%iJ>nr46oAAe!R$_72coX*x9jHOq}Mvktzo|~QR=<qot<caZ9-&aY%OU{0o
zrIC$DO2S%M=~OZ1EgK^r>F)ky{?hq9)WPg3_pCFe>=ms;&ho0w|A7}k#>yMMXZKxj
zpdS91H}4PIoAGx%ElDuctY=1O0sta3bXof3*zzwLI>dI>)pc?|+`>?xu<@V58ZVtQ
zkT~$h)tr9(!kJH58A3#tbEc;Jje`lH*?g6o^i|Ndy;19pg|$HZIEic}tqYK!ApUUz
zGZ!QMPjH!>!mcjtCr|wH*m9ZDmK-kYceb_`6&8F$La_b)JJEp6##}uz;+N0fhs@`5
zQ0=0oWxu>wU(bjPS_Lo&bV#%HCwnZS^96_;+b)kNBu$(g<s|HG=~Oyz6`kQLoTDsb
zh`_i(RD?45)THM!`MA4FdWdW<_lLa?D__%LI(qO3^Z2$3s$@%!WJs<3B@0h?dD@$p
zmV!Ic(fj9X`RA}2{WZx=zmHh_5Q%qo_DR&v%o0W)F=dK*fcW*_*xg?V_cjP8s_wh+
zLHBB3zy74xwJ}9pdzdoRzs$c0B*;VYXGiHf*|D+Z4!2aga>SeSdimsoL9?sqA#^9^
z*4S<7phHibme_pr#0MYT0iG-)9CmRaIFov*UbODB!H>E75`Z2Z@hP%&V%At)AW*1!
z2nw~sLT-RK*g&HFSd=m%0|3B)fsY1^zg+2_{|CEUC8+?rtEpkS+k@SX(g_HdQ?+N~
zBjuL{(f7We14M83N~Idk?kt<UX4Uu9)`;K~y?O6a^Sa%{9MB~mAV4&8d3=!WypWTW
zTj|_M;qhIn9P7sSWeFt~HXLn4rOkfg#mxid>>~Bu&zu<H6PTq*{@DA^?TKe&t*Ipk
z&DGR`a010IbwRxq8*Og|@rmvXk=GU0q?!Iq$RtS!{XBCxo%*e$ij1oG_N?}pM!|U>
z)td-94i{~Yuns%Irv-VzKPMIv#Rf>p4FJ*=5;|7;Df~5EN4tzv3(XP7DIM>WYujE&
zk7|fCl+%7&j*`SH^&NSsG2Lt&LF^>ujT@JEzGnw-j>4UyMQ$?=V4X%;GFEPLeZk*v
z0fr89!E&RedfuZA!^ii{j<11*7<N`v_ipD+d_!q!=8*BN=t%0JUNV$eUYm0Elxm0+
zRQHLq@|!=?HJmQgSC2?>JE3WyJpZa{3E##CwpP=<idAM8R%pgU@S?Lavj(}%eR?ll
zk$z$Qp>lfaD5Wau2syXyT@FqQYs^yk&)TT+gN(kEyzJ?cJbkcc_raHJ-d?F8P0Xl7
zxSIbRI37)|g{^Xj=6p2*VBow=Rn45PT@kn|QuNe~Wpy|w|5I=Nn=Sr2ZD9JG6M>%8
z#RZ8ehEXiUsfKJz$sgCA9K0F~1M_^Z34n~FqCyY|a8r%vGUQ9#?EUHjv4Jc9trj6Q
zf7cTqU%%^DD>^!wdUlm<qBtfg>3v+j=Jkzmdl%ahP6=P%jLrq%*$4HCZsqJ3Dzko5
zW*s_760#<H!pkc?{+paC1f+pM*0U&n4pOhyPv2};Nryn3yuIJmKvh+KA?u<qd6~m^
zf~uqW5lpT((6Fsfv-5sNdzI8UPE0mL4BRps9_OyBHmm0ld!|dvqf@TH>S(00+vO${
z(hwJ>Wo%~_QC!(Fv%c~&P4T?KM^2nsECgKa%sXLL=AsvI&Y!(m_C?rTlXLTeGH}~L
zlKaKhC+i`KFr9T>wG{M3JB=3CtJb7g%i$E$iX}1q>bDw&wi=%MU9Cm@`*qCl>zG}B
zePnn`o#){#din@}p3y1D`%9(y^%m#=0vgS}(ib8B)m1p~-UC2knwSRH&s8PyM!-q2
z7kElvEi4=gh(A{OfGCt+xVPsFZEd;7!FL-R$3U>^)pXG+Dg?T9G=TcmS{+XGML{W~
ztc-W+at6D<<rrXv1O!AaE!W9E0;CUI9#CscP2~tcXI<Y%Ffx^j3Q$5yL<KZ+U=NTK
zF!46a`}ry-%gmIIzuuHg%87LQdIKU9??8|lWj7bLtRUoigd<KuU|0TYal+6en~0aW
zJGjt5;!m8u)Jl4?HtosvfOzj<hL|YhzK{N_-p=Gq)6R|Uf(FVb%UQYwkXr}GnqG9b
z!lMGjB|-gm_G%oxA3ljZ#uYMTdMG<^$ev*<_+44Tx77q$1)piq8Nr!^*La+djP7rm
z-U->xdEx)JNG+A!9j`BA3D<8=ZE0J&J%Qj99=P58Tvn)guvwO`bl74V$_ZAPy%#z<
zcQo;F8v)z~F@l=}3_pAJ@eYm2R=qdPKHvm=LYqHHE2u?W%MS=BaX9w?Tc+DZNuLw%
zIS(>p03@*Eq_OQ#N`Y!y<@6mwJUw3??ld)KJOGp<qtDqa5J~1(V;ALHM8w3=@$p&x
zJ)r3I@7g<DTu51U&#E8ufe&FNoQzU@)+})Lo|K-JfJ1K{t9D7`&W7Pez<GmUi7|}3
zo3k6>MrkF*0!CX!K{Djkixb6pKi;l*QKR5E5bCF{52kKv;~ew+>ectobnD#SkEd>7
zZZp~&jfOzxwIWJD&2Z|CfcD1~5@lSFmpgP(j}d~i**3=!2ifqYb9aXOCO<8@Kd*7N
z*xDHfw~rQ<FdX4Yi4(q2)$3(n=yO<Cn8KwU7l*)8-RFiJv}wyr0W$T^m<3SpHx=gC
zj+xC(68Ac52n75oQi+d`#{-}}(yWXtHHp{ZK0AzQg$}?`4{Cs$L5vLvr{|SPryut@
zS5{X3rebGm8aP$Q9~DDCzlaBEuX(G&$AzdcIm$7nG52(P*kBs4xYW>AadV2=D6A}-
zKYL$n>l|~u%Q8=9k|LGsx!L-N%)Epvlq4xSO{R$ik}{?C(Q&y-cM|QH7&b!BXLj$^
zmhqI8ugiFDGB=vvZI<rK)33h^qURPr3%PvxXog-`d5SAx&_(t2#j_zw44yK8^7ZX^
z^b{5S6qm64w4D$FAxy@6Bf#d98z57quLIf&kO$h7+zh+>6aslzIo$iq4LH!-cT4|J
zK;UpVUuUq0XxK<xhgnH!X(M3L(5{E%aOR|P=gJrW)OXnEoN(zo&Ff;}!~9iOUiS;r
zsomY8W}g{a1e%B_4#lf(xMnDnfJ@jFOg?65ke^@b>Tytk1G!B>DVSTgG%KX-)pH4v
z6@@eeo?ec_y~&YK*<9><9?`3cvN`#&a;m4bt5H+%WbJE3CPNLoq)2lRucLs>tjrD?
zeT{ze!8dkb;#9*0cm^t*99eL7bu~fVShnb5rPt*NIh?@j@sLha4rK-@?Zo<Y8U6=A
zZBG}C!iL1C%}b*;(TnF_Mcm*%e*1}fiN=4DxolUB!l-!Oeaq54bGw=SSXu9<HBpAv
zsuHL2BQ~~xhzNz)F?&ZxDIJ|=)Wp`=?~sMm!8_#}Agvr?0i-T4UszbSHXmj?Cqq&?
z5%RW&6RGI@^-{C{?1X|l4DTLf!siv`tz%b7=x5KmYsTQ&IVCudzB0FsB|{4{R^3DQ
zrTr<6s^N_K)cu<ug2^lpBQ2O&ZeoIvNdop%pNq=i0dw20U&P=>G^b;EEi4a=Y7(>p
z)tGfxr8MK0M=Y`dcx+`gI#X+gsCXT-^!7n316dt&MS=}%GZl;%tI3ky%BE>Io%B)?
z27V&9HwhowGRE+=%)zS@kJLH-W!s#x)BP^$^IddvimUZz$4_YS*3<>3MQ?JhKxPA`
zlbybjANrs(#Mi>6oh12G6cxq7qu<mJc;{a{i}qgmZ1n~D_<zB|)Bx_7`~JGXyC5TA
zq*WIjjnTNn&u$*0i<_r=i-%K;L3E%(8cGnN2*|cx317nZX$7oa$7azOfCw7pDj8Cm
zFroYT=5Xo|i5zR48lLa8dCti~^jN4lZ0qpS>OL?-Q0ErK`|hGnKn}C{Atc*vW6S^I
zprT4mm2F7noUf9*`4lT`X049EQZa>MHUFT|g<*^N)9=xP#&vgI={2A7w-mr44P_ku
zoM<94+&a0&fR8gD^2bz)Bd?jqvG^jWAxc91hb;1W;A#nxq&vq}X)_CSNm6jHmW@0+
zvYEx-=AE=p-gB{@uN~07u?4ux^iA!coWk*U4^Y77cVbJd+e)Vvm)5`61-tK+xWn^v
zEQolSMdMT>)-Dx?pDoHKG&214h+IL)4^hSn+vgzLrc5LaF7PsNaPXv<WA87tM{Cj!
z<NATL%k#~dFw)eI5KtIS$A2TF{_?J&@9{&iRoggV*Ge8ez%lgMr&y<_+wNKaT}Vl#
z0t3T*Ry)|S+@*F+J&{(&bY#4q`$@d<o1G#Vs%QL5KEPDU{Spf}t6eK8qay<Tg%|10
zLq^xvn&V?@zGq8KO#}KlyQU}Zvtu8W5{#SUjX-&({i>Gg*?s8xXt;dJ{IJP*$wO+N
zF3P#GxV3jRJEqp7m);yiM8u!Z?kWKDJ=#Lu9+i2Whx1Dvj!fH=FI~c!fA2=z6(zf`
zhh|*tGH9{}Agr&QebL{t6e@^J#2qZ9tw-8Ip4Yo03T92Vf@L*s&<WLOQU59&A6a)7
z-`;<Paf4oZ3=(E7W@-P;J!*G1#Tb>4Y^7B3IFZnUf~6cCxeG?+Rqh;8w4)Ob8eKV=
zdT()Q>m~~eOYPz258$X4S^>b!jL!OOh})Y2c$QQ>=Ez)JlbW81_LFL=TQM<fe~fNh
zw!r9vU(L1u3%p_ep}Tz=E0ICsAL<;mdf2grl86?HsmKMr96IeZ^KP9Tmce-@2E(}K
z`^YEH<%N|B7#TNMAE9oV>aYMM>kIl9$>iW>uJM6bkg`P&7#*?7AO)pXQ<m1CVi(iK
z{Fg~eUHS{+lq;Xy=E#|W5s@{9fz>GDOxnMtBwqGdUBPNc7F*V5C|^|g$A<q4SBq}!
zvS-|&A-UT>!4wxdmJlfW{iBBnmH+OMZIJpapdw$4nWlWZfV#X^X6GyG+@vCQf%e@-
zU#=1)J!7feQ^n7H&V+ECIb`wumQz5MyX53B=fySR{U@yPK!AJWz|E_U1viS}65sb6
zCQD80(gb_#;1!Vb^yvJM<>+|iDV_YUw%K{DTeZ$y>nA(Tuv~kZz<4o$m>osp{GpnX
zdFC*vc%|GBkGCj6Zu(OYdiAwd9fYDoG=@iOc}b5e?dHKA;X2a;#NziAa}qrp7*5Zw
z=BfT*7-qW+e$JFA@+-muh9vOMfh%ByaX{w^<_P_7+kDI!|B(s^=I!lmNw87myQn=`
z%?AY1wu6oFzuU9Tcw2xsaJvp=*$dJe>^^_>>Qy?=17_x?@p0PX`jfo2^Q5GtF)w`p
zJoNty+YJC4|6Oo35UQc%(EjGY2|hsrSk&AC0&Vr)r`&+%A0Yxxq89fOnwXsA@z}FU
z(@#^rCn7?lp`iiz(SDVcm2GwI?(W-rdr$`l9*lNkVIkA?(jOaN8At=}@LJq^@`TKw
z#<_n$jhqzei)otA4b9EikdGjNh(}#U1`e-9x1_6Uh1Edb3)kn*pO>1n-ay&_`hE})
zdVDV{Q_<Hag27<3y02c{m5HFig&fU>FwD%&F^>6aYg20$>0ftqJH#`6z&IfL1u&a4
zaDiVXr=ajIJ)P<7><ma|2mW~ea9#-jDg%(DePb-s)pnsX4&t}Bw}(e=0QdtYxF#kh
zASOggNqIo01g?;bL0PQj6SxdYT-G$zH8iXr6h^2-ECOMH25T4CU+HsU_`5*LfwrP0
zq~Z#Si=z;*9UH1`cpvIajp<`fM+XQ0^z?LY)p0UF$A<U;;R*k7%KjQUJumMA;Yeuc
z1z2xG<+|D;>BHvc<}4=D$2!tBy(VK~$b>{h1Z*00rt7`2AXHrXR+^VkXw3(JQAEds
zm;=OWWH}E$78n&3rOcerm0VXQrIsani*zM6QWYOOd+*N9PKo!aD^yn26i|(OK_~#|
zTdokEIOnY?gY?(n<(o)ApvP9;;`QqZcqxES2{F)%uIo|{J!EzFjm9bi8;|p{^4`X{
zmUB%22;Koj&*xz`HANF9#sL;g50<Y3{NCQpjjB!n42cNRiC`i+p<9mCbRd~8;(b|S
z7}euzl7Z*sq${*s%HSkKOAGWi71ey*;gpSo9*8R7e{=6^DFZ#++}ZhoaQNTt0Rv2B
z0tpETZ5<tBz@_K3>+R_w+1%V5t#UAp&0={yT~FWI+NzE6@eVs>W*nfndX{^<yySuS
zH}N2VUnA(*RT2!9mTnv!rB0WLXaW&V+JZ{Z9vr&m1mLNH!Fbg?0jJ_(keMM+Q<{K~
zaFA8$#S1A{SAl|pf{!UFMZ=G9FoQ@*=_jB|^Z1;-1%1-g)dj06Qf9TEX}|)nV!FRN
zJX&HLIC6Edt5Ka(Ri*7b&JEN)EiElZW98PmwEcx$;03gb^q-lSFp@Bchbtrtq~+z^
zfEbyZzbiHU+4S>gi02*|>HFpFJrHbB#mA%zgAUOM{qxn9|B%7JAW=G(44IE$@!X2t
SNbp53h`h8Sv`ES*;J*M|;f&k>

literal 0
HcmV?d00001

diff --git a/docs/zh_cn/reward_model/images/var_len_atten.png b/docs/zh_cn/reward_model/images/var_len_atten.png
new file mode 100644
index 0000000000000000000000000000000000000000..3e60777d2063d925176799f49a1e821a24ff0b2f
GIT binary patch
literal 41318
zcmeEu^;cZq@@0hJG-*80xCXc2lAs|$6WkkjcMZYaf?I$D2=4Cg5`s4F)<6gMxqRNd
zH*dac{)Abpf1qz)J5}e@soJ|Xp^Ea-Sm?y)Po6x%l6f!j>B*Dl2Tz_n^+SCI9C4H{
zc6joH;)#sJJ7u@02Q3d(kWKGxnD3W>9rD+OR<dPrMJ<fKVo3I<b-?ES>GDqglGQb`
z-Na0lxccX_Cbs46vQDJ!@(of*Hzt8e8X82wFMI3l3OCFk)~X;G<g#w;rihI1VPaSf
z;BRL`-#)VXROw8qO~ab+2DWz(cexx#Q~0-=6_nOew)eN)g_oD^PA3<YN=3mhpZ?d!
zPeqbKQ)PNk{lAU@zx!!Z>1Mn|lK8KW>Z_-rmnzfxoNE8qVVboml>a_WLO%f<6QXB~
zMdv8_Usp@KrPB5OzrG{={~P|>9{(>Jsy_pZUUy)NdLMK@oK;B23A-*g-+aHv^;7vk
zG3R*E8ZGj7*N|RH^pdg<v4bM&wGqnZyW^_4w=~IXS-E|PeJ5XZ)}6!x%}Eshx4*uN
z_?b)I*4<y<-iSIa(?6`t<?9Oh?j_VIWHl8(0FJ}qI~#;4w7A{lX59=vo>QkQlC^#H
zcm>K~^Ft`Eo+tc_*XMLSOWSLoR29CabyN?{Q^fcY(;Mw`8$YA#yAGW5@!}z(T#l*^
zHRQ`bpSRx6K8IdTVEbI&V=rE;RFo;r^;15x{L=2<*LwbVs8_V*&QI{jcdLgiIKG5f
zcvq1@X~SJt?s-5BxozkG%QGs9$4@vFdqT*=X$9^BqQdmRf~5|%80*<cU*=VUxO7y;
zjsstX?kZXHR55yqOzzbzSC)v&SGUdoOm=lVzh4r!k7oN5zWv%}<E7sl@fC|c`HSU?
z`F-0BcN>MEfZUJSi89{AWi9u;_ny9Ec?nwoWx8JyUX|N68UA`)l`8lZNb)Uz&qtcD
z?|1q{D>7H>KNT2PG7yViBNS|wyvOOXelOb8<*u*|Rk}0@LqPe%&Ft=Yl5m$McC$cX
zr3-vkr@=M4T6drFB(PbHraeSj>6@{JMmqD5wr%8k{e+p5K*od_Uf$lt4XaNa&4`I3
zJ5fOF>LlESBodi9GzS%NbLXSsbu~ltr8e0yE*>j>;M3Gbp4releNC`!l68ZmJ`1?s
z{o-9y86-1asURbc?t2A<s`ZZTfjAQR$CyI_{7U1Pj-b2th1#re#mI?&ON@H{WN1#o
z<nHOsofv<VvPL4U%wIN7OF_9whOjlh!!w_I2GE)~-&zjdFSlRg@cY|8<W$(m{xmW<
z3?C-HNmgsJxbYB2qIaP)(JFj=OSSYn#<LcQX6;OX6wE+pcXH@|o=O8LK(CSg=T(+c
zkuJ=*EqKzD@ZO}2uscm%y?x79MNiPpniZB66O9kXLwcr>_|se<QOX6*7Qq7nQEp~%
zOE)!HPE42u^q=ca>a&NG-QnBd=N2h1P@&4jMN?CNkG|||W1_w9CdA+?dY?Gx?Y6oT
zmEoDl$h%p!fW)xYC48AWW|Jx6>VEZ`;7~daZO~Z^+BYg3Oc=#lq!M0uG#KJ0TQ-hZ
zMc&4wUmgf!{j(pK8NaEQSXkbcAgMccZ=O0G+izo~X%Mqz579E$Wmr2AA_oO4;nire
zY{eLBK>G0etnA9#b+FkRAcnw;L`(iaWs(A7AC=E0s-i(DI*NoBA0E6R_EnI8l8PCA
zpcs@PGD~IWL)azY?>_qsmQMyL3erVWVN{~FY;_^zky52?BlPL+<%QP_LM$7zKcb!A
z(rKFO(g9Ddqg9otx}ulC>8m)4Z4v>(^K~(vV*}b{kG|3-t>>XGzJ-;1#2@+FbZ!Yb
zE_B*FITBD}0%lE4VX-!9V5Q=$XxF${C*@SCh+!b}2AeDI&|2`l5}tF@;;LOpFt}mN
zB8_6~#HA_2cn)w?9gNQ_LVRHL7cwFJ5a$Aw-K)m1F7^v2>b3E$3dwEvCAiC#GHsNx
zT2_n;ttNF9vn5Yj?gXk8d`hxx#_+Y9(Igye(nR@bC+gN}2x?}}fs;U#{~L-iN0Ful
zE1s*DrnwXP;#^lovnkEXN3d;)Np9!b!0(-TrZ!Vl_j@g^+;Z_W&|7q88%d!<J{4XD
z3TY)}DA`a=3vF1AR4#aFv~UM`I5?_u;@fruEu_!BYR-syjq^7t$KZ*OOq^71$M9<V
zuCGTeOw@Dsi^7F~2>yAKNPM1o*-R1j+FX`k2I$5yHBVoJ2y-@x9od|Pv}?6|&dPyK
zCho_S1^*DZ0S2!`-P#J0$=oi$Fw8S2{)_o}B9jwkl<Bv;P>>RV#{Eb6w}6|^vuW2#
zbkluZ893lk&TckFf#Uoy)o*FWbES)(p&na2j`v?jD4|lx?W+>dWbv|@s?MLxK0DZd
zUWVWKWEp&u`%+1<W=})R<Xtw?6*oYm{$Y|EhzW}}7iD+O0vUW!O^c;N<0?=-$oY-8
zQAhwDw7g-$OwkCBAY+^V8*BKXOH=(jB?gfG+qUJtrx7Mqp$@N0zDsD85ctO+6#(&L
ztt*)?%KCK_mn5Si1!d+(P)(&UdYthqGK7(RfpQ#8EnlV*bkWj4I&xf~e&MS%1Gn7g
z#Am*`4UyjE32Ocfop<{-4KdN-$j(FC!WNnR<e11)8CXA$p>g*$ZMtRB<d#wzyyf;g
ztR!+2!%^KLT@~@hDU<tNK1nHEPg^i}*()a|cp+DuH$&H$rSkmO+Pjwn=O5yJz~I1u
z@FHLLgM6JFU_VcU`uwqE!#P53=#jpDou#_7^nH0+KE4&&IE%^gWX>^0D?`^W`IfNi
z)!yvUMQjO6D#xZe=ot>;BikA_8>D5mH?5q>5w>(kuXjCHbV|9_WXgIVXos4umMeU{
zaq>sq4&r-w(E2=>$7-P)66GvE$JZy2&tp3`3Lzj9F%c*CxHS)2s&#!pY>&>5Gla>v
zI{B=w8dlpa`-v}d5n4eS-cfu@PL$;%?ozbrt%g}+q!Xj2L`YHBUIs;`mDiY`9A!4N
z%LigHdN%uQikx0fE*P706Z>~i3~}BN1$pvt3{?h``OdrKGY=i0+g~qi9;RTpl6pL8
z0b`7raf^vCUz|1!O>5Tg2LuIldF|csDG||aREB-8yv!CgN{Nr->(!rOq;aBmN$(vz
zffRN+^bVLDP{A;}81>iZY}{at5iIjMy%a1PV$Ka9QY?QXr{CzjwCNmtjn=OOMj1Lu
zk+6Axh#!nIoo_tO>0gUx>Bxdsjt(dvn7s7o55sYlcZiyBu|<@_jOA8xrkI7hi8Mpf
zS;V_L@H5WX_K~5P47|;u91svk%Ph@!2HrybP9<2ggy6giKJxUG+o9`Crd(jr3&AHt
z2PjWTCP69$P_F#6yObHX`l&C0Fc1=^g=b@^>4TM~_5=OQz4mKX-Uwp)KKqefP9$_K
za8lTd1pc5dXdF$Vs4MzlT)(e55K!4FLfZDnpQqJX1s{l9{l0>oAGQIWrZ5AsK^B{h
zZ-8)L8pbl1%Mey9C3?Rwheg7>=)Cjvy>CGsLs-lPU{QKKfi~zm9r}e~#%B^98w1d|
zRz3UMdEi71QgX=&YtM_5-_Hh0Myuv6Gnz8GX(H`Jn-2KZhSy6}Edke?jaMyS2^I^u
za4uFjVry;yIS{a8@<lb+P;QTC)!Hsq-u>~tA#bJ~j3m%>exMgIUY4DKyTsGH^IeZG
zmhaA80M2SVHuVkdLGxrR$KnUB@ZPT_P2RNUC6HWys<CxoLaV7VG6;4_VLqo<T?kW_
zM^5O$zOyG38J=Ttd%rWyRXl#Kjl|m3xe(K=DXSKKSs`gyB*IJ#JINOHh6r5ea@B`_
z3K9tRT${At(zpNha)e;DybD?IdyIel%%r|8H*xKYlE76&LMo?+2UwH{vmWDPyhWm}
zbaxWCi;}=Iwzj}VkuTK_`jRIHf1%z}AD<o+z)g(h>5XsU-RakL_@zC6_@47wJw<h5
zJypR>*J^}XNaJN`ynhh<Bm$cPvTH1iOIT5R$``d5$oki5J9c5kpds7=&xU}R5k<08
z0SJ??C2t3kMg^6YgdHivsQP!hFr-T$!Q%~pFV9K&BqkiYBL8vReZcy0=LFy!Xxpv&
z_Ma*5cZ5$xyM6;1E9SuHr4k+@=UFDU@I80!_A6=7%kSOQ+bW&eDqwUS2Vmv*zciQ5
z1s|R3dhKV#vt>L^6kru+@*Ad-MN0KwW%p$*mD|T_EwS%@mT4NH_cA8XF$U6s;S?$|
z{-fCj`Y@^{Wr{2|uU!qIh*l`|hPw~oPkc#mn8~P<^7t2(10DCwc=mXOb-Nt)Vf2v+
z?~9-F#t3m=$O{mc6NYZ`OPS)fs*bHtil)PP#YRfLOI(mv0G7l3;TyV$P_xo#j%u;k
z6-fpjlR(&jswqh!IePB+GSqeM_mh25+H+2hJ{Dg~7r9r~Be^5fPemZ!1<%gy;ccZ|
z^(;@)vK<8@Sn<MA{Cfh$dVV`9Eq=G-PRMvBR`N@O-|=saP}l0qKILK#)yc_a0qsQ$
zI62JJDi=1qD)pCL9@^B@Z~PmQ#y9Qe`qN^hd;%gxG7?+&%c@$kLC2XOK{b{T$lw4U
z35^OV=JBCOG@mTEuDs;!%x-kM4dSq;{c`8r{aWbB^O7-TJ0T#Ls-AxK;n2)JHQn)W
zd268gY;~s)Bx#tvYNaXUs**^+MR8Pm0N7JIe%wO&w_AFwB>OgrP4j@wUXvlapvz%J
z_In?|--!9|_VDPrwc01)wfA;a3#lkps_NZMo_+EXl>*tmbDdkaT><@**J{kR^<N!Q
zBUC3ieumUtk2Q#QbcjJ+%(Swp-WB?^EnD%pw_+I2-J4V#3s6>5(rlF{&FSBD$*R`~
z%<6>7YMGsX3UxcEtYTJvDo(U)%P8a+ZTp@>CB^U{<r3Y1kJWh0N#l<(?eDgibRISP
zF}DLz!DNBO$L$QxEFK`2xy2fkR_!5rwpf;v?nM?DSFG;~zc32bDrR{~Cc*}GkYnwW
zk~Hv2`MdrsA#s$+dF9h$mk6m``Fwv2$}~s5#G(_W9U~w!PWsfnT<a=|-+<3;M<_CD
zA&j9S^81sf^O*K55Xi82m#UACIXvZ&(+;7#<FU#Chf+^D*-kOxdbT2yz7+{-+K0+$
zhdgol7(D@&X~U(w3uC=!M8D@{mJb7u=38c1<uwvX&Ak&h59eEWhv!;9X0i#iZ*->T
zG6?*>E>5#2(sb+1V|NOaZ2m0yKMdS7l^W<^I#Cl6`Q+joRVFs*+`nzWWz8BTx8`kz
z@hcJSu$(<eJ{sYDOmdfvpeUBzAoFkl$F^!Wy^O(J4!U6z%6Wf-@0f)E3dpK{a*0R&
zMTIdA!p-qx+fMXU$8c!YxxRO59v{<J)Pn&&m1wwyi2JKJ<^wT&B*w4(<;?OKmFVm2
zrm%;9nUTF0)O+7{Yp9tPmWdDY(I}p>ST*+%K_QRY4D-U{<^b~EpLhsh8lf3;zlom$
zQg(c*%-`t2yM$ySLHTO%+zLKYC|lreZCH%WJOx+*Z@ME|xPuE~$m0NIknpN!iF}<f
zU|&0BjaQM<sezA1xu#T_ahd|(bxbM|HNtY81MhKn0AJ=NJ3m(q^K%b~;J=)%V|JpI
z6sk^b$%od|ShtS1mA6u6u~Syo+?#pSeGV~0bzm{T^7r%rlW*})=MVGDbi^pkv6h8f
zRp8l}Q}g0kzwomJC9&)q+~s$-{_Qijo9RD~ucXuWX1VR09OKFJ6#0!j=T-caN8T;H
zSJ!qnIUDwdoa#@rUwfnrrfJv>Fq?u2Qb+P?kY*TbPrmb+`okHy-K?_Es*LBu?zNQ&
zxU}=uxs6NAM_i1o6te}rx{GoS1Nr+MKRjkguh-P{=`+R9Eyx<lQFsmpAS5G=7-t^Y
z`EoTPi+Bi|e293IETb6ZS_4aWMOf9=Y(?)h_*@$7$Ey%im8POmdbZu!K5iE*GUkZx
z6r?#2QVp1Ym9@!RFJNSjKqpJ8Uc4=iifn$yM!0eEjTLG7lhrZBVy=FByTWf1>ArIu
zs;Ja3s^JcSJTL-(t(^eeID|(dm?`)5`co`>zcq^>GXs?@{<YviAWwaU8+T4iK(2~&
zx%e;kQ0IU#Bn97eWHx&HvVLaEXO@@?P&HS33hTZP0?mudKYJ7|xQ#S$@$#HUg8A^H
z9U5hi<z0p9$0#hWX=DzHG|^E$x*h!-zQ3Gr+S5QM-Qj}j83&+KYN;pgZ{T0Zuv^-b
zF_BEDs~#AAQYEx>?9txNL6|~q+0ryqK19*dP=MRMQTYII!Y*scW=y<AB`Y*PG3WUa
z5x@5#r!GY1K}<SZm+h2UwzX6!|G}UrE!7<dxs1J&y_QY1HwD+}dFe~_>SO}BmFaAk
zrn^i4#wqDS`&x;{FW$}dM|eAI<1_#hD2mT%F!hT1Hah}29r+5y)ug&lsO98q_=4Dp
zhN6X!v?|K^`PY<REAjb4-mG#)nAJ1EIgLQmN(CRa!4XHd71de^{>dXIRUSvRvhN5O
zao9JyLKvPkb|1zGz&bm=24uJ8Dw~NjtmSurNIEUQbb5jUD}z|Xq*LbZeA<qkSB&M_
zF?ZYdiQ!F!MBldgiK&XiBcJmU_eT<MwWBdgkeV2KYoAb}@&q3-_T=Xsh@kQ;l0*jW
zSMVvVXa14Sg>ATJ1|J!-S%&Y47^2OhcaolKT@og4QmM5W{031)yy4Ap4+*JqnoF}r
zwCcmU150qHqeNulW-ss%!pZ01kO<~JhnE0|_yO#D+fXI{vZ-!TIs9^Ro3$xv>aE(G
zrN1`{Xbu<5_weNch9{D9y%Rrn0%yn{({b6Nd=OJ$>~!8|08(=)J_;R%Jd;vcRCs2C
z9k9XJdu&=pq!75O)r;L8#B*=y+zZ7I%3Ag1sxaK`4WT5T0NG<Q(#NAI+gF@;efq^A
zM@|A>EI8N$`cQEZR+jJ}Ln|{%Xp+ob);Pp%EBe8pqxZtUv$#GTD9t#6jqwp~!m)Sr
zLpiMiNZ{V}5<aCBBo0gB>)b0hPO@c{$n{wrWp)&SW4Qv$bf0VNTMT?O-Z*#kS7X$T
z{i8hjhUzvuM@~UQiD(%Q-)6HDk6V6cO#?MtcwI@3*u<HnP$oZAuSI~!o{&K%uAINU
zBT5fX?ea*U{jKVy7hPEU$*Hj|1LH7C%g-r`v$8vcp~5I;a3edDN8?tmC@O5sJllzh
zqapO=x%!{c7PN1U6r!y`SJQKfCZd1rAbk)UmjVa8a&%ywYXuM_;!E|a2rp&NwVekS
zCDF*Zg`mD_<mv3ov@7oC!QD4#^L&A3!{z`<smOShn|y+H81tv$^V9XSqtkzFZ0oz9
zZZa<cH(4)Ll3MiIG=5&@Q)s5YKkigI=*#lj6`~<kc)uR#&nN15y<iL()-Ui@5EA1B
zDlymLkKMSQq|!k3-0I*KMu*w+XThRo9)3P|bh`*d;b>7;0$bhf<~f_Xbv~cfvHQ5~
z++&b*>C4;+Ga!uy3WGk!;WD42tSe*1;Ry!9tZ)HD9~<=3ijhUz>|Umk^H@WWsTFJM
z8Aw;@adr_dd!QVJ<{xA2TW;qHWK%-Zx(pW8<u4(qkl8H9Au>lGRg5tRiC{Czo+2_G
za9yj=2s9*OPFi-T+b^y0v+>%1p#3^er!d(N^Sy~Zztr<t$V3yDE}L05Bw0XWQKoTw
zpM)Qv+>@D=&^%hJ*w-F3zPua3Nh0R^y-T|0IH&7+*8c6YZHp(OL!9|+&|~#Qf3y1B
zuf>;aWG}Wh^tGNsdR?&7kZg^i-tjbkvz7L}jTM9-4Ey$=&$yF81+VY~v`^+x?|jH7
zF~Yj&-K>%7F|be9OHn=PKF<kdhjxK=prUlb=U&=|6ex8MBRH$5H!QapLyQqZ5K!6(
zU3l-AI&`_5ns}mIy6cV_55CvwXU^hU@;RRi_WNqe2;Qw#IJB1-qGtg11w7@W?by{T
zckWprUzYUx2FR)FmX9#o8VB<=a~T@1-I!hob5`qFwOkhzVX!PXv7HuEJI=VID;msS
z2X%1*u8fFu^g6}U>6Gif&D;mdWT0y6L5He+@Y+nulC!SSytOLyHaP`m$>?RIlp!Ca
z7G~{v(sx*Cm{{lPtVm6QGJ}5*>7{{1cU|6^D3dWBQ$7N1X}}=-dxi|%+-Evq8q=mN
zC3tq{v#je@Mo=}Akns;5P~sKqd4^+$6OWH}u~HIOn~jIS{jtf-9c*597>`&RpLVU_
z5$u$0ttBl{Ewa<;_B*RgL$*vCK>s5@tM;TaOx+NvanZ)j(J*q#zD7+Nfb(>b!q9+W
z#Z#_+DMJkm$w${y*1VwEm;EgQq{D~!uiP^qYrFyfzn?WFJE0<$vnYM<JH^KWFXm?0
zgYipK%gfm9@avFKjWkS-(8~OXOIjFZnbs6q{_?yDgHV;5pr(6l!qZ&IZzrOC4J`4Z
z9`+QQO_A*h(rN{^6VQ69zVN*U-H?0#br&(_FphkL<ck%+dY1tjsd@GgOA?U-<&T@x
z2OlLrT>&WkF!q08-zAwbT1eB#enq5$i@wXhlG=v$0=A_N$h@Eol3vgqZvHZL%?oCV
ztjgq8ssD|H+xp~nCREdSbz+Y37yrQ8FK3481xJ=LhPTht{Rfjrlqo7a;`%}DchQ-*
z04ARrE!;&Mi2Kz4@>-HaI(v#clMhidjjLB{zd`VC3aUUspYb1_WBDQbhD_-EO_ACL
zf!~z;N{6AI=R@$J=X%zX)CI^jV<^*Tvz_|sPZp!I;qVr%qh{MvfAY9w$iJw8f3pK)
zC(RdX+mH7WK{S|Vc_5lc`#v`gR1l**niF6<+Kw-|jvmIneedCp?wP*(Wn;@8+ujQ4
zm-CioZHQd*2PE;K%p5e7G%*%j>2+z*`+>*f-JxZRz-h}KSIv>H|HEsv=b!>EI}!2)
z-tm4#TEuyXw9Om9xvy=lac8fKf_)I@*i8cWoovfhoBnN^LT8_m(9sqeA6oV9N~mpb
zCt=~K`FI-tW*FMR$*SFkMSmyq@Ap&G+U~Y?I)5?|o-J$A9s_?q9@+W)a5gP?bT_T{
z;6!q!<?DTacXM>JvEBN}Xy`pWl-9QzyF$A@)1ueZPorTSbdw8Cv^YQ9<v)B7MATyR
z|A$AR27E?-w-Q{p<i1RPk5HQ_v{Gq7N4s4m_XQr{>th6=H|_K>Rrtf8%xy*9{uxC8
z*EZ$<T>ePI_-@RGjzN4I6}JZ*t#L1hKP&hiQ)gXA&*(1U999yF44K`~;+Om`)l$6>
zJs<9Z<JPME=LS@YS?$VH_fgANr@*jUYNrY<)ZVWQ7cIAL&K!B~^?bzzpBufO1>;+!
z6q^z=-DX}ggw4Zi1KX-?yk7p<oO=XQLQC&9-J{qp_b@`DeiL8LS=6lYMddIIRqpNm
zQ)<N<zw&qy_YaG*fs-T6p@}DHo*J}_%W&YvQzW1|cj@5&rS&@|=<}#%XZ^x+3r29(
zL4(%z3dWnm)C(_1r>ErH&ZOV84m_M27Yzs_V|vkjE1WkaZvoOFr(n(s04lx0oCi0@
zBj4Pm<37ufrw_}K&!0TH5NEN8gkU1q7g%x=!(9QgMG8;S9)gYhoaEJB^PE*HNyVZS
zPggPyf?yt}@u$dB$3cycL`mfe1SW4vLiA27RPdRpAGwm!k8|Fe$LSGbEUXj#Cqf5e
z1QQmn;yNltjS0on1nUXR_MoA`N~;0eej|;<yxrNsD5>1b`7GRh_e|hnNOMbH@7E7u
z6ee*r%<@#33>*C?UIU(&g}4Z~i{}AWMa>_TM`k7i4J`m&7Iy4n@iMZ4yOY%@zSzYT
z;2v^*(arv(Wmko>SlgeZz#JB0WG1Vl%-Tgf3Ra#LE+6HaTZm&wdSz}e+VF<u=#~Ws
z7^bQ8nQv7wvBs->T<TKG2+_0keE3|#F_76bRBNu5^uRmX{8DlPGT(f2uG3_@{3bX^
zr_o{JHM_;M9rhC0PxYc~n?2?q9Rn}20W#*XVoe{!nB<(-M0eH+an>dQrF&*(H4eih
zdj{iQMM~MEvxZ_hD|3d~@T8r6WX46jRB6tl0PB(OrbahFguSLhFH;GpEc3YDcPrFm
zw{UiPsD!wAa?(JE_G!MYea~stNQBR~Br$p|RvPpqsdis@v~bHIDD+eFGUS6_Y`Tvk
z=WPE&d3YUj4{;vYG$z*z12i76Uy|uY<P=Pptt1ONR%s53)S`M)G@vJS>6Z0f%EWP}
z$$91(a1*~tlOs;adp$)G9+gwz3&+mycaF>dcP~JZz16(PQ#9BCb)2Cs7{HA?F-z8T
zZdUDP&A7kNMi~-V#PbqkHFd~VH7&9-fqpFLm_TL-jPL9H4v7II-jkq5rH$E?zz-!>
zLr_SK*o85vgk{2HYX9uuya9{RvFt~PDWO3)jm+1<c`~BDiv2A1*5xByytRN_>Y>UB
zvp*I!vnl)`2QS}GavvbqKfHwvdSC=OW^QN5qj>-QsCWGdPZDrz0FWLaXvB$-$PXJ<
zw)NRpE<eNt=(Mo&7w&$w!{~{8(Ql&}@`-QE?vsJ487PX6vm0;QBR7MI-CyGps)r@5
zST@6yUSk&R{GFs7*#$p%Vl<w|I{RVP$Jnv`#|CP>mXY%oBx(Eh;aiBF2iSKj;q%9u
zGSLUB%s8hE(IzxGML9zNi36(2cj<N(1@cXF;z^3#;xsid_<fBiGnzKBI9QZ8@;AVY
zB+HZ^8)nyp>m?eyUA)hoc)JC>e$+aRO!0he__#iM`!1?QZY7GA5`@eLesR;!FZX#1
zlNumVN|N40tB7M)PXMdb2X4y;5%BChBjG@>ogc^`t2S)LsGAAV$XCr-)o8{T6e~gc
zios#+b2xeMGmq1X#BVaz+t}dfs@)jU=A2v;fT?wRmJleAZ9cU53%{TMcuS|}FoEOX
z)?k<lm+uAkVs2x%>-zn{!+V@syheeWTDTi4%5AY@?h_$?$on!FJkNYEGrCe+!Ook=
ze9^JdisG+?s!ZIeM+;xnM3+Ws(VH8ohj~2>+jp7uG;eu{BBf;F>?qVlb~cRRFC_a@
zLBU;oQ3Fq@bJkkX1K4iJ2Ccb?W#Z~PFg)0w|Fk<$#ML6+tr3L{Tz!#{#O}+*3m8kE
z3;!#`mWE>wH?l4#+Le-2mjD<i=L{Gj+kR=bZ-!6L@74Bh%f4Leuc<BloBR)=v};Ma
z-M(l0^wXgsbxotI{3*`(#F7)S)26R$Fw7OZ34>Jgb>g!>GOPjRplAIfeTJ3ydov<}
zc0nHDo@oeneg?KwFPbu#7aGfm0eCs>fSB<-2T#pe3@3hS5D#^57o~F42of(+F*Z-u
z!aXgcnTnxV^hkFj<#-7;fpR!T1Ha@m1y^@};yGcf^2%q=jNch)`OXk<4_r};&phqK
zWiIcY8x|^g&LN)A%L|(8cUE?o6}AEe+R&i&d`>hz&oQ$absiQctTbNb&^~b2I#Gmp
zJ`0KvN9Q+_CjX>C4oYDNOD6e2v%!fWsR?6xiFZ!WZN*Kjn%Xr9LZa!rqfNvSI3VC0
zteg|;(;`n2hxFBsRq;0JTI9dh9;+9EOQJSrk?tmo&)OLaFT6HS4oeB>dHT&7Ly)!-
zRZZ?z!WNA!ksV{sMfx?AZnVv$4gI7;+4iFZWJ52X5Pyq*EYni1TlDX2>|))Kz{%h*
z>pH8*oSoRdsFj5w*JZCW;IHNj@1xKhv&plW#ghvEW$zOevQV_K?lw+7&#^v`P2+td
zL)g+y>jgVG9Y1gfLC`_h!o|`>%l3($7v{S|n<(Ll`#)TIUsROv7VY(yQ`F=+c=<X}
z9b0h~j2kp-R{?l%n*R)_r~zJrlqJD9B_mTlt7$>l5ekqW0AJEsf#i8ZboN+LyR9-D
ze(XJ3(u>O30aEs79r9$@G~C3I>th)|L$<3s8@2$t1z5uFJBa*fXL$~lMC0U7|KMr>
z*=ifglQyzy=LS&pjEgIJVqjafg-*1Y0TFI(B3#E7eA<_J+Iz`JWzic7b*P9XR_cJG
zQ0X{!(=-~V_3I87%>$u^Wc5*aj*4D)Q!zye<SjjIEYh3G`4{(5Dz9#Z#^Pxz82j))
zbft?9)~q3OR57oM&|blMje|!0<UVhQX3dUOm8eS71~2BaJMM?l)MmRMbo1=brtc@K
z5uuRw{t}oe6LSt{tS@RUXSIq_ylPM_TGmnw3w-{Hm@MrbODD+i3ijHV*(mMhA;2x~
zk|y2B#JvrKF`?(5b$NaFU&XGnPO?LQJI;<3uc;FR6_tBmAuHpBoqSZnoAby<)oxS=
z8PjC=j`(aAMrQ?PG!(n6rhiV}7V<2W1Gm8wOP2Xgz9#g;Y~~t)FXJO+n4?R8yK1#+
zKlKTM9|l*L1MZpptCC8!Z+lY+YZoieGNVb2f<lh_4W(nA*DhswY+?NKr8DC)TGZ_!
zZ@=tjZ$?48(E+T6msZ*~y5~l+`a2<9LHH!_qV-4^c#s9#bF+3KKA?Z)<r?RLS|*HM
z11ReLMYf5aE#IU3K-&ThD;A6bD9QOE;)tF6?2-85Kv9yhU99y#ifyM9D|UH__zv{)
zXw#5+-0lJxzxhcb{9H!06>TVS$iKO3_1IMU^{6(5snL*1Aq-9hBNe<T&QYZppv6O&
z^|>@zPbFlymmD~XbxUEWy=zK@C66w%h&fAmji}+yi@o+fNVO<3p%#2Fm+*39>QW)!
zObo>&c@_87dzLtTyCH!YnjLXaNYE)QyG_o+%H<RlM%#8wH8_Zuq>@d093bqn+_&|7
z-I}2mpRY)H!^)cB4>yn4D^uxc8Rpwck%+>HKLQ#Yz9jBSEicd(Rirbw1M7AVrbxqU
zK`?VAQqL!gz06MQoI#UHC-dYq)MXB_)&<_yL5avb%?(nsW_Hx#^lP!tSt@?zFmz_T
zrD297pjP5WjXbq+T{Navk7&L#pl-J=3LTi#9p(~=fVSndIR}vXD7EyP)G#hsDt1dN
zNeE+A+yKjS*6Ggq+d<5P%gsXQrKgC?xQr>pF@Fn0?RDF6bHrC(Ci}SJNhD*NO6%9#
zMd)5fJSj@2J)u~JK9}&C9fH43b;kfAM-Yjv!=y&;M1A2<HnT#jlZXtCV37gqVQbNk
zuBHM2`q+L1WV&-(&k5+wwOgLA&?Nyr$Z`zqt(45V8(o%EIvrnESeHVn+!P8=e=IkX
zSgiCL0(oV&Xq?a_h0e&y-&WIP$o^7Kd@@)1)dp;k<Z^EVE_}UKVrDox-GS}(SA(}e
zpeavN$#_!WK{E8xrY?NZjESnx;PrD$jf5V*O;B#pGt2E;9Y!g~dxrVxuN)a&o3`TO
zHy9}2<;37?r<-<s8^P)Ng>dZS)xOMQj@L;nj-b0;T(VH;OVYqI=Tmg*+|P6J28$9p
z^h{f})WOxSSc;qE%yJuG&tf{OM|!JmVCMFw_=YbPY8vyZ(N=q9#Ts7qaV(UmhL=D@
z9t0pDaEC#4boPlSw0_j+o}s!t_<8tHj3`w0#da9Q^n68*Xx_VNUMtRCCu_ATJ&p$F
zfc1(-^is35sY1IlIk4^7HBhLzF1Lm|q)=V1V=pOqA4quNNeYp5T>`6WQ2foR^Rrwa
z+vQjY_1&nTeq4OL4NjXm!UEB`BCL;9-ZBuc$NSlxYDEIkBhrck=k2$$7u?GIaRn|n
zK5k=J2VJ&$48F+LNSuqW{C4kxzwr@A8$g39CQ}5F*SFsNc<glmkd(k0x4$?xX;nkh
zOUSQ_W92OQ1Nhlo&7g$n7$rws#F34Qta{s>7?~=JYgbs`Rf__p=PnQC|Ax9j%zB>^
zi71kUJ)hB;kTMSc8L3Tx!e9Sl3d97%N^kqvlRPR~{K8NmcxX8&=sY;5qc@o3msK$4
zdcBu^v-J);no2AzxtxP(uhT5h9@5g;KO1_Kbc+yclvf>ImI9e`5%0^Br^RI2txU<-
ziBC@Q5XOsI=l?VX(hW!0VE_@ex^!xuI<ws@>AU)P|ML%%e?)MecO4&FJ}intmMiDh
z_Vw#y!-X+weW2JViLNIn4}DYbAHAcg(vP0aT2LIB4g^xk=fz!-xt&c(JgrRg$R+3(
z^LPmPMPNr1*t12}_+>3A#Iv|q%3q4~J%5M_Zwqq_*c&lw_|PTa8JjK--D*r@l4d`v
zx@lV7jdDlDTK?;TC~<8xWnC%u${xW(pw<*+l*FF(RU3=PPRpClw)JVa1UWc{n7bY<
zIQ}@(kfrrP3d;sC=+?_=K^~aUm`*`j&=w>2f!OH0xsL&2{gX8b>d>d`9CWpJ!AW%d
z4|cSuuC0p>>MnSGk}GE*?d;VfPWu^pS(5AE^Vxo<b#99SQkHf*|B_*zOD)y##u_;K
z&)8AXfM~!OR@x>zkl8J#%uD_``&6MZclI)Nt?Hzf{$K=DydKC11I-FJ$877XZ+G1~
zmnWDawC$t5BE)M(zE^Cr7S?@x1q4fP$U=r{KF26jumq#C*K2Nw=~W5|e8kfCv=8QX
zfTe|me;yO|0io5lpvHKx+V96jIZ_UyDNe>Ix`9ga+7S&EuHQ0oR+|tuii6t<bd~?v
zyD-3Gg3{T~sb<66>bsJa@ZN=JA#@6|;hRsX4qCm{pg}^0pXsKD002@4I;QA1GDc8F
z>M+gN{Fe<gFeLZ^s54YqwczU8cw+Fvb%#8pf14Yy>z2Hbc*^t1!`X)5OxaPzB}lnt
z8;0c8t!t@R7psI|zN@~*+PonN=xZob77K=H<v_D+Ne~c>#E?A6g%)0l;p9)0aI9KL
zoV9xd8f5U-GtS^40w*joIUR=0(VAGj7WSoyu;_!}@lmTfbgjhwwMwYu?g9^Dp@tvx
zNs<(9%}G^3oo(eQRyg%g*=%_LXh&F!2fP3680j3vL}*@iXAccJ<6D-8?_QIzQWm~g
z<EA1b6s`TJ?s<@=#rVN*wg&|AaFAqT_nk4{Rgk9S4svK-C;eR8y;UQcK5TSbIjT{Q
zJEm+y#W?&!vy-#va{eRnWqj06?{a)sl1|;o8UFI|)VnZ5Nv@sHbe5x)YEfqktAy9;
zzH#wJ4OQ#0TVte9hrE|t(2|K+-Ie8alNs05E*{Yov_8uk+ijTV54^i;3qWl1>-%^}
zmS`ds54s(HuXe-tKKc$9dg(-zu&KTH;jCxXBOM4~@>oL@7kSo|YV(E$XS)!m)+x>e
z7XJ!&Sob!vqW&TI`IGJ}a!$QHik7`sPmfU%e|&n|WH_)*7eq81=dvEEu(;uo9MM%K
zI3rKWZ|fC|zDkKYhBNJWLT%1}g9j|bE;;#0Y#7BkVEd5Fa}8B+bXdS3kw-rP>EZYI
zACAYu3#wz3pw_P&pqnW?gjEK+7Al>mepgC&R;m3exr_NFte?muyT4b>dDhOkGKYhe
z(<i6~!!!J2Kjv!tDuE}%gruD){0X*b<(A0rjo*LiqYx5G5_t9Kfd;jtm{oq3%%{AE
zFJ@0pnZE|R@OjqXyF1ItO#W%P#3$tO67QOQJt5D>oC+c$cX=*L=aR#aembtCA=f(3
zIXr&GM66F5NY(1YR;dBRn8ud~SIDgCG7IhpU8XfQore0|>-~k?3f2LGv2Ne{@aPsG
zU$;-0%__Q<Z5GZ3yIeP*Z~iqbsSL&gIwZM?dH(FiF%N;3PGHJi1!Wn@cJf=mrj0(9
z3sC7<Mw>A67#W~!kvlovi0}fUjM2VBubjH&NqXJ05iLeAcF(Xv?=gwE3zq(5u9Me`
zAicT{4abcYgZY~D#3sx9PMvl|?xxo6VEd~pN~=*s4W_!A6$4q}`WyNKd*sRY+t?HS
z+tfFAB5bxq<RBt*(fO}4y-f?~6#@7XnNXwRLjVr<g-2%{L2aeQLWq3sk;$Kn+3=X8
zcwpSE^Yso1E3$rgeM{JxGhudYj?_#t$cRJ3jE5EI{udo72iS1e#Av;+s3OV;6c_^M
z)>6sm3@mY%Wmvd-JsNu@72jjuvvR(z=sLae<}OWGd{f29RA*xQn(4)ySZkBv4l3jA
zjEm;R@K{zhc~B+aFiM@%fkZc9EuD-3*d7PyD6K_zmL4V2yWPTTn=|62M)%~arB^fn
zisYun-asVlElIk3)SJ7eio5YhLazg~>cyB2wN<M(6S>79+bn&1K9cJt&xfysnUSXs
zBqQ21A^T$8!M6Fp*1Xn7-khaSCIHr}w{HaiD<KB*O#7@AIJe&2HJXTxR(%t8QD@hw
z>fP8l)(a{A{cp6$?ahHwN~e>d2a9SW?IxZQoI{mOAmO)FuSfX;S%kc-0M?}6&p-XM
z5DuB|VLltR0ggo1?GUNQhLvd@kOLpfoeZZ4B7iU_RVy^>&zCHa0xvXsq3#2;1p=w@
zq5rBI-|(Ri|EAy1Z`YyZ4$@p2er7WcwL?HBOR>r==i5tuS1(T7zQ%gDrX_q;EEfS^
zuX0k2)r6^9foVDWFMV4SZ?tU!)%U{Plp!<xZjMwIrXyd{mQ#CHBq!~(DdZnL>e)gj
zTD9ULj1eA0@uZr0D1A>X?wX1Wu~@fE>C|b-NOY7JJO2LKNr3ckfK1>tXRCAlzq|-F
zf<f88rO%yJt$ni`DVoz7{P0e7R?u}h(c$N8R%tLS7WImRCq)j#yR}_TNnr#>>QN(u
zvBj+1NvbYJN$N~@1$Sj<9RB9UDsiY{K;k}p6X{+2?Bb~+8>Whv!3#BvM9f1$gtJzp
zl)s*4&4hMP;N4(P-&lu!yI(sN7-=rMPvMxrq%ursCNQK8$gI2-ye?B+=+_oK*SanS
z#T{W6-6y2SUU1AM>4A%lfHFVVo4d<eJYAZxi*A_O&ah$`A;BYZ?{YDdNz(BoxeI9b
zJbdWRw}CZ1TsoQp&1PQbg1*7-k%F(JHM17W?QxeRPxdbE8y0O~xPETvE<llUD0h;4
z2p75T@bpPL!s|jm1<3uQQgQ5AdJVYzZBN}&c1Rr!rRw{?z%lK*U5{2(%>BV>Ye9xJ
z>-h%j7b4zG0aQM}wiY;FH7Z{*kpmJO4KUFS!&Ncn=LRZv4!<+1wM9a$MU>AO{|xcQ
zpy;M^AIKfygj#irl?iJk9v49F89hSc8PVqx<9q>c(#B~{!&&~;L$9OvD!m8A4CBq|
zu<%s{y_IONPRz~RNtN(0TkVqFy4=FzkFS01r$7rHIIN0bHn_bP$09(X(SZ7?G?{d+
zQljyKw6wQ9*SDYPUc1enaZ_P#UKRNK>Txd-i8NI`Hsfa8Yr*>4!BxU@deA9MNw3x2
zPa@nJrqLrj@e2Sm<0$`q>m<Wb7bAuKSya0wE+SqArr52JIM(<<5=?%0uZhL0mi-dp
z#+6Hg=4i=<@AYZ`SP=`(JARdlG+ATgufp6G-0Nm+cd!r1JZ5D!sGm`l6Qh}t_`Io2
zwITx=)$Z2-JJ7Dt*WarD)U$Fn9pn&m)BR}&bCVZC6mpXB1gZf2x<Hq&gaYW2zar%v
z&ml8AWc#zWS1Y<{{2JdhPqXElJ3POsk59Hs5LV)D6RjTmv_q0lkn->mwo$CI-ygPE
z21rrPRoaA2mM}9Lgn@$#%{ME6<^`8Grruh+Ajp(#FB3HY!Btl%QY4k@$Lrj_1)v~6
zLu!+^!W?7a#G^b07{>+CtITeT8qGr`u9+5q50cP#2DfuXo$G`Z4f$v=J)1(x!$Gv&
z!?Hw9_UcSo8S*)vcpe|Vnl$IsrZ*pJzJ`;MH(u1tESE119o`3R^KMac;~`G;r*hN*
zVDtijhWISn+=pCz#uBh9G%Zd6Vaq|=#f4ez|DefyH$ri<-1l~!skBt&KCV3Q-|JN_
z3;YFWuDPY!u-a<#S}QvA8|AlmY0qU_)#D!NF=_>`l!5=LrU7(E&1a*Y5697M7MmFh
zmYMC3bdqh$$rA7VE*yKo{eg?^e0;TFwbR?U2!zi_?4BulVsP}_b?H&sMEFMQb>+WB
zLQl|SX#p<nVOlS%?QHbM=i%3b$JUf>S>y@AQunKZb_cIC;V58$F`={564xb|=M_Y&
z?QFyGQLXfc$(GP&d47`gz{QE-8^h(flQu}hw|?eH5&7LE{Ue8oOW*3UpK;;6zj*>B
z-2^7kKY}7E#rXDWDv_}h+swOBw!5w8T0WcOfTm}!&H4R`o@sF1{e0gvv`1wV4b^L9
z^hWeDR>Aw^z|~?@F^uj%w|EaIZXT{4{^mdQ`2~84I<6_)uRX(72AELbA%RHu8*Ho#
zHz{mKLgmLOdLFHIAL8==J?6KhkHV;=v6kAC_RbU2`)s?K#D9Lo!*RPG7Citi`P(qF
zC;KVvKX3dJC;3CS@9C=H#r=H}2BEoo7@)Cw`@+vt^kj2c;5R7$2mi&d2N^ZOs^z0#
zTfU5w!Okrg&$Bn5U(7OuO{+&TGB%ZdRb79nrTiQV*&pCMd{&R2`-Tr%cuVyLH_oUp
z=Hnp9FC+8hCzSi~UW9%@{RX~_Z>zW4&lpkR%@Q!zg|>w^5aS0)cuC$cnk}Ig_fpm}
z1ldFK&F7CU&UWV{?G24W-EuBdsOKDA=g?VvA7zCFbnGy;+UwooC0qItCY-5gjo5+H
zcN&Ru%?M_&sPMy^#OICO!%T1JO^RtylpAJl<^(g;D$Nq!j3MRChbD!TO*0e;JN-+i
zQ+lr1f8Rh0^kF@-?BdRsrdd<_?x(5-cC1X|2d+1Qoq4pVxRTk@UKPVCY1Y`rO1e`T
zpg0IF+pjg&UnFiF*o1N?-cv2T%R~+tvLL+&z<nj|1S@qMGe?8oxIaHl=<01IlrB^2
zrxmJyN#(k%&D8eBh3LWHf+zDVB_Vwcu>gf)#S>4{v91R&ioD;u_ySSnukU;S;z7h7
z!kacSVW#Z%K2KiuzykJ$=*EK8x@Za@7;Ih;gMvp)xT4k{)l@eB=m3Vte}3IO*V6(n
z`D0Uel@4Lt8J#CM>~?KhWC%+g$r2q~sr)cm4TEbM#wa|JZAr9rGQW47tg3~^U~nE7
z>kw!7P)~R7h8<t^eb*{`SPnakTh5$a0-!0&Benc8&Vi<FQ-!9*(MrrIQ|@%lTGfOG
zsK2Eu$t(D1dOXeR>M9@Uy)zGu#O#j%U&K0sQ`GlD;s;z_btD@8+&y!q+R=ck6RcF#
zEi^_?Vz5)8k=OyfsU@|fh-l5&cF&)Dms=&MGKn#eD`6Cl7SI9^qmA!kjb9NHtBHhL
z)Lc{Zj-Bi+1+LmQ0T~%+Ti{$utSEHmRlECdlO3Ct?=FDAX8FkQTgB9#f&|Y)N<Hmb
z4X)k?pz(N%)76@!e1l44sFGW*ih2Lc<5B!!QKC8|UuK@#Yrg<@0i$pU=G%4$C>yJ`
z=#~M5TnA^<R|DfP0N@`M3+%`#p^cR4(i9}ZT;r3E3qY>ZEa8g+P-pJ2NUKGlS;jtf
zgzoS5o(OT|lk7zs>H-+tz`N#}ghz`L&}S$go2P1?D83Hc5*ka1?up8)_)2nS@CNM1
z11szx-8+-d2MB0O3I<QzAxWixbPZP`?U0O-K}(h)6e_Y7kWT5${=0#Oq3m%tZl)9I
zPQIuVb}QFEqrH^D0;Hj8dcar~Z3ypD#0ovM-|l~R>F=-@uke0C)f$z7n^y5x9CMzO
zuzeNa=kTxse3e`obWRCk`t&j59g+OJs^%nu>#%DZd9XZ{GPh@40r6luivHNF!+x~3
zygxJ5fzF5}>?OnriZm1k2!V8P*-H29e=Z(fgnk8D^o+}(<IwV|FXkv!BL?(Tnu55&
z5?;RobYRz^HT#(j<wxz=K&QQ=QZ;lv56E7}g-@!ET;6q)13DQ1^DTc`Ww@OQY-`4v
zd}Q8#AEg)40l@iiLAxCQrm{{C431s|(ut6A_n~t@T!Hz&IRdfR#GhFru2(?&KG4+s
zN7!){NP))1NGIuArViB|z(=RFdCCW(N+ejpT{JQbyDw2NF!NUcm0u*}dbYS&s+hb=
z5h|gy!R%L~+8{<Y_Y>7qo=6ez)2Q<84-Xdr6${j4A9Ku4({bqkV1dCIx;~qNV*L&l
zqX3XIy6%~|7S$|KhOP|>G@&g4`hZ+#=DwKTWg(ZL^1Pi$>Uo{OthG2b|1}y{Ftej3
zmUJ?NHc?(Wjs_}*^R^8THX2i|nsI+9z2hJcS+5{PQ{x1&&A(vSi@;zIWAfKWmWbw>
zI`g}Bi6YH3Hu)}a;QZ#Jk*L$TIG(B={b#};mcWg3V*Hk$6K3^aEq;qTXf>_c0i-Mm
zMa)+1{u<S!PVSbwZvH|lwX4<9Gb~GhQcqS$zpFK5BPb(LT=nK>Hv$*vq~>XpzU+qf
zCr{ymb4R)<!EH;Dw_7YQTUZl$<qeRa+SViTfL^<IqyRXwUE6X4KvK`=y#M|@0mzp*
zK;bYo2S_ANfp#lE*n~XxsD!XSx(7U`v@Z&DD*>cVj-8&{W%39V$YX&9I%%UHK;|NN
zB09~F|BoPsT$c}h4Dha`5)EnF{HEZ69teCy&@VJbS(xr+Usa-ISndbMqpzyOFuF?)
zMDHpQSh!wuVRLD{noNz1bgW~YM;#tviN|+u>>=WzcF+Dg22g*;`r>gkf3W)Lt)3PC
z+Gw7g-1{=dn46t##mtd^mdj8Wx`^ftX)K35l#50r8f(PSa5FnSd7m7^F?5YzCuBV@
z@aF5U)gArNKoBqM;2*~fktZ_3iMQM)!$u(qXHT_BX0NI8;3)-Gavtw(`UC}JX7lR3
za@cz$I+Ub3od-+!9s15orE;N4<H&WMV!i@<>&7_y?U5SS6ukEm@XmdBkG#aN9UUn2
z(wSwAvg38?qvT{n4oof?yQnG(D32Jz0E)w0qJ2y{%WKJe5J*`8i33U|(7x~+o)h@f
zbxgY+fMo%T2PMBmG#7l?2Bc;-^$0pWv)J5!_W~p(_>@XL`^dV6F@v_GSegu7{HsUV
zxsY;s&7lJx=a$#)5^D<3pF$(lKUA8N6fUMfcmAo2D#3uQ`;*;0<*PIj6NCBl^M_-u
zV$W>RL+k*?V0Qr&R(^GlI#{3+&>PY0^TaeQzsHiR`W`=rCX7lql25LEVenQbK+%Uo
zEO&yT2@{oEwKkCpWv0W*ah7lDuzdA9s&!$icl(^7BSTe|{a$j%M^!<pL5JG=&!CF1
z(zV8?KA*I5UCwqX8%lIE(i(+0icuYvHiI^-9f&`hpkdVH24$vqGa`$5H}hyUE;_+A
zk!q@3KgQTq;DW=`LdB_kbiywIWATMr)qI~?JAD-8)Ptkq*)en0C+M@Hb5LiH0UO#Z
zRc5@K_RH|u2O2~-A4};8Z19+lSGN~HCYBuvH7H*<5Th|Sgydg4#s5kh1nQLh$qw1b
zF8o8Yuy9*P$9IROlEZfV)gL<Ps_$*4DkeJD(RuJ}@b-YzCG+qr$^e4>t((4o<p8IX
zjU{rD*ZgksN&|EQC}x08zn;A)+2dmRg-EIAe~EAUjssx2FEpB~yJci49Gq7bT&g(Z
zjPIFktOCM(Hf6DVEAg}@_b_9P4(b^eLZzv3<aZdVgEO=5;!`@7%$ibY6;%x+wf?*^
zz#jk#FEr1+NgxmiYAXh|uXZWv&7Qs)a;E-<nc)isM?}!ER#jcMf1>2xfFJ|(nn~3_
zU$_#LG*j6tv*e-|cl602$u66uIEa1CZrj=V&@drozd-C-%`vfr$M(#FadEH5XbmiN
zmxF~9fV+Xo-2r{mgo!=?S^Gvnq8ZSRP4-NTsl3ytqasTitWOx;)~nzjc~NN~=kV?@
zKchoOv0lIUOXu5c?NO+1{bfQ*<c4Sq#gU8sdXpV_1t1Uqcpu}aeu4e~G_348*=;Ji
z*^!$gY>TEkqu!k7L&+9qvu%L_=U&$P;4V&7jDcEXOJS%_$}S@qZPiZ2PW08>wVG-F
z^TDb09I<X7&B#_~tF9V7X680j3BMFL9XI8Ze)>!|@DJPddY~nA((rv0OWjsK3-eca
zc-}TwS#|~f%pN{<&x_TLW#%EPpMVaf;(f_Tcxs4MC%((a?lMcTo_FiQo^zPkz6RY|
z8QD;1PKe+RAo}y2$woOoE%fHDUHc8`!QVR*Q%I%nW27oKpjwjOR1CJ6v$IsW6X!fL
zY^1WVAZAw9j)wnz$-4KiNpouuJ*n?Wf<4@Tav9SmT#PfXSWsH*ZXC|AoT-#dFVrwG
zTE%?~NbC6ah(%&eewK^?<J7N{NsrK1Olb>W5ug-D1R&7K^M7%%qE9)Fh$vQdDOgEC
zd8@1~5SkBRwNi>24-!H6)^J@DJ)$a+mht4Ufj-q*Md%DO;e^<;(Y8_h)Z?9oH<iOW
zs93Qqx<cFQbOKH@XfA6R9L{u9Nk+83AO0DAJ)u0l8KC7bO$n(JM97AadI+6|vABD3
z_DSV70Z4}jXDFm7q&+U&?Lrf0_{Sk8hc3!4od)plC1kg#n3@>>MDHfHq^9N>lxz;&
zYwH{+(70NX5Hugtm|d>qebja16lpz2_SQ1n0qsZe8Q%|Q8=#p&mKe$zpQe%pZ#CFQ
z(Wtg_;Lbqs$L{B8jJFc5xL6oRCvQ*lpWh8?L@8d(=2WHg^Z6z#+s`SQYU72u2`yNy
zYLOH-fv(JlQJX$z;KUXli&c-ld_LQmSM7<issW1o#q~9LVQ;r`Aeg^PiDAvMRx<<-
zz~(Djiq>9Xo0i;Y;Iidt@b&k#*bz*q2mCMg-omZQZfhSEL;(S5q?C{b>F!2AN|f&I
z?gr`ZkOl>%L%O?5q<c{!x#;FMm+#*D_wDl+oO7M;de`N<T&^{rXO1yujC<bWe!w^p
z3z5Z_OwElEziq^W8jf{`gjknUl{-=i{z_EBS(}&wdzW(VxER%U$7X0JaiP@VU?=h%
zlO!BD<MEn_%Ubc>(O2{`I1JzOwYS3)&4xp?O+LeR9pCbqkNNxPpJd(!#1{+754K2C
zL&$obDHPRvE6!Ra<ZBhXEfVpIJPnZnXOMh(>JK{QFItP@3yqQZo%ALJq7CW4*fo4i
zZKgW!<nny}^zn^oW7XG|;T6KejVH_A-CR#qK!JpgByXinL<NV0RE^xgfxnHesa9gx
zC9ii*LFQzteiJsg8oaSl0kMIZrE5sh5v8bcOCFT~#Wrz}>_I2j8YCS(gJn?ME7L*j
zOZ)HJzH`P!78I`MV7p6#U56&_D%YHvvWMyyW6@wKI)hM#N8<zu3*{`Y_&YDVU3B&x
z?XypjFSnEgfBvaMv<S9y{HmXjyxQb0T?qMS|18UQS*^u}x(cftj#0jV?$opa@&nW1
zYR~qEgvD_PoSfC1fj~71h02qZkDm*`hF$vGXO^xf`30TrWfgaLs5(cH-rSasYY>0*
z>(iAxG2%ti-WT~80YNL$`m=k9BxOeR(!(Wtrdp&uG_N}wyHIx(<^~aQD<HurLO+jn
zy&hPj4X@kba9`>Cn(8}ft?W4*tspH#?!9J%UpXc5*$Y!muRru3TRAVkTQgLUj%R+(
zTB80Ml#`!KaK$43F%W}}S2QV#m#<eEs=Y5n|AhmE*>p0m{~cSmS%o!iUNAv&#hHH-
zl>p90FCP`jIXVAan5Qzsw|Ct$v|QrhR${&?4FC~eamXV`<(lm{nkeopMWCvjbmq_*
ze&rJ1ja^oDpx1xfrE#P|n0?)<k|g*(Dl;^7{}~?zWTqyxkN)MpCxpv(&Iim!IVk%D
zd`WpHff&AEQG+sm27Buse}dLp2SBy)!sZuR=vKVRa`U<zXf<qwjohF7JSv#}!0zhF
zYw7!!NIb10JIMl8@ldp<x1HJ8^s|NC9Ktvfw~_*h8R-RAJY(UhYXM5io@=>c8Ug8t
zpE`2vlw5;F7$Ab5JcrV6e)7@%{zYcxLjKCq7i`6T_lJL7#pe}UiY~7iXCZhVQg0nl
zXk1~na2aH?2On(k(qYnibSg2P#;im@9(%l*)mPH5%Y}%uIO_3ZBu=@>PNu1u!?%~*
zlt-fZ3no>$#_RCw5j?Y&JU3sjE6{&Y3sn-7@l7$=e<tg#t({`pYm78c$3Lii9qPcD
zF=E9;mjTU)t@#@B2AFvx=)QKXb9s|W$H|RrO>JS7>PMy9zM1ibqqC_EZcY1}Xk@Cq
z@yjG=T&V60`8}W+okx@&#5PXuDu(24SB}#9RN}4xSqQ0nuOeJs8;ENT2B`k<042ek
zKVXI|Bj2P`me%7Dza1P-KT!<VjNR&>m?MR^3=_^IQ=uN!FsSNu#$2v!r-aNjg%;+;
ziy^3_0`y=OnNU~P!E^2HN4M1``Cu3>qldhT7Xl(?6eD2}wc$-}%u#d@EhZBt6FfOH
z5va8@q&DN1F&ir*MIQCT3%}uYr;A1keT2NK9d&MmIVKtN(%^kscRlg*wU@HabeIqR
zD`xitx>{;$OMJN00O)**bPN7ke>S=`OHky4;Cfr>)5#MPhDf~6rs<YA>D^^EYDsn?
zu4jV`BKQkyN{xDuE5kBSx{)FgbU8F_ZGL(hSOn0nicW-syWtd}<L~FNh5~EoASfh1
zJ};<J{<3B2!STN$(<6y!zSdc1m$Q7%@_pFYT72zgge$ls071FDQiB0fB}d2Q5TNy>
z!r{L?cH4#MW4Km6+QR7){2t-o>QVlpFi4Ak>|s@Q?C^5@uFdJF<Kd)Cfy(n>+LxSJ
z48HDe%lCR@<*uvT>!hhzW3|kB{o!8_neQqWp5VpTZInWazp&O{^`Kv5%%!2eYv~W(
zCrl3^W7+&y$Ckd1JA)E;)bMN+N5`wSLcYfvZP#_K55>No7XZD3wCUF0Wj$CQlHDg7
z)*{@WmETU6`96dtFrEK-r|^HFG(|~R=|S`99r$|J*!>lk$+v|MKajZLVeetweRL-8
zhaDk-8V<hOwb#C<Nwz+Rpx8gq(9d)ES<5wDS5hu?>+4}%*5z@n?#=H>Uk^GfOjxzW
zKd)aDfF}R2H0HfCTRp~${56}L*%7`k%lmx1P_XSx6`jNvx(Lddx;l<xY}>>}G%&7{
zMSbNr4Je}qL?5<SbQjMpNxWfdOFEHS|AH%l1QHpn(IheSR{qO{G<r?~Qh=xrmEVoL
zX}jmU;6Keq!R6m<65W4#`)u;^DDJ^)<<Q%nB_Rm{g9-n>ZZiP)UU1H_4qM@x7T!@*
zZwL_`H!H7_uWyZ(FzDXPrvJqFD-`tAc1An<N6&x6mjG$sPrUh6uXZ{DAc)xo)g{+-
z*luU?vY+;-$Xf9%Ivyu)(uDW=TsiO*mGi@tzx1rJK?z?vmWzCCm#LX#uJD&C)Qi>q
z4y>Ru)TmC*&u*RCT_m2LQ^Uk>yMHGt!s;Nlz%_2GLN5XP(E~7L*b*y@C8U(tvxAce
z_$%A6LqS;pH^cUuA1jAaN~fLKca-J4&#eC;A1>PgV*bkE>#gRgZ{u_`lKL-IdkdFr
zOx0^yjZ5ulL5fw)kglho%?G7lJfM>Yx0ea4FvP<Ysyx*?2;5cfR(_^@IKL(k0*)ep
z4v1cQ4W7-VSa(X0N?~TmI1eV&(hbFTK;qWO!K><neqVCTKL&*%k<Y|$!Rf%8sq-sd
z9H@%|FF!R~b}rNHoUWu20;aCMBaQ?(HL6+X(cLOkBD({5q?Z22J6MzEpDn1D&+<C6
zLG2nVkDz*yDA5M(Uu0Pr%pAqS-L_4}#p1UHhf0d1?_JB1t-;6wL`tP@$C95Ngt!*3
zgGYV6=Ojz!*`L?tMvJlLPkdXl0eV2%H?ot7)X~TMV^yg4w=>_bZ0ao60Wnjj{GiSR
zC`aW_gs!JbxbnVK?ckMU4dVbBH$bL}FbZGe_o}>C(~|roPBN)Ae%s7$Wfj0Q38TPC
zo0+(Hh=8i$BtRgoI7_X6e*>$@gc+$jaUWRad-X*e{vGzDm_m}>>rOYyV_s`5hn@;J
z5`fAG|7!zO0NC~Ud#lC5$cn|fLi!&~+^AZ4@4?wk+w~bXj4FkyH2yl~<yt&4BHHCa
zdb!~7ET&owI8Q&gn(a7jy(+57W$yG?2rJA(bB!+abvXgb5FnFmzHadax$d}yJ+YX^
zoI|OLzfV;zsDlU*+lr+aSPRO>fWn;NfUn8Y2!JgBl(c|fxR7Txl6bBu;Q-W0zEBLg
z34mWffW@WIi<|hpDh$VhVJ2XN?$3D;u38lc5hnzd`Gt;EG008ol+PSqfYZ`IhMFBf
zFdv`{z42)q%i$kOA2Gp^0EM@{`{gsc6MY|rZb7w5KFweis>H?3ok}ropZ(_o0Ah8k
z5`_&<h=Hm`4$W&oS08@E{+PUV&I%ZJ-zBbg%A$jl2v}c2>6aNh24V?>yPyVC=1%UY
zkpfu<o}4NK-1*>)3GFx`R6rq^e%kPl(jJ}!z)qV(+qMl6!I9uuQir<CM!C<OIF3a2
z1H^EA1jhB0o6)@h{6akhYDi!217L(QBZeN*gsU0k9xuBLO#{NQ0ALNvsxr0$LNRa`
zj2G(M9OcIk2yVs9^q*UjiQRfAx{bLh_Gmwci2<RVpxZxAyyLugKE9y*5&%y64IsP*
z!Kq50;DgkG`7apY1;m($pNGKsh5#r8u$3<W^W~)r%^RH;qQRy|!27`AV)#j@5MWS%
z!b}N%3xrM3^VpNh(eE&T5m0{4BqpLH*Y{lS55QtiL+DxQ(&QnnhP#ZXW&<D@rOyQ=
zwprWU7_ith{bqK;A2PV-(!^<!wT)4g+ZGj!@>o?SYx`5jBl6TZr9rjVxfRy=q+CvK
zw(xjUg=)AVq+*&VR#OYaKyL;|;`{*ak1^M(Ju;z$Qt?!kIc9?En+mv&?9ebvWF6$$
z4Tp$-ElnWQb-FbL(9&)I!MZQg_UiTp$pT<*;&jM&$3&4x&H{^$`9fKY@ab%|IMoL3
zm$-Dj4{w&cDD(q?i4AmD)aBD>Qa_Ok5<gw!Wc-kPCYBppG6-tPkAdirP7*_m;1&O1
zPc+qHzdP6x0%)%A<A@2i6t`7_yfio^0JhP`&g9H@qM`c|6lyxEghMH9V9U)9w1#H0
zq=>D?=N-x=dv!R@@=%Po?A{SH!jmr}OA+V7;Z1*4>|P%Xw=a1ujYZvLZI-nS;Bwp!
zq}iduFOsjLIY6NZPIBnJyU8ow!DA`07hC3R_6jH<S;OfdoZD~;QB`)&dd=lY)^Yjo
zaZmQw7_bQow;l|5Q5=a`_`(N3TAzAVTS@L?<t4t_%6IbRH4YS%|2TulihR@uXcoTU
zpn7%L`=$&=a+Pp1K2r8>C<El)@ZJM@1%f{tD2~fO3<DghFQ5qpztZ6*ln^3ug=v&&
z2SZ9wMyVC0f!2j9FplEz4+5gE`!4C`P!*Zv2yc&Ip!0SG0%p|asL%d(cKabzl$d8B
z&srmS=(Z4D8xB*!a!0rrYMQm38`_ubp4QrpS2fbnK+pJSG0VdYwaVdsO{FeMg)B!~
zaC$VHRMIhq0Q7$`(jZ<mq{Z0ib>h_m_rD?m9FjGp(nKh&fg-Ynf$o|MN6@QpkxEU#
z`IR)`ou4(ltt4(|FO4zP0!Ik400Wfq7i88<Dv(*aMrcTbEgJXIpqS5Jn8OQX1&b$f
zE_{pOZr}32obqGVtWZdjL}AsADC=I;T2F(1JXfwOg8*^nkyiwzW32?6Y4H4<-t~9`
zE)*}Ily%W?(>93HnJ2|~^LFVI%K<ghIi7J5J-p`nh)Kd&TOQ6rf_^if{Cc2|+2h7M
z*&s9AoM$I-|2mNbYnhM_^>Wzdk-l=-cbsVz5X81UuDQX_ad~7>Urrea3>Q^&bF8Y<
zV0`E60!O;lmcbA`-rxj)bpZXh-7~<K@C8Jd+RE5xMWHgEH#NS8P|JZS!k7lJ0{4O&
z+xCn&rXSGCX(84VoThlzmfcu8YVqTpoCcT?&fFXX;UA3fuS58tfr@5gsh=WLa^<u~
zjn#R4Z@A5cE;Zvb-7-!Fs!AqzDgt@?%!yq>;Dv_#60csEkSqMN|C9pr-vs&nchgqI
zdTiUEa$Kn4x<UmuNmg#%SKUU3%6BCxS4|t1tA$`Iu@^%0_+aVJnOG$@$=0ypfqT}<
zLzsf~Z2b;cRluAiF{+Y1chRxnKz4T7Gc8*BcuEJ_qZ00x&L$3Hb9Btzfg=viMCv4W
z?KM$qlOQi*da&zIk2tos_~}r46?mT;n%%X|zGhfe``-AVqAJoH{Yh@?nn{CLxXn)9
z4mVod;NK%+`gi}_^f0}UcZh2b<~f)cG%$rY?-5KPfauM{e{7RfH`=Y~-*3JCRR9k@
zsC$^lX_CEXh_@oF6$r#=GCJd80qPOJ`q<tObEnEvW&*V}JGZpR8v}aRxirUb6W{mr
zHQn|+fp}TudPx-*y;T%`6pcmjJf{H_HeUvW!ZbzyH%U0yTR`)iQt*TlD8Bn}k?WxA
zrcMil6B2&<Fxm!@p6h)7JIt#S(RArkX{d?qI%x=nbnmM|k@@f@I9~a#t=PXRqshr&
z1D7Xp3)b|ms3X;<ru#&%N<+)Bh4|N7XX_?G^#K%+RL1HiSFL%@&p37g*nm&>mvxgS
zc=~p?8W2YF*-Wi0Kz(JR=NZySzJlPGqK{sv$`a6F1M+!G-<E4UnUYR{19qHv&!d*!
zMpU1H)Nj$aFtpD0vlQm=IanTU4lZ3)^7K>g0FzFX*z)AGL`~PT)-(&~wV1obeNufF
z>@ibI>Qb`g{6d$-+_g$6Tn?q~q;I|AT;Of!v=!%4(@Sj?048Msc@r(lDwoThAA%MZ
zo^RLW`AI5jC0<j};x5Y+VO((1|7+nLL;`h&fS=>J(q_LE1lG_R&2`ucfIbq%HR5|Q
zq6$q&zk_XXTqhJ$LdyMDz#8FYItnI7j!3@{CwIG*gT-<It|t(Kx{!Uv@A45b0$|PC
z2KrRJ@<u>OAYD4v1+17bZG)i-s4*q6q26d;HeUn<=W_&J7YNkl<aC}yPy!sS`AcX@
zAYqPWG_crO8c8&JO9<#(dV{pLZht9;n-h0&xM54OX(X5V%|d@~#;gETo}Y8{N`!{c
z&k*tt)sMupYM5Y-hcp5jq)O4(msLWwweVYeUewpe^jX6i%6EI>g>q%wt%L&FWyKxA
zbE9&WBaX*6A~gJKe_|m|p9n=tL`5^c1VGIBFHFz`(y++$zhl!s7DNUG`8_B(IY23q
zC*T7i-g;xO3pUP5(O_&shv^G1ja9Cj(5o6LU{x((|F;YZoM3Zhcphxm0!2K003QYi
zmkL}(6N-?&=L^0WJh$sN$^`-7W|l<q!?fYtvkX+>(CT<wbCTz1T}Z21?KfRWL`%NT
z=nl;@15*>w*IS)BymL%8R#DXOe<ipbe%+|;qIn8V(_qyNUJS|Ck;+dk*$L=1b+>q^
z$$gmdS9@;Pc85VN<|n&$DvGW@+)_Z3V-xOL67Zs~Oy1S4>+G)p>B~R(o@;#{azBl!
z)h$^2)N{#Aj$>WS>3O+v<3L+<Vy<S>A5GU%NNbcP3+TPZB;kzmf~h9!dB7k7@mnK8
z*E_S#fy)wMI-kAob_=sv2S`#%BZ^*?Q<UXtE>|U=Sxv1%HJ;RZacF2@9ukTQ{YzZ=
zZ<dt;5*c`U(#hh8oKSQ`X09ZZGn+d_8^@~9YxA_g=p+S{8|{X&_l$|toh-)R#GBW@
zXaAc4^S2cH6Uo(5itbIi#haSvVo+W}VmsZCsq7HCHjrhOwp~wvT-|eR4!x<I-}u{P
zlWb}=QTtFLTULiav)w7!(1DER7O>$66vmHJ&sdWpsLk4rM-K1OeE|_g`3~nlwf@o|
z;eKfHJ^E^>do{=(T6<Ng57x;%ntAUVMnQjGq1sXs*=T{AJs1^-MDSQ@jXBlfBy|9K
zn(HkeC!%a~#8f=Ocme~-5=sDyJIsaK>}{gQV{r=7b0?7LPR!;1yI-X6956yrNJ{HN
zM4j*DoX|t>wNgmY&Mmgv`B)nirR{c2=%Ltg;oSSw_x?_T59?cXmw;Qbda>98VEtQV
z*|C`hX~Y||N4L4z?3TZ}*Y$mfA(Q?e**z=~A3M1Xvc131ZNBS*CH9*gE2C}A2|<?W
ze^c^G)9kw--TNbymD@wWd%7zuti79;(p2!e8j)S8qZaG*&s7{UQaAV7OkcSHx%zFJ
zX-M|H@<XMX<tG?z?w=Q6gyBMkyjxDc9zLvOd;WpB3dlnI`U|Fz-1Z;d7mmFo(YY4n
zyL-64@zuJ#dvFDrdq(a@x_PWiu7~@Ci@W6)x}JAvwQ>KK1%{0Qlb06xE<^sW%lzwl
z%rG&nunF4Se^ZqHZghu(!PvE`FaGdvl;^KUQ(=JEzccs$D1e)Yf#r;I@ng|{weSwM
z$Qd}v(EQiss$qLZq64~*|JFeGtA|ZR1wLirrTkCaz$Bdg=Rmfejpmv}_g^hw1Bf9T
zKNIUe@CGaiZMOo`@mO$^4DP=p?{5oXKs{0+QXcugRLS)?fnJSim-g5HZlN1|N-bF!
z_#ZdG-u#90s@KZi|91-pU<705D~tZOanOJv=j>F!|91=jYu>*O`d{<@9p2d3T0b;D
z{hJH$zmW`L<9{cLzw_Jwo9AIQ9n2M=ZSQ-PQegg<+9<07=!#Vgn#Wct%v~^y3wpcI
z5cpANz9f$ZsD{g?gB1JDM<^WRvuztZ^kSI(ReQSf{`WJ)qZ)t$IA!tGTt*YH{G52#
znzhxMI;FGJab1whaE4ggQDhVU=B6cZ@bqL5fUeyvwV6~CNB}X0PaS&Z(2)l211DYp
z2k>V4w2_nr$wnBIZc$;jsc5>O%@iwcy`LCeE1BANDRrODqzFr|qqo?ma^F&(oyLL&
zsr>`BP9dm>b~g!(cIaXB*~e2%CBNZxdY&St8N@L3lf6dJp&Sl=zEk&cRDH=ueAK6R
zd2JNZIhLL+%T<IG--b&A0M~)Jd1hCsq|e^!g7o*8^sA=JZ=a(WKuKJ%T1y<zK``l=
zTJWm!n!-%=2RjC0*`A+eI*yExN%qXHo2_C{44Txz8^WeKQW+lE;oU$Ju2DLDn+m{K
zyJe<u^1Ah`Z62Gv#wPvTP$}n}jXW9%1rw%_5L!z~at;bK&qjUd{T5hbsxhI8Lb9H>
z_M;NF8R>;_%y9lX<%I%;$n~Ws6ae-L=af@!Di&~Lm`cF|qU<3{rGoU>&^0K2*kxMX
z;WK={B<*U*vm6yn@<;F|Ad`8y03_9Io<vFJ0|v^%xO%dEI`!CzBIWv`{FzLev3k>z
z3lC6%al9w^PV>j5380d@$xA9E&qJmw0H<n(r%cC**@>O7X2DueIJv;QZlOjS+Yw^N
zy?1f*t-j@W=i&xXYxqe15@y%z%o6flm}5Mx00t#7bh3zPXLZ6~UY7snSL5!w^V2Wp
z7M9`sUE1v?)jWhzKwx^7d>#}EO6tW^7?ON>&h@L$VO^t#(|I?!>ti3PgJTQ)8=~hS
zm6EJtrjfhZkz2$qI%DU<@`dZ<k^4S-oVAB*Zed-U8p%(B?hSVX3Oe=$=d&rK&3v=A
zZBB!<%NF+E7g9+FGz4tPt@`0e)*&z8#Hc$eN-5!R;+?Vl2{rdue&%4tnQoLtAzeN9
z19NcLL=FXw{#_3Rqc5TK7h>9=t3YUR{-*oEHMNPA6S{m@kl~7RZk75A?O$?K-=0g-
zsOD7>)BqF%(ZKx~$#at)Ys*)BE~xdLL>#l+G)*C0Y2!M?4pk1@JnkdCn>jPP_FF)6
z#tsYcIZJRF@dgcOt<1P*S(ufuHHiCyikmG(7%ccVdltt9_HaIMt3C6K<A*)|AvG$r
z@11x@c(z*E^_`%Jsb>}z9h=-Fg*4zdJ<O975Sc*4ZvnM2{)<M(wtJTLO_zF#PPwV0
z3pULuLZWip$~JF^?2~Ug=^4Qt*pB3jsb>`#3xxWjww&o#O}?oFEVg-&3_F%`efQdN
zGkl8>RY<IL-9pb}WS3<0H}+P2(@<^CyNBI%xk>hDLrz9TML1Gj>YZ5k;wf^3Z;}k?
zfCl@mT;h)$g9o2nvhmcSWWEwiu>LWBiB-1hmNH2R5LA+m?-uy;LZj2?RiAy+k_j$p
z(rkXLT}fOhr}~2a+n-O_LpXBeFHm@djYzqgUNt3C5^U=hGUBuDyDYd+pPCk~=Q?S>
zv{tpp98c+$r)6&)7fIWU6fp|NEevFEAOF;C#NKIikDAwI^4^Td5HA?dBRS0N+L7xE
zucOK!n!74ZKCTKjT6e%|1={Crp2pY+Y>BuiNdXNt8K7aZpBm%YZOM$USjzgGxa^@q
zk^$#U#0Dd%{56ly+?Ec*%AU23rwOn3+Hj8+5iuV7p(btKMk!`lS(_(0J63hRKv!+5
zv7_K;Bg0_za(KHxp<(TFAD;NdtdG2wLufl8DKimfwsG3j?^*O}b|8$hyeX={@Xh6U
zz}rn?BMJTw-_se>lAQH8D^GAlSybo4u;}=;JW$Y^x-{q$OWhedMg6E%?BzeE5=^q!
zBr6s*K0&XktwU1ebL9vam>+dAoSQDme}V!CA%Hn9jn#CtifTjL$EiTQb5vrNKT%;{
z(PA`UXkI7C>9a;pW$Dlr-`gu%%dJLyE}-=>xVVb#saO{Wm#WsngC!x4`T_?bS~P>&
ztQF^6O>h?(BU=eHXo*Q0ODZd(Wphg-d5-si*IDEqJTyZE=)j=(@fhYbForx0>cFR(
zek&u)b&$HPhPW#q&xtI?vkW@>6}?!DY|4z?&M^njvlDq4<dii4z*tYOprFaP6e|)`
z%M$IFgFSykH!CfnQj4Nw*g$-gkW^$cWVYx6RDQ$*FS6DK(-wwR8ji<(iTlhv79Ry#
zXJE}?q@GkZbV<Qu5X|@_ze&iiW+7}$=rKU9mMr7gIV^xTvqYJOiZ{57B9ffKSyFi6
z1N|QZ>Y&GXbSv=UI~fjebVF4Yj*VH<hZii>mEP{(h8bJ;<eQ86@lmaAST$>v&h+f<
zPL&a+U->8DLHRiiO%p^*Y&sa5{QG}=7&yb_XM7Ci@E<)O+~eq3!t)gU^{E`*KO$ey
z4jkvl!|Y?BR`l3%&jO1T<vx2zYVL<1p)bhW-os(~`-`Iq#osxecJ>J7^o}2iMu<2i
z5@0}#v3NysHz{7SZeRNcs@<kazeJpV%!rQL(rGWI&Z(cZZANH{J4}NUJuT`~*qIAA
zU)z;fStB`pm_3b(QSrVjXnMy7*V%tFIo(nn&-Bi6`bnv|2pJbU-h%PoLSE+({>CSI
z9SyY=eO!u>iUC?|YZ?n{DVj~~`X}v6%IS}pl*zoT!$#5Ww&yc@$Z#H5$v*#lI<dbj
zCGA^K^E_EJ&Qxs2gx!)N_~~(-VF>@+B=e(uOT^b{U{4TR{h=2pG@2PYkDTWI)8Ts7
zj3&CCTLO53L1pmV$R*g0DQwFAW_K9z$pnRo1oyE2a~<q>I#SI9e1hjHh2)x4{=C~o
zG6}O7p#%Pg<+yY8wZ&{9&rsAGfq)_>Qb}Ow_9+^ot4I+tqyhn=r-D{fH3CQHuJmkq
ze{*3`^rdn=g2CvskF2A=7ru9gU4|2?yZf81C0ehmRcqHK=SG>_X=4$g9b7x3PA1P0
zK+q;TnVay{q(g`oRRqpt)@Y|HkS+tCs0(r?ogflelGP43FwFhnrduysCI6uHNY8W3
z;)N+{b=5s~-Ak<GmvBs$DhnyGpR^6k_4`XvBhJdzQz|%=%1-L4QiIi{)B~TfSsjm$
zc#l-MqY|?0@uzamoiexYW7@5^kFc>nP77HN!elnqKb-TZ;^vEDga7y)tw-I0i=GvJ
zQUg&o7pFGYiH*>hG9b30_+<~f{|h>yVTyjZhTP)W2>r|Is>D@1K2y(Rh)MA6(oEG{
ztzpzNYWGD=y)*aZ-b!Nx8*)3-XBc*-y<5TGdoio8>=<hdV^dhb(XObTs0wun-}<;n
z){(z-Eq5<!+%z!*4AxUChHafr3>CNqv_i3{Bln}&ng|=7)H5S82=)7)nCrjWHOGxP
zL;q;9w$aHGA)DxIw}h76w8^W#uE`ZiTg+W3lpbiq7-(EmeQzugD|yGuiu|&c)T47P
z0IUUo&HF}GrQ4>(kE|Gv6XRriLXBTuqkfk8_+e%F>U2=)%Z{Ls-w0U>QkX9#-rHn?
zbkSu$D`btYJTHhgo>10_YqFyi^bQ-ZJ>A24j6Z^_Tu7vLX-314nv40z0;8p(44oI{
zMqairKE}1DH-3t)&WOxiS#%3~<B8BPpKTkadeX<Vq$J<0*F)jq;gN3ZFEDNS32+H$
z9h0)0Bdz0SyFW#?+68><?_P+Y@HL_#k!w%15tK1_nO(3i&cdEP;?q{X=@x&$mCLB?
zYZ%vPriHwVHIpsZJgOc(3Acp2RAKJqoy)}lFOe*|WMo6jD|wo1$IO`%V1E`*<v5gt
zKj-|Kg=*D=oTV4vF?W#+!QZzxPKLnQ<~Yu%fZOlqy-IJ*Q}Stx3SLxQOsV&bTq<cl
z*W=yg!i(cnZXK}UH|g3Y3E7U6RPya{H<%Ww=3h8>zxg^~0q{&jni28Gi8byNNSTv3
zR38wTFzf`Hox+f38g&U!Mx+QBbi^Ym{OK@{ntqj7gjSjom!<O%qb={XoUQnJ@I|Hh
zEfE{V|90Y4^GxL@c>7!f`j!HIi++EI;_C;yp^L)J{8nmATgEZ*M%Db#`fx+qUoFpA
z^6K8<#uj3JwwHJd*>FyjrNA)ttArb4)Ax<z*Fjy%=Zi59Ku%jf3P12|<Mjz}h_4-g
z(z9s=35xtR^GmkxwN;TG7O5CEEs5ODRz6pnTK(qD<ef6(j6c*3%)A>LlmyDsjVpnT
zUofGq9=ABu&SAS_R6Y)3?b-d0Cv9@ElSMg%12^pjE|B3e)dDd!rte6?;HDQ>WVoV>
zXAq2OIs0v&crf(W{a#>xPfoc|FH7;YpCN1luE8)PLqTgP&K;DgR7s$EU^a%!zNagG
zlk#0kCPsG4I=`MU7pdQ-o+*Dc3hny!4?&fx*vH@UylZk&FiS6>fI6X4Jk@ZASmCjQ
zprMt#QQ64ECa>lE;~MdgPC0d2`(?eQeK-z7+<||LK7KF9*QcG#$RASMk0toM5?rv`
zqdLNvn~NCUx-E!3U_^QsvzB{KHhoPsx@LsWn&-gsk<v;X5b-??ciR_pHb5rYB!hwx
zdpD=un&f9f6DMgVZ=Bv7-G~&pj>?=aW>=7fDQUAb@`IPm(R$!lI&WQ0eTE!=e}8j;
zDAeyWr*NkmR{Oq0{G*FE_Dcg!Zw(`j^42GRWEzD&I{2Zl35#X3lyq}-T8%!1e#M^=
z@~lyi-oE|hxRS*soa3nyxGVpO8X40ka%|38pFYt;j4!IcNoVqOFK*z<%=sGNnNgCc
zX2}e16q!f;b{@*xhVTkmg%*2?Vlp+cQl-F$<lMZHNf~^Gf$(nUefW^LPi))>n>~l<
zXnmbX@lFLxyH_Bb8)R}xbkd?`!$+7qRR3CPkNY`|yigIn;F~hupLbuxFR~P^Vi&*Y
z9PAr+c98Qh3&Ve;tnk8&(LcQAD?voMphzei)){c^?P4Oj`n}Y;ckSf<+p!3*PrxrV
zL?)=tHzM(w$K$eA@u~p}g{q$(^5jvWA=dM*yNPU56IwJj9&|d6A)mZRX`T9*<uFwF
zW>6h-rB_9xhvvFHn;1X$7JpEDeW=QeoKXDkXkPv;`&|-?B-XB*QR)}p!A*NOIa=up
zF%R1be5Xd^5xo#JwUz_|0e$TsnT?E8Kvkeki*?Q9-A@YSe!}Vx6bD7z(2)I5)=-EO
zR#sR@&ad~<HhKli$-k+QmVzhysAO=(XZ}nl+ByplbJ{4{u5&7|{~5wee)rB^Qt~y6
zhE|MsgDi{n#NovzH%`<LMZzFq7tF3Vk8?Ii(Qs5><B)Ng9B0KMM{AT|J!Y2QOn&De
z)?Bd9@pyn;m{t`2Ta(}M*807g|4{CpF<;c8M(V4)rmsIdfmvNx|7BM67(<o0HNjK8
zUP$Axkr{faY9vOt!fY842O4tuA1X0>h{d`Jq_T74UMBie<qs{ktpR8YFPhCOa0oJ-
zBk3Lr<z824uI^r6Vi%P@iDNGd%^t+-JCUyKA1uvhgkKAEv;R}kN{q<*c~g=t)6H_b
zH3wmLCE!C!Kn2PikU2tQOCF~0M^lzx#7t<EZA#8Ta&%#hZxB+UV0VbF6+2`-3;5C-
z`gZn?u<qeq9cEDOiK{2+L6=+v_4kN{&&w*o1oKaCEto<Ab6T&<kgtGREQP4di}0!O
z(n}#ytO;5~@zEc%VQ-8uKE{hV-RR3G$<F6?h@P}9pO5c28QBu)r27rm)Zu)zc>CFE
zZLS-&Sr?gKJSwf@U~6?r#oq|{;k{dUHx5szD^fdMToEhh5=va|hw0g^WX86Z{`@si
zh?<<h(at?ZFF{IAJ3ss){vF~fbbNPMiC?W<QtY=4gGs3Mu~A&42#>nF6~^3~E^$2>
zL1xs@?rL*Fa>$uB%`zl}lEVMHXXuYOC+9=G2{n8we;F*b*3b~csYCuw{~F1BVXogs
zM`3yQ+DVb_8n+R3cNL^FR-B-G)P%v^GBeH_GW1TcUUNKnO@giLMS?Uuk8nZi%ysIE
z4;>%fso(yRZ!>x}najDrj58#6`=J(jm*z^AluYLxS=AGFu$H8K*soc-p-~3Zkdh5F
zK{xwU&lF0$!OryHy^%RwXlhrtl|WsP{<oxOI_H4VQ&2>+gL~YJU^BePmKjeJHsd*v
zUf6u6{5Z}NxdDN#vDN(D&%1GJc0>l*{3v$2*zWZB%@N~oIW)|}zm<2fxQrIvGQ!_z
zKT$C^{sc$4kAnGHRNN!FcDx7aiuKA@1g^-&VE>(7Apzn0vTrmAVf`x71TV(_j2J<#
zg)vj!2i04By;|hQ%zvq>MP=_S^}M$ud~^~mP;h5H=diKmg@NL#s9~07qu}aF;*>?r
zskOCM)EA^B%G3D<ile7Lj-~{qR#*pqeIJ{hm>$$dd^GqbRuC{+pS^#rzL(-P&f?T?
zDbvO%Q?zQGsjhpz@jRI#Q`t(-4`EP3?~44@D+F14?ueJgFUll4N^2Rmg<Vi=RJ+%M
zL_{z`8{@mNglXM2Lq_Lbo*y7krQrk7MmI6P&pc0eg|Dw(U9=pO9mk61l68-P2Xs2Y
zr;xQeugl@0YEDYpL8PgcHr&n@nm6yNB)_Lp*;)F}@05sc-y)14>c4Q%8nxj^U@>7u
zL*O?qEWa2Rn8E5^{d&NPw!XCGK_$sLlUnaKS-*U59?8O#?CNO}e}K3Yyn^9}#H8Ws
z|C&qN`1jLhQuPJS8ph>e3A<f;>p!7lZcgfIy?Diot$|fPB-v9UdU55}(aeg6(>Klz
z9s0eB_PV5XAtR9M7?zK|2n5#r@-L5EM1Q?i6)`pLv>65bxP|>jo^%u_OW8q12e;af
zXnR7)AXVf@J<5~N$`MW8=!N*vPdn00p#vQ8<UgFUnn@{B`holb4~ti^uldBzh#e@P
zRudPV`VvA^7hFd9OecKYiskBaW$<gi7Ji^5^3H-40+|<B)03X%LBsH&@PCh?Rual2
z7{E>W!D~oXpA4Fum*&8UrOawNol#j`ACDEw2lqNfRwI*V7*Im>1#F_N+>wz#_&t)i
zRp70LM^hiB1M;TT+R2u<61NUl)u&jPu{Dk@vR}l{$BY2Td6wlxsd;5O7!@X+^Bg>B
zk{0q}b`1Q1RQE$&xjUtH$IMM0K{9`0dOF!hkZdhFKVwhx9ZQw7&U2KXG3|q8Sv@$T
zTGBa*|N5!@|MRCt09N_hl)mFw^?Jo%;~m{EjR)Nj67u7qk)L#;e&ku?KK|Ni;14_5
zbPfqG^YrN8ZE<0Fg3dTOo+J%FgCnF8{-=JAi7Xt@l^@-*=XxRhVT26L82m^@c<cA?
zSI#-Oy5-6fr2l*Z-sD(;4~M1=ex&^SYp@Uh_nZGeuo5c<bp6nmyTRq$x5ci=el-Ps
zPpH$ELh(I()Z;d-yR+{JmZp6x52TOXZ|~3xr$=9TK`EMCrPDcb=fD@d{%}z%+>O=>
zW@?K1y#qzA)WRgKr&Kv`_6W_&ku*^akHNEuxqRn*HpVL-?tJeZ?}qjrQ(F0auc#j4
zgr$))nogIlQx+a>%O9ZqxrQr13kooaZ;fvzOru+>@1a{s5i@<Nwu_Cuz8@*=yuF-H
z276qmg?vuaQd+Nu#!3~oRen~!tKOCN6KpQJZ-75@C`@e#YH9`N_9VU!U9Ur;Gu3XC
z1@4|$YJY^Mp1wZ?7xBGYaE%dk+`I6-EP+4x^zp6yDtNm$Vd`Vy&0AOW+-|z~7~ywU
zf+v5$aN45+{JT$Wc$?M1pyC0T;$i2sFv%6ML=O$wN#6H5Lig^xl#OJQzgb-i6REl8
zeuBy8Op~{c@VS>t11o4HVpZ1sQ%bw$E=i`=jhjl#fm}8y;Jd&61$v!Ca8iPVjO2gD
zR7b*d3ITmx#_aF6z@BY(upd)qnrIlX1M7;66zJQ$PUEkdg=Pon_ZJEt5?3X0-^IPw
z+nE~BTb;_Ho>G(99sG2KInMj&t4pFIk*~Mcng5;C@5T5?oB}d`1FRhM>;<^vP6_OG
z9JXSeivk6DNCc0EIWF=S>ZA7i*AcG<hTAF-Ud{Sac)u`sP9Y30OCe06qlPHk1wQMh
zA2f~n61r{AOR-Abj+`^<lT7XtHXk98AWQ|ixVzVdqT?Fw)^;>#w82D^upQIZ0Jxe@
zj7P25s1?5=um0MZITwPe!DVV-AQZ~CsR(z`K-fqW(@g8Z*TA>a+-rv`SP_agx~iSy
zeV1<yF%agq(xh!}vfWBjl|NF|Or|MzgZVW^xMAz5{G<7G2AX%6x@F+;r<Sx3lqupT
z?7@OwB0iq>bMLzy`Q+@szEOps{NAKa?M5aYRc^aj+bO{VYSq-I!af{nV6Hb3yiFG5
z9}O`jYIEPwQrW%6fx`alp;RoB4#F1B$oJjs9>Rfc{uxgCkxbqiKk`}s>+i00FSlA~
zwzo6ryULeEd1bBd*SAov!^WG|)$^5g1;M+486lFdF-eylTT@eAQqcXFfo1mxl=bRS
zFCYE0Bfw!-r3k6H=gBlF^*!had6i5Y-Bwx9=6$ymv#|i4mUBJ4{o=Gv1jFn2>>SE9
z-Q5WqK}=V5CJYwINyTozr+&rfd|h*<AXI;G?-@XPd9p>a#>a@p#DCElm&A976tKE{
zbI3w8vLpZnHeB!MD*i3ebg_56S6tnT=zTslx8&&}4UZF;nHbvr@P52*raXi9j<7_d
z;Us!nhx>T{jGkA)7QU?8v0jeKZtDakM(B#2cR|~ArEuw`_-)!4)83NnBjob~K0W^B
zt0AtKwn{2SXj#*t=lx~+{5l$ExuECi$9>x#!|XOM#|rn4w^z}b^u3y8XJbpw7v8)f
zp#Sx9z7C(ZuB=(S>(9e?x{mhKQ${<qsh#={vMvYM`lB>wIGHjNXxBkU+z-1ZnROHp
zMKFG!Wr9w}N^h*LO^wKIzw=%~&<fj@xsC6+tCrIBKFSRtYZAKhxPek<x9%R?tkz};
z-EVzD7ib2-HM4SuG}#5dh87@fdCHcpJ6q|vIhS8<K;D6ag#gFM<9oG5t!oRO0!hv`
zr}3}|?l}Ij$?HY@)p>BcrzLm@9l*11>hS6_Vir8MlMLh^fuP)LcoDs~T)q4AD1f_L
zwhLu^B=i<|!}W0PY!T(bQ5-z(OpP1Xf$_QBY@Z5;?U&$xqy#?aQ|e)C<!#6R<^uR%
zJ`WI<IlYE?W`DUF#pe3wQ77;l@Y+va9ctk$aIHx4<%J7;WOzGQ@}?CKPWz}A2Yg4D
z2?~cxuEMwb%if4Cds-8O-e*TjEh+)Hv4|udYCqczBighDyguA?fxljYSPMo85}KJQ
zJM*2Ffc-WisBc9LCBTojmGotOs-V`Q6&2j&Ik<~~fag_H_tVb(lHCfKBC)WMyZV#m
zQ-7lUpHs^NnOk~l9i4`g=M3l_YvqgXA=6Oa4&by3)*8<-MsK26<U9kD&ZJAq2<P&I
z>y2IO*MlH{KasW*AAyl_h8jqG07eT4Ys$K&gVsafhe0U>;H`q;cd4B3;Jut=g}}N%
z(ewp9K-h*3<LbJJMBiQKC-Woto{UGNUJ?pj0s<GEnCq7n(n|cm$OrH5JV~-(H)QW(
z69BvYpAqyBd!ii58xY;UrwF-Vcei?Nyz^FuNz|`!5~l+lU)uZrc*HcO^%~fri<6;!
zH^}i)SpC=y3`3YDkt;m%W`?Nu20Av$FYHV2IUX&m@g1F{*g|S^kxtFTSiIe8@6mKV
zoU*gJ-j&+_W2i-h@(j1D)wPc}+u@fWAg)6UDEQpmJ~Be7{63m64&kZVJLHEl<%Ra4
zm)n<NlP`M}o6^2RHW};~hORTIsOd@(3Y5PM9=|@uge%dO{?a6~k#ry5I_m#Vyjfjj
zY4>735g~o1**f21n)2K54P=c*V&(N6=w|4c1iWK5>D+R?^9U{!M9jP9>My@Ex1o;%
zyn-Ix*5jMEY9C02Uy?_9yH1H|*0`>~wOB-#SLi=oYq@!j9~9+Q(ke^nx+LBjPxQ2V
zl~MFr5LYsZXx!5w^7t%&N-UZJ^s(t`5#ik@IQ72=p3j9|erJ1&fv_bgGSi6*1=D$f
zYEZ12ZD!xHY$kH&#j-ED#O*sii_5A)2HU4c3pN>J$650EuUqIy8U}{hXPd8A#n25o
z-X11&<TVX58hlDK8AgBdb}K58*n)^=3t{z{8AD_G%KNBa)8zLp=n73j-k@i4F16j1
z7$Lj!u)$P@H<P8!n(_`KG=gj3_~9*P+o`|r)=4(rNu>C+iu;QR=Pykr{zEmRl9zqk
ziLg1inmg~s-dC;tuWQ<y^a#yyRA@q5VmYaT{F=4}OC=lpQ)HHyLe5XaWJ!A(IEne2
z|5(`Lgnu8*6pz)^YEJx^_DB^{jkBW8gL60-YgBgD$n{<UJ^m4#02`R~2J<yMnhN=s
zj9;-&VCO!U#aw=SRa*=5Tv^N~Qzs&AA;T{c{KREk0)h3raZxKy)<;{xZ?%Z1r`I8J
zVOPFiLUpg(xll051m@`n>v&n^yNNPeriG|Yk$+36QlDDjm>#EL&~CBBQu`ZRA6d1e
z&1ft9ZlnS78sFJM@m6%!wTUYBCoP{!@Z-PlexcO8U4d0DkORa3GC4g=qvZ>_bXpFg
zeyL8eX$>Q{<Bl(n1oci?V7A@wATB4lH5f%Fxw)dZYy=BusKrEZrEEXMk^AiA3uY&|
zY0J6q;l>@`6w2p!;Xpr2!gckWuFZL{sS2;uJ6jitO_TflL1pceX>JNx{xzT3A4|%(
ztfk{0g9DYofhu2gH!WzN9`hFj4tPSerNUkl+Qol%H=B~%o7!D|dokbGgX0K)bj{5p
zO*L9$j)@+_uRp|hdlQIIhK7h<{-?`Mi>;oz)uxL&;aYwP`!MTqzA7YPFO18j6WOp3
z0w(4CrT0wWVIe5t22a6KqJ5D{)k_a4<TZK&0lP&`=cj4T3b$Z(Xa$QU)vpLokgG+|
zK$gC~cH|u~!t}jAmUC1?<Iiv9QWNm{HLS_1<W8dGy?w`YTCEN<7(;>b=wuPhiGHp2
zyPsX>3?Hy#1|;3zKw5U$Q-WWdhQ%Hv52`B^HrVgZgIO`T&)WVv%KiG;WK2_<Pr&Zj
zK7CpDHqSI<(7|_sy>JrW)D2t`{94Ku#0(Jd3Ba}j`vN$`mIycx${vBtr|7kRuzfy}
z##&+?(Rn5Wlbnz#z9Y#h#XbRgf=-0%u~DJfyr1el7q_`Hx2nrW2iI$tr@YlRZ3@Pc
z)kOX^yG(w)&bS2gD~Z={eQt)9-;9jI;*g9>u*C{HyHsdh?v-?RZ$vo>Zzg|Mo#<6O
z76zMdv4;FDx0zN>|2W%zQBzwGW}$vm9?DgC+gJ!sglRh(<~o)Xvej#ku%goz5L3nZ
zFc_9{*z)j-#$jUfTFv0%!>Vn$u03`RBM!9YrGYYiv@s@6&|T(5+a<L09pvtZPD;r;
z3jPYM<Ix-4RpXcOuXd>*wj+*m*BNdXBSbJ;CqnT16NF(CEBQ})nD@4>HND?8VM3Kn
zG0bed<(o6$eAlNr-qCPF-<cp>;JZ~X$JAjDdOo$6Zx)<n&{UKdiXv~#A-})fOt-ha
zWb=60IvjNimLK;)i@s6Zi_+rb50)4RLakwnThvD1R%Wl)^UJn0%XKa!w{|Qd)K&iW
zxDB}3eI?6XzX%h`WuwhzYvr8u!^g1LeQp`dvi**-4hQ&wP0f{4S`O>eHK5ONu7hQZ
z?tQ@dl7co95=XUKVNCtse%V@a9E|>U#@~~)NVx=aU3!mab0;!rZDruuTJ(ix)ADTb
zOiSDoe|0W45V+N6w)W&cRbrg5(PXeZ`=@o|CJCEqokgXk{C!)jGYGpdofq)O@rR^H
z&tK`Zus5&De~hm;fD6BbgkLuD3>Hn-$zTh6^Bbx}XOXuSt$#i%LieY;Ix68OI<}Ya
zs=Z>X`~DYikk+e2C}M{Lnw-wxso*5MyuZA%)c6Sh@p!FQ<m>HQ%xo;Z>!Ap8$e8wZ
zS<!m-SrMB*$kaUb9^!V_i+B3|{B&$bsXHsrkmdTm`c#WV*IVxaledo7^U-w69&GQ<
z=9&sduDqZC)d^NwE|njxNdZ@NFQxoh@AQ;zw#I`))1Fx7d4iv;9>@V(LVT*-H}hV(
z8KWBZZM0mhd@76{R13PbUIsZ1NtQ?4=*Y_#k!1c=7~dkn`U7Wxw#;w%*&1`$n$CZO
z{zdln2$6lGSHUHC!q~?AoUg})&>-s<b82J?>Z4=6>|2N`^|Ho;*&j2bfIpa*DQpk-
zAW=&meRJieFM~xx2zPz-$9-$?*Kgg?+na$EZ~g?e7}*S&?wGcRwK%ap`I5WuX89@S
zvBHG^oim>NJbXA(Hzr1;j}L4IZ|CyInonoiyrs62O)#JuSZd^;Ws25l^JLLS%EOm8
zBoL>ocTkY3;YbiY*xrDq@2_+pKn`qeL#CZKW-zhy?*zDS<?a;r&xbujN;uvpDf?2{
zJm=@ved(%<Du%wd@4ZxLpS1|KHo9*@Kq@J4JuFM~aHhOF+3=M_dDtJ}7?!O)Gu2dh
zy+mt5RrxA~yEHT=&kSVj<rky(>*Wdp6%R-t;r@<%yM9urb5j>{_U0JepM073?XL)G
zS#2i5?7(*AHuv?7Lo3iKrAQw;g(3$0>Dm)>q|a+uc7>_j=2`rLk1?~GnF}c3ut@*M
z>#c$xxuAg9mM}cX8~^iaz{&|7!{tJG{Xf1*>VFDCagxmlmSO(iS1dh<W-a?_=KoJO
zOBxjFd$W)9@Kdvc_-&3E!fZ>#G`ahH*YsEd=McHS1jZ_j4#LN+G-UrhUejloBnX)I
z#_uUjMB#(T+YhwAZS~LL*ZPdMs}my|$@}GVJAN5vriY(%mJh!ynGd}B5FzYhSfZiz
zM~SG`M_@j!ri2wP8bViRGq9Viz5OMW$z3()3WYp&312hTf7rzzVfGp&aHvF@^aVU|
zPlKKb*$GB}ndAHJv99cy581G`u1ktk#cF-d<`X%Jh(KG7HEv;l(*L%lR6Y3%+y+;#
z1N&zQV{yOIY|OVNV(_xB+m)KxPZ`_@K^LecZrP`ATeA?Vhz!p*R1Y+&&xV$Keq<r!
zWuq@6y$>aRPvQRrv)?9MdZMkG%|IkE^$qf{g5Q#-kIQCTe5vj|rxDs8(9!w>ft>cU
zG1$Gms82JW$AVoqSq&LafEup$K4aC502wFvk(9mu=YU|fzxHq&O3s#ErG}r2VmW3Y
zfvfeonvcXwXkWozrMdTfb^C$zAFIK0(96+c`0w?q`O`@#V`1s-QX}*jQ7Ebbg9KR<
z68<%tsVzAi3HVMblWQ+-`XkM@*?qCyw|cz59y^h#2GR&syBX@CZk56RGm61`|L+Lz
zSHu;&5g)2=-05tn{^xl3C(3C%r2h+HH9~ks22m#JKz*BIhp_Ry$u4;!rw3bamnn-y
zmejuo3q%Kl?p9WSuuCRe#6CMMU8k&cP7chvPfAHgJ`4Mm3c)>ZBL^7wDh;Cq+tg$~
z%m-cqo*0$_QNW*88q&9|Zout5me?BNfGWC5lNI$sdHXGC)2K9NwqbSUH!R?e$fR#S
z+0RVjt-M{HVA1e_epc1#1?~T}ziNGYg#|4Vu>X7f4%rmo_cn?nhGYa*;P(sy8Iweg
zzkc_cw(hzu5+QA3Q9*lkR`f)(O=o{(UL_7V30Jd*+<6e@OVuy36@&!X;DW|h_|-%H
zk2Hl4`xycd@MQe*-E!U1+zJ2j68nC;Ww9)*Ep5rEQUcAkM(d%^Q$e6DZH>mjvtd|U
zDHT`d4$onx;K3gvXQgVN9>b@pe(k@AY=^TQo$-W<G8PuL3WFIte<WZ{>DvIc&?$?L
z6#-Uv&p0hiK0XW3v$UkI-b<YbK(fPOLb3|5;*l&6{(GkiA|vINGO(X1`+Y6r)NK1T
zBDF3}8crT<;PR;x>+h&}Q5&aB(0jA`;ZlYb4mG5Uv_fo_C@r~ne|KTurF>?HqBzeq
zJ~QNvhn%dLZrh5h?}NJHw$`E<9oc17n%26~-pl99{ijibEUf5Vd%T4oAF-*OoG)PB
zkJzCl#0yI!CFXkCT3OYTFB7F9pPE#wzrfTDo*w%mH<T(}x8N*s8^F5_kyIuaNEzND
zj4?R0JCZE0RgR8x$2XbbQ4YOyM=c-XSk_!-MtxD4@Uw~W<BRc+<=b}LR&@&<`#`%X
zmxG;r)3=`xR5nfX0!4>8EArMQ8z7y4#)K3AWN8mCh@zhU1Sg|Vw@7xtquq$i+B*mX
zJYVq0OH?u1{?jX<g|OoU$ZTaZy}iMl!E2wff1i2$M0eI*nxALAt8l#;6PHOLijw?E
zGy;wx%`8;e=LXFje76^8XJhY)gULMVZMp~6Webf`5#ljIiC*jI?njB!Hblkwv@v00
zJk1|wst7cskZx)+r-5iw5p>trHLpWwo$m+|aU_<VzrS-Zp&FuOeihTjym#hd=K_f4
zNbsfWYoDXymiuk$h=R#}9J}!!oO$TuTf8Je5V<*DF=GcRvB2r5VoGmlIuoB0e3=A}
zV@3rfeyk$kG!FIYW0tC!T@Kp^G6()wuDqE~(9mcT{jo6pcty5}qa=ZfFRGZpN}bI^
z8ev1GBRiqFKub-{locmmX0%r77kUq=gq<=VBa7PQ(}w=q{u6KCB~}+css8{%K{Tc-
z_u?}7wd<r5js)l<ptzc`8&y92R5@q$$3D~O`RB;5Su0>Jo;*-|rDoI3Pv$-0IkaF=
zGn^Nl_<2ikwuvz)V}u=B0`u@wkh?lO5fRadiL&Oxj+0ZQs(yW3{zU&idhA}jI|f3O
zq?n+j5<5O>1v`1%LARa?cqX4@-9pkS(rR!AD#n}Ld>RXOLJjRS;>$`E1aKmM@alI9
z>z$NRqGC5t2JsFdkig_T9BiSN3}+nL4?i#q!2cO>CxT&iGOpI!PqALCQaU_XK*sDG
zS>HMj9)nBpx_+o<F#4B9@;8cLLvh7noSe{X<e9NB%~F-7r7ZIitN=|dA=beS!P#NF
z{<%=jWyG<`;hv`6x6e*6&l}@nCki-CUgz(gY6BYP1o<n0MYlgR;YUP?--u_LX~Gv9
zy48kaoT>D9bu#)Uh2D!)fu}}v$vsGZ1_qxrI}GPgQkKDyz}Iy+v?%hEP|S=;e{A@c
zlm~9IV_0fqUGk|&Y!|+Hjm9#^XPdg>nJ-!=i##+uvDaU|gl0FC0`<uJiE)D`mdTLv
zmbzvg%tMP5p6PH%0ege~s`t^bwbn0`$S=$!`ZYx0x#J`86r96mXt`bQJ+XIq)!#st
z94JXzUaISoBR9t#zj@3-{m)&Jzl#3gmGkJ=jE84t*y*{20Hagz`|#q4Gs^g3BH9%|
zqZmL=Wib#WC&5lzz-QT%`AV9xk?+dw_IeZRsn7fLFR@vE^y5o9y9EOlrTu&nzstsM
zR!`-cONfu!Gx({)<<t%Sdi}tLac+^4PRbR3U`4#KMT7oToz_6%ltozk&y=_DwqLx4
zgiDbk`wST-q?tHgGL5cF|DWovGpea<3quG^U<jcY1O$`NCJscwNRu9#bZ6)wNJpfX
z0OA;MkQO0=ATYq7ATV?z1QZk`p*R#3A!;B55C%oW2pSZ1=EQk#%Fp-Xt@Yko`Egh7
zz2}^}&p!L?v)8x3tz4A$ZDr=U<*XUzntd(ReJyt3kkEQv(ew5j*ABbh4n7NA6tP!A
zg^Z{uON%-nm&}E)E1|CWG|Gq<7O)oBq~fCp;KN;$?1s#(cSe;rzsl^RWWV~*dY}Qf
zkx2>v{KIlay-QkO)e|~KFjY27;rVAaO5esb`Es`n6%|fjD#5&uQT21iV$1lF5M5p^
zB$CAYQ&S1OdIk0P(-)N&6E|d=zKm;N@lY0qbGL{%CW@8^<LMXrLsYw9U&GAEYW<hC
zQ&txgkLW`ssO{Gq%3=&1+_(cQPfI7ASc`1#z~+tg3yf|BITCN~$`x`lQR)h<hWWL+
z<S6Thxx#a{(K;5!uc`>1E98s>WB914geakU9?O$aIFD8Yk+o9u-QW~YNQUKX7M-a+
zFRfjgEgy2))<^)2C7ND5a3Tw&9H4;uk8&X~`SxMl?lz6Zr4icDQcj&?JxXc9H#pfk
zuR^V9huy8|FcwcTn<FSIGKc8fcaOY<C<M|B+U~}n`i@8_B%(2GRah{xgSo?LBl#}+
zD{usxB*R2BNb)Q0G^~kDHVUQ!<qu!1Dv9d!T152qO<3N%I+{%o*WyMEi0sxWb~cNy
zZ8aQ3v*9)et$dEJ`+9GEwyAiJAU?#d8R}J9Nh<Poh@18=6hZma__d|5J?UjY!}dMn
zX_!WoT~W_Hd-W4NJCWT!$Z)-LQ2b{iz^_H(7ItE`6R6DTwlMl9G>$pK5{RxfR^K~p
zU494=ZBa22LIKNSfS)c;&x~IL^Zr;|qITy}540x*wJAISFLmTQ<LWpD@@HO0d;J86
zORTOE_iD_s)_+`ishzKquG>jvl-X>A9bAH_3j_0u$lQISjIuti%6>`mNtM2THbWYG
z&+GyWPgVdcWnf_>&Pv|V31>2zdOq{;b+$w?B}Im#BLI?dR<M!ojc6H8*{raWET>np
zs)v*pOFW~*&no8UwzOn6g9vi<!ZHb*#_3gol?4p<@{xs@>~DhJAQi3>+GRahWcjR`
zGjnRW@!GM+_tZ%Qc1J<61AzrdKZR!H{Gpn<H_ctm&w47SwyZwdqpz5>GI~BuDLI7r
zu=TxC0-L-GJ$p-q5;w_hz~YPYEIv=-<%g$NcPEA#T5UmVol-AF@P_TG6~I=XH{V5p
z_g~Q6ueDfN{Q6C5&x5{Y`h<TocGU9XXnK3H@4bX{U?`v7vv>OIScF$Il`?T!^RI}8
zK?4wr$$qGEV6?WUr+Z8Ge@cH%FiK+7<sW%$_8=11ahtxwPVuT_6!MAw>t`Au0-y}n
zG{Z#x{m++MXq2f`b-wg5@}Ccai~o;jX)r*8PKb^j<L0Xof0lau^yx0hMAdgE-DcD+
zAc`pYyFV7n&fSOVea!tEhT$wU*XAtpc6Q6f9EUGHVemy+oyL8{g_x1fp*>w`<4v=G
zlb_-`zQfmKQpPtt5@f9Z{ZXlEoftW5$g)a%M~9iz@e9|kTzVn}l7VD6PZok4Wzqcy
zTpO(FvbDdBM;}xZA8cu9v9v8~5Q3)m`PEMWO2zdg!|qRv2#_NIc!mI_(}OUm87!*?
z<AB5A07s`*IIaTh>E>eOH+MePN<#-`X56Sui9u)d?Uy;qB&z@fU=1Dpqu6*PE>V0C
zU07ubg96?!##-54z-B*7RH2iGmmKReSqfPJpv!{c?JmZe?RjgyC^?hxug+VD%jKi`
z=BWSxwnst}nF}g6$v}yFoj-VI*{Uc5X}yo}i_(<g{IYBk|9pvqFppTlq?qAFK^@|i
zb(63QU#cGfpOAJ$*Y3_dSWLg1Y=EtbYbZ>${S+=vQ4ukq{$}~X(gls&OO*&f+@op;
zCd;T{00c)DnxPU#1t|p&24fi!LK#|h0As~m2mRv#EXECy6_=SCWCpv;^v&K5p){5{
z-pE!i7)C_7!tBTe;)VULA~E^<1=A*eaHVO(2rJwH0CjDa9Z22qBD16?JFf`hHm_5%
zp3Xp&3<JhK03$P#DhNN$Wbp`{TWPc$0%bMK(#SZ;K<QZ^jTAJ&Ct8b!L5a_EJv13>
z^MqCl8W)uO*4^X$S|c5cfSRdre{_*a!T0v?4@sQ;2v=1mZ~9~Jb2~JfMt`yOj{>5K
zzlwSAgA7M3vKc=s94KgWP1R$=7xNHVbrLTYxe^?zRNuAp={Bx~u?7%uf%e^rJ5I=@
z>(l1SIdAQZ-ogEccb`EPip#4I#YG9zUChBu6|8_6lYR48beGj?UIfE4@z$JM(;=R3
z`{*#-KmyFRI)8NN3SyKZ8mEy5;68=BW0hY{fCR{%lO5TQFm2t1yskG?gIhLcXi&;w
zD}x+#jDRQ}l%8j|($@O6XKj|^LhRL`y+?e~jEWauAbfAGInHgdZh87NQ0|)B0wOvp
z)1WE<?jgs2fgDVpr1<1+3qE`|A*{KPIp=A;RrcBCu`I_$CblJvNb2it{#B0i>CAlO
z{fB!-ntW%q)@3;!1-d&aCDz%h?uGrQA;EfzK;lMX7;ovfk`+;s+^?YSv^?dNgQo@4
zy!D%C2V5tTD#jynx1SKki&Jm!3(Cd*Ja^ji#D?xtgkh_!eAo2$Do(fST|_u?m&MQd
z=Z_0#5+kK1VpR81^8yf#k4IHX0LnKWc6spDq9^dDQ5J)V_3?QG#c@zl17u%%I+a~b
z>OBX2D9uLLPPT9BJJYn=Nh=6*7a0dXOkAjvdv%SS-zni9o(6TLzVb{{c*^8Hb1~XN
zQW&ZjsR9cV0NWZGvI6>9RA`8h^9Y8p_O^a+XTIF2kt|7yoojR~Q0zU5?snch1+0)y
zEhvI9Nm?xyFETZT>RG*6=$&=dSlm7G{e^wHi7g|QiGX;yRrZ_qj?&V%vq(S?d_o|9
z#?)_p(N2+4DRk%E2&7jZ9-30<5jQO|w&05o6%$K{^SoRgV?grBI4i}`dkPo0r<fC{
z^~@SRLgE{?iE1eIfZ;zv7A6m$8;f_b>Z~YeXEjlMPN~Nu7`bwfU-X2o<F+2H>lVI@
z8+r#;TGKQjZ1bWmoq;#KMRMzN36Zr@LK-%@eKZ|Yh^8p$G}craz4PVNG#ABqHWi7Z
z3tVNU^)<<HJQN=!LkNPG^2>`4Eia?7%KTH*4px1o(e-J+%D^DwM@M^efnL7;6=8mb
zLf=nu)1_wr>P1#*@$mmjs=(KKg<cY6T&PDFkp3w=SsSsOZJXdoIa)uz5r2%-enJHU
zi6^*cBQ~?#izZ|AilM8c$MC(kjtbwgjy>k!c53|e3d;XO5?pHXz(M5~s(yH>JEq(5
z;x&!j0YV_A1Q%R4Y&ag~FvK$kIux}(BQqc#0Hz)GUJ+?LnvP|OG)UG|f4$h_cB4Ne
zWae+tMne#7oF!9C{~2wx20=_E6F-N)Hem2dk2G+AmGs5_5%DA@@Pe3}rRl%g1kp?Y
yE&NZLEg*Z(@$vC-Y%R^c{V#R#KiDk$H-X<mk557Bm}21L`E2oymaXRg8NUM&Da$qh

literal 0
HcmV?d00001

diff --git a/docs/zh_cn/reward_model/modify_settings.md b/docs/zh_cn/reward_model/modify_settings.md
new file mode 100644
index 000000000..c56b04115
--- /dev/null
+++ b/docs/zh_cn/reward_model/modify_settings.md
@@ -0,0 +1,100 @@
+## 修改 Reward Model 训练配置
+
+本章节仅介绍与 Reward Model 训练相关的配置参数，更多 XTuner 配置文件的细节，请参考[修改训练配置](https://xtuner.readthedocs.io/zh-cn/latest/training/modify_settings.html)
+
+### 损失函数
+
+XTuner 使用了 [Bradley–Terry 模型](https://en.wikipedia.org/wiki/Bradley%E2%80%93Terry_model) 作为 Reward Model 的偏好建模方式，你可以指定 `loss_type="ranking"` 来使用 ranking loss。XTuner 中也实现了 InternLM2 中提出的 focal 损失函数，它通过调整难易样本的权重来避免过拟合，可以设置 `loss_type="focal"` 来使用该损失函数。对于该损失函数的详细说明，请参考 [InternLM2 技术报告](https://arxiv.org/abs/2403.17297)。
+
+另外，为了使 reward model 输出的 score 数值保持稳定，我们还在 loss 中额外增加了一个约束项，你可以指定 `penalty_type='log_barrier'` 或是 `penalty_type='L2'` 以启用对数约束或是L2约束。
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+loss_type = 'focal'  # 'ranking' or 'focal'
+penalty_type = 'log_barrier'  # 'log_barrier' or 'L2'
+```
+
+### 修改模型
+
+用户可以修改 `pretrained_model_name_or_path` 对预训练模型进行修改。
+
+需要注意的是，由于 XTuner 通过对数据的末尾添加 `<|reward|>` 特殊 token 的方式计算 reward 得分，因此当切换模型的词表发生变化时，该特殊 token 的 id 也需要进行相应的修改，我们通常会使用词表末尾未使用的 token 作为 reward token。
+
+例如，在 InternLM2 中我们使用 `[UNUSED_TOKEN_130]` 作为 reward token:
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'
+reward_token_id = 92527  # use [UNUSED_TOKEN_130] as reward token
+```
+
+如果用户将模型切换为llama3,我们则可以使用 `<|reserved_special_token_0|>` 作为 reward token:
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'meta-llama/Meta-Llama-3-8B-Instruct'
+reward_token_id = 128002  # use <|reserved_special_token_0|> as reward token
+```
+
+### 训练数据
+
+在 Reward Model 训练中，你可以通过 `max_length` 来指定单个样本序列的最大 token 数，XTuner 会自动对数据进行截断或是填充。
+
+```python
+# Data
+max_length = 2048
+```
+
+在配置文件中，我们通过 `train_dataset` 字段来指定训练数据集，你可以通过 `dataset` 字段指定数据集的加载方式，通过 `dataset_map_fn` 字段指定数据集的映射函数。
+
+```python
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='argilla/ultrafeedback-binarized-preferences-cleaned'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=False,
+    is_reward=True,
+    reward_token_id=reward_token_id,
+    num_proc=32,
+    use_varlen_attn=use_varlen_attn,
+    max_packed_length=max_packed_length,
+    shuffle_before_pack=True,
+)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(
+        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))
+```
+
+上述配置中，我们使用了 `load_dataset` 来加载 huggingface 上的 `argilla/ultrafeedback-binarized-preferences-cleaned` 数据集，使用 `orpo_dpo_mix_40k_map_fn` 作为数据集映射函数（这是因为 `orpo_dpo_mix_40k` 与 `ultrafeedback-binarized-preferences-cleaned` 的格式相同，因此这里共用了同一个映射函数）。
+
+关于如何处理数据集以及如何编写数据集映射函数，请参考[偏好数据集章节](./preference_data.md)。
+
+### 加速训练
+
+在使用偏好数据训练时，我们推荐您开启[变长注意力机制](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/varlen_flash_attn.html)， 以避免单个偏好内的 chosen 和 rejected 的样本长度差异造成的显存浪费。你可以通过 `use_varlen_attn=True` 来开启变长注意力机制。
+
+XTuner 中还支持了大量的训练加速方法，关于它们的使用方法，请参考[加速策略章节](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/hyper_parameters.html)。
diff --git a/docs/zh_cn/reward_model/overview.md b/docs/zh_cn/reward_model/overview.md
new file mode 100644
index 000000000..84b5ab14b
--- /dev/null
+++ b/docs/zh_cn/reward_model/overview.md
@@ -0,0 +1,29 @@
+## Reward Model 介绍
+
+### 简介
+
+Reward Model（奖励模型）是强化学习过程中一个关键的组成部分。它的主要任务是根据给定的输入和反馈来预测奖励值，从而指导学习算法的方向。在RLHF（Reinforcement Learning from Human Feedback）中，Reward Model 通过整合人类反馈，帮助强化学习算法更有效地优化策略。
+
+在大语言模型训练中，Reward Model 通常指的是偏好模型（Preference Model）。通过在训练时提供相同提示词的好与坏（chosen&rejected）的回复来拟合人类的偏好，并在推理时预测出一个奖励值，以指导 RLHF 过程中 Actor 模型的优化过程。
+
+Reward Model的应用场景包括但不限于：
+
+- **RLHF训练**：在使用 Proximal Policy Optimization（PPO）算法进行 RLHF 训练时，Reward Model提供奖励信号，指导模型优化策略，提高生成内容的质量并使其更贴近人类偏好。
+- **BoN采样**：在 Best-of-N（BoN）采样过程中，用户可以使用 Reward Model 对同一个提示词的多条回复进行打分，并选择奖励得分最高的生成结果，从而提升模型的输出效果。
+- **数据构造**：Reward Model 可以用于评估和过滤训练数据，或者也可以使用 Reward Model 替代人工标注来构造 DPO 训练数据。
+
+### XTuner 中 Reward Model 训练的优势
+
+XTuner 中的 Reward Model 训练具备以下显著优势：
+
+1. **使用最新的训练技巧**：XTuner 中集成了 InternLM2 中的 Reward Model 训练损失函数，可以稳定奖励得分的数值范围，也可以减少在简单样本上的过拟合（具体可参考 [InternLM2 技术报告](https://arxiv.org/abs/2403.17297)）。
+
+2. **减少显存浪费**：由于偏好数据中的 chosen 和 rejected 数据通常存在长度上的差异，因此在训练数据的拼接时会存在填充（padding token）,造成显存浪费。在 XTuner 中，基于 Flash Attention2 中的变长注意力功能，我们在训练过程中通过将偏好数据打包到同一个序列中，显著减少了由于 padding token 带来的显存浪费。这不仅提高了显存的利用效率，还使得在相同硬件条件下可以训练更大的模型或处理更多的数据。
+
+![img](./images/var_len_atten.png)
+
+3. **高效训练**：借助 XTuner 的 QLoRA 训练功能，我们能够仅对 Reward Model 的 Value Head 进行全参数训练，而对语言模型本身使用 QLoRA 微调，大幅降低了模型训练的显存开销。
+
+### 开始训练
+
+请参[阅快速上手](./quick_start.md)来了解最基本的概念，若希望了解更多训练参数配置相关的内容，请参考[修改Reward Model配置](./modify_settings.md)章节。
diff --git a/docs/zh_cn/reward_model/preference_data.md b/docs/zh_cn/reward_model/preference_data.md
new file mode 100644
index 000000000..1dd296053
--- /dev/null
+++ b/docs/zh_cn/reward_model/preference_data.md
@@ -0,0 +1,110 @@
+## 偏好数据集
+
+### 简介
+
+XTuner 的 Reward Model 与 DPO、ORPO 等依赖偏好数据的算法都采用了同样的数据格式，偏好数据集中的每一条训练样本需要包含以下三个字段：`prompt`、`chosen`、`rejected`。其中每个字段的值都使用了 [OpenAI chat message](https://platform.openai.com/docs/api-reference/chat/create) 格式。一个具体的例子如下所示：
+
+```json
+{
+  "prompt": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": "Who won the world series in 2020?"
+    },
+    {
+      "role": "assistant",
+      "content": "The Los Angeles Dodgers won the World Series in 2020."
+    },
+    {
+      "role": "user",
+      "content": "Where was it played?"
+    }
+  ],
+  "chosen": [
+    {
+      "role": "assistant",
+      "content": "The 2020 World Series was played at Globe Life Field in Arlington, Texas."
+    }
+  ],
+  "rejected": [
+    {
+      "role": "assistant",
+      "content": "I don't know."
+    }
+  ]
+}
+```
+
+当进行 Reward Model 训练或是 DPO 训练时，xtuner 会根据训练任务类型的不同，将偏好数据集处理为不同的训练标签。
+
+![img](./images/preference_data.png)
+
+如上图所示，当进行 Reward Model 训练时，我们参考 ChatGPT 的训练方式，在对话数据的最后添加一个特殊的`<|reward|>` token，只对该 token 输出的 logits 计算损失。而当进行 DPO 系列算法的训练时，我们则会屏蔽掉 prompt 部分的 token，只对 chosen 和 rejected 的回复部分计算损失。在配置文件中，我们通过 dataset 中的 `is_reward` 和 `is_dpo` 两个字段来控制数据集的类型。
+
+### 使用自定义数据进行训练
+
+在理解了单个样本的数据格式之后，我们只需要将转换好的数据保存为[JSON Lines](https://jsonlines.org/)格式即可。文件中的每一行都是一个样本。
+
+```json
+{"prompt": [{"content": "How can I learn to better regulate my emotions as a grown adult?", "role": "user"}], "chosen": [{"content": "You could do emotional regulation exercises with me, and try to notice and practice ways that you could better control your emotions in real life.", "role": "assistant"}], "rejected": [{"content": "Do you mean how can you learn to be more emotionally mature?", "role": "assistant"}]}
+{"prompt": [{"content": "What are some interesting things I can add to a grilled cheese sandwich?", "role": "user"}, {"content": "A tomato, a slice of pickle, some chopped pickles, some chopped green onions.  Also, some mayonnaise if you like.", "role": "assistant"}, {"content": "Sounds great. Anything else?", "role": "user"}], "chosen": [{"content": "If you’d like something a little different, I recommend grating some horseradish on top of the sandwich.", "role": "assistant"}], "rejected": [{"content": "Maybe some nice fresh basil?", "role": "assistant"}]}
+{"prompt": [{"content": "How do I bob for apples?", "role": "user"}], "chosen": [{"content": "Ah!  Do you mean the game of “applesauce,” or do you mean the idea of “to bob for apples”?", "role": "assistant"}], "rejected": [{"content": "Sorry, I don’t know that term.", "role": "assistant"}]}
+......
+```
+
+在准备完自定义数据集之后，你需要将你保存的数据路径填入配置文件中的 `data_files` 字段，你可以同时加载多个 jsonl 数据进行训练。
+
+```python
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_jsonl_dataset,
+        data_files=[
+            '/your/jsonl/path/here.jsonl',
+            '/your/another/jsonl/path/here.jsonl'
+        ]),
+)
+```
+
+### 使用开源数据集进行训练
+
+与 XTuner 配置 SFT 数据一样，在使用 huggingface 上的开源数据集时，我们只需要定义映射函数 map_fn，将开源数据集格式处理为 XTuner 中的数据格式即可。
+
+这里我们以 Intel/orca_dpo_pairs 为例，该数据集有 `system`、`question`、`chosen`、`rejected` 四个字段，并且每个字段的值为 text 而非 [OpenAI chat message](https://platform.openai.com/docs/api-reference/chat/create) 格式。因此我们需要为该数据集定义一个 map_fn：
+
+```python
+def intel_orca_dpo_map_fn(example):
+    prompt = [{
+        'role': 'system',
+        'content': example['system']
+    }, {
+        'role': 'user',
+        'content': example['question']
+    }]
+    chosen = [{'role': 'assistant', 'content': example['chosen']}]
+    rejected = [{'role': 'assistant', 'content': example['rejected']}]
+    return {'prompt': prompt, 'chosen': chosen, 'rejected': rejected}
+```
+
+通过代码可以看到，`intel_orca_dpo_map_fn` 对原数据中的四个字段进行处理，将其转换为了 `prompt`、`chosen`、`rejected` 三个字段，并且每个字段都处理为了[OpenAI chat message](https://platform.openai.com/docs/api-reference/chat/create) 格式，确保了后续数据处理流程的统一。
+
+完成了 map_fn 的定义之后，需要在配置文件中 import 该函数，并在 `dataset_map_fn` 字段中进行配置。
+
+```python
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='Intel/orca_dpo_pairs'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=intel_orca_dpo_map_fn,
+)
+```
diff --git a/docs/zh_cn/reward_model/quick_start.md b/docs/zh_cn/reward_model/quick_start.md
new file mode 100644
index 000000000..3762a4e8c
--- /dev/null
+++ b/docs/zh_cn/reward_model/quick_start.md
@@ -0,0 +1,86 @@
+## Reward Model 快速上手
+
+在本章节中，我们将介绍如何使用 XTuner 训练 1.8B 的 Reward Model，以帮助您快速上手。
+
+### 准备预训练模型权重
+
+依据 [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155) 论文中的描述，我们使用进过 SFT 的语言模型作为 Reward Model 的初始化模型。这里我们使用[InternLM2-chat-1.8b-sft](https://huggingface.co/internlm/internlm2-chat-1_8b-sft)作为初始化模型。
+
+在训练配置文件中设置`pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'`，则会在启动训练时自动下载模型文件。若您需要手动下载模型权重，那么请参考[准备预训练模型权重](https://xtuner.readthedocs.io/zh-cn/latest/preparation/pretrained_model.html)章节，其中详细说明了如何从 Huggingface 或者是 Modelscope 下载模型权重的方法。这里我们附上模型的 HuggingFace 链接与 ModelScope 链接：
+
+- HuggingFace 链接位于：https://huggingface.co/internlm/internlm2-chat-1_8b-sft
+
+- ModelScope 链接位于：https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-chat-1_8b-sft/summary
+
+### 准备训练数据
+
+在本教程中使用 [UltraFeedback](https://arxiv.org/abs/2310.01377) 数据集作为演示，为了方便起见，我们使用 huggingface 上已经预处理过的 [argilla/ultrafeedback-binarized-preferences-cleaned](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences-cleaned) 数据集，
+
+```python
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='argilla/ultrafeedback-binarized-preferences-cleaned'),
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=False,
+    is_reward=True,
+)
+```
+
+在配置文件中使用以上配置，即可自动下载并处理该数据集。如果您希望使用其他 huggingface 上的开源数据集或是使用自定义的数据集，请参阅[偏好数据集](./preference_data.md)章节。
+
+### 准备配置文件
+
+XTuner 提供了多个开箱即用的配置文件，可以通过 `xtuner list-cfg` 查看。我们执行如下指令，以复制一个配置文件到当前目录。
+
+```bash
+xtuner copy-cfg internlm2_chat_1_8b_reward_full_ultrafeedback .
+```
+
+打开复制后的配置文件，如果您选择自动下载模型和数据集，则无需修改配置。若您希望填入您预先下载的模型路径和数据集路径，请修改配置中的 `pretrained_model_name_or_path` 以及 `train_dataset` 中 `dataset` 的 `path` 参数。
+
+更多的训练参数配置，请参阅[修改Reward训练配置](./modify_settings.md)章节。
+
+### 启动训练
+
+在完成上述操作后，便可以使用下面的指令启动训练任务了。
+
+```bash
+# 单机单卡
+xtuner train ./internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py
+# 单机多卡
+NPROC_PER_NODE=${GPU_NUM} xtuner train ./internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py
+# slurm 集群
+srun ${SRUN_ARGS} xtuner train ./internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py --launcher slurm
+```
+
+正确的训练日志应当如下所示（在单卡 A800 上运行）：
+
+```
+06/06 16:12:11 - mmengine - INFO - Iter(train) [   10/15230]  lr: 3.9580e-07  eta: 2:59:41  time: 0.7084  data_time: 0.0044  memory: 18021  loss: 0.6270  acc: 0.0000  chosen_score_mean: 0.0000  rejected_score_mean: 0.0000  num_samples: 4.0000  num_tokens: 969.0000
+06/06 16:12:17 - mmengine - INFO - Iter(train) [   20/15230]  lr: 8.3536e-07  eta: 2:45:25  time: 0.5968  data_time: 0.0034  memory: 42180  loss: 0.6270  acc: 0.5000  chosen_score_mean: 0.0013  rejected_score_mean: 0.0010  num_samples: 4.0000  num_tokens: 1405.0000
+06/06 16:12:22 - mmengine - INFO - Iter(train) [   30/15230]  lr: 1.2749e-06  eta: 2:37:18  time: 0.5578  data_time: 0.0024  memory: 32121  loss: 0.6270  acc: 0.7500  chosen_score_mean: 0.0016  rejected_score_mean: 0.0011  num_samples: 4.0000  num_tokens: 932.0000
+06/06 16:12:28 - mmengine - INFO - Iter(train) [   40/15230]  lr: 1.7145e-06  eta: 2:36:05  time: 0.6033  data_time: 0.0025  memory: 42186  loss: 0.6270  acc: 0.7500  chosen_score_mean: 0.0027  rejected_score_mean: 0.0016  num_samples: 4.0000  num_tokens: 994.0000
+06/06 16:12:35 - mmengine - INFO - Iter(train) [   50/15230]  lr: 2.1540e-06  eta: 2:41:03  time: 0.7166  data_time: 0.0027  memory: 42186  loss: 0.6278  acc: 0.5000  chosen_score_mean: 0.0031  rejected_score_mean: 0.0032  num_samples: 4.0000  num_tokens: 2049.0000
+06/06 16:12:40 - mmengine - INFO - Iter(train) [   60/15230]  lr: 2.5936e-06  eta: 2:33:37  time: 0.4627  data_time: 0.0023  memory: 30238  loss: 0.6262  acc: 1.0000  chosen_score_mean: 0.0057  rejected_score_mean: 0.0030  num_samples: 4.0000  num_tokens: 992.0000
+06/06 16:12:46 - mmengine - INFO - Iter(train) [   70/15230]  lr: 3.0331e-06  eta: 2:33:18  time: 0.6018  data_time: 0.0025  memory: 42186  loss: 0.6247  acc: 0.7500  chosen_score_mean: 0.0117  rejected_score_mean: 0.0055  num_samples: 4.0000  num_tokens: 815.0000
+```
+
+### 模型转换
+
+XTuner 已经集成好了将模型转换为 HuggingFace 格式的工具，我们只需要执行
+
+```bash
+# 创建存放 hf 格式参数的目录
+mkdir work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy/iter_15230_hf
+
+# 转换格式
+xtuner convert pth_to_hf internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py.py \
+                            work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py/iter_15230.pth \
+                            work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py/iter_15230_hf
+```
+
+便能够将 XTuner 的 ckpt 转换为 Huggingface 格式的模型。
+
+需要注意的是，由于 Reward Model 的类型并未在 transformers 官方库中集成，因此目前只有InternLM2模型训练得到的 Reward Model 会被转换为 InternLM2ForRewardModel 类型，而其他模型则会默认转换为 SequenceClassification 类型（例如 LLaMa3 会被转换为 LlamaForSequenceClassification 类型），但这并不影响其在 XTuner PPO 训练中的使用。

From 44749c2adf214add084f01514ec2ce463c506e8a Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Wed, 3 Jul 2024 20:05:13 +0800
Subject: [PATCH 06/29] Support internlm2.5 (#803)

* add internlm 2.5 configs

* update readme
---
 README.md                                     |  31 +--
 README_zh-CN.md                               |  29 +--
 ...chat_7b_full_finetune_custom_dataset_e1.py | 226 ++++++++++++++++++
 .../internlm2_5_chat_7b_qlora_alpaca_e3.py    | 219 +++++++++++++++++
 .../internlm2_5_chat_7b_qlora_oasst1_e3.py    | 219 +++++++++++++++++
 5 files changed, 688 insertions(+), 36 deletions(-)
 create mode 100644 xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_full_finetune_custom_dataset_e1.py
 create mode 100644 xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_alpaca_e3.py
 create mode 100644 xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_oasst1_e3.py

diff --git a/README.md b/README.md
index e0be695ef..4e729226c 100644
--- a/README.md
+++ b/README.md
@@ -39,6 +39,8 @@ English | [简体中文](README_zh-CN.md)
 
 ## 🎉 News
 
+- **\[2024/07\]** Support [InternLM 2.5](xtuner/configs/internlm/internlm2_5_chat_7b/) models!
+- **\[2024/06\]** Support [DeepSeek V2](xtuner/configs/deepseek/deepseek_v2_chat/) models! **2x faster!**
 - **\[2024/04\]** [LLaVA-Phi-3-mini](https://huggingface.co/xtuner/llava-phi-3-mini-hf) is released! Click [here](xtuner/configs/llava/phi3_mini_4k_instruct_clip_vit_large_p14_336) for details!
 - **\[2024/04\]** [LLaVA-Llama-3-8B](https://huggingface.co/xtuner/llava-llama-3-8b) and [LLaVA-Llama-3-8B-v1.1](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1) are released! Click [here](xtuner/configs/llava/llama3_8b_instruct_clip_vit_large_p14_336) for details!
 - **\[2024/04\]** Support [Llama 3](xtuner/configs/llama) models!
@@ -100,16 +102,15 @@ XTuner is an efficient, flexible and full-featured toolkit for fine-tuning large
 <tr valign="top">
 <td align="left" valign="top">
 <ul>
-  <li><a href="https://huggingface.co/internlm">InternLM2</a></li>
-  <li><a href="https://huggingface.co/meta-llama">Llama 3</a></li>
-  <li><a href="https://huggingface.co/meta-llama">Llama 2</a></li>
+  <li><a href="https://huggingface.co/internlm">InternLM2 / 2.5</a></li>
+  <li><a href="https://huggingface.co/meta-llama">Llama 2 / 3</a></li>
   <li><a href="https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3">Phi-3</a></li>
   <li><a href="https://huggingface.co/THUDM/chatglm2-6b">ChatGLM2</a></li>
   <li><a href="https://huggingface.co/THUDM/chatglm3-6b">ChatGLM3</a></li>
   <li><a href="https://huggingface.co/Qwen/Qwen-7B">Qwen</a></li>
   <li><a href="https://huggingface.co/baichuan-inc/Baichuan2-7B-Base">Baichuan2</a></li>
-  <li><a href="https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1">Mixtral 8x7B</a></li>
-  <li><a href="https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat">DeepSeek MoE</a></li>
+  <li><a href="https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1">Mixtral</a></li>
+  <li><a href="https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat">DeepSeek V2</a></li>
   <li><a href="https://huggingface.co/google">Gemma</a></li>
   <li>...</li>
 </ul>
@@ -203,14 +204,14 @@ XTuner supports the efficient fine-tune (*e.g.*, QLoRA) for LLMs. Dataset prepar
   xtuner train ${CONFIG_NAME_OR_PATH}
   ```
 
-  For example, we can start the QLoRA fine-tuning of InternLM2-Chat-7B with oasst1 dataset by
+  For example, we can start the QLoRA fine-tuning of InternLM2.5-Chat-7B with oasst1 dataset by
 
   ```shell
   # On a single GPU
-  xtuner train internlm2_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
+  xtuner train internlm2_5_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
   # On multiple GPUs
-  (DIST) NPROC_PER_NODE=${GPU_NUM} xtuner train internlm2_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
-  (SLURM) srun ${SRUN_ARGS} xtuner train internlm2_chat_7b_qlora_oasst1_e3 --launcher slurm --deepspeed deepspeed_zero2
+  (DIST) NPROC_PER_NODE=${GPU_NUM} xtuner train internlm2_5_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
+  (SLURM) srun ${SRUN_ARGS} xtuner train internlm2_5_chat_7b_qlora_oasst1_e3 --launcher slurm --deepspeed deepspeed_zero2
   ```
 
   - `--deepspeed` means using [DeepSpeed](https://github.com/microsoft/DeepSpeed) 🚀 to optimize the training. XTuner comes with several integrated strategies including ZeRO-1, ZeRO-2, and ZeRO-3. If you wish to disable this feature, simply remove this argument.
@@ -231,18 +232,10 @@ XTuner provides tools to chat with pretrained / fine-tuned LLMs.
 xtuner chat ${NAME_OR_PATH_TO_LLM} --adapter {NAME_OR_PATH_TO_ADAPTER} [optional arguments]
 ```
 
-For example, we can start the chat with
-
-InternLM2-Chat-7B with adapter trained from oasst1 dataset:
-
-```shell
-xtuner chat internlm/internlm2-chat-7b --adapter xtuner/internlm2-chat-7b-qlora-oasst1 --prompt-template internlm2_chat
-```
-
-LLaVA-InternLM2-7B:
+For example, we can start the chat with InternLM2.5-Chat-7B :
 
 ```shell
-xtuner chat internlm/internlm2-chat-7b --visual-encoder openai/clip-vit-large-patch14-336 --llava xtuner/llava-internlm2-7b --prompt-template internlm2_chat --image $IMAGE_PATH
+xtuner chat internlm/internlm2_5-chat-7b --prompt-template internlm2_chat
 ```
 
 For more examples, please see [chat.md](./docs/en/user_guides/chat.md).
diff --git a/README_zh-CN.md b/README_zh-CN.md
index c5037d28c..16c1a2af2 100644
--- a/README_zh-CN.md
+++ b/README_zh-CN.md
@@ -39,6 +39,8 @@
 
 ## 🎉 更新
 
+- **\[2024/07\]** 支持 [InternLM 2.5](xtuner/configs/internlm/internlm2_5_chat_7b/) 模型!
+- **\[2024/06\]** 支持 [DeepSeek V2](xtuner/configs/deepseek/deepseek_v2_chat/) models! **训练速度提升一倍！**
 - **\[2024/04\]** 多模态大模型 [LLaVA-Phi-3-mini](https://huggingface.co/xtuner/llava-phi-3-mini-hf) 发布！快速开始请查阅此[文档](xtuner/configs/llava/phi3_mini_4k_instruct_clip_vit_large_p14_336)！
 - **\[2024/04\]** 多模态大模型 [LLaVA-Llama-3-8B](https://huggingface.co/xtuner/llava-llama-3-8b) 和 [LLaVA-Llama-3-8B-v1.1](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1) 发布！快速开始请查阅此[文档](xtuner/configs/llava/llama3_8b_instruct_clip_vit_large_p14_336)！
 - **\[2024/04\]** 支持 [Llama 3](xtuner/configs/llama) 模型！
@@ -100,16 +102,15 @@ XTuner 是一个高效、灵活、全能的轻量化大模型微调工具库。
 <tr valign="top">
 <td align="left" valign="top">
 <ul>
-  <li><a href="https://huggingface.co/internlm">InternLM2</a></li>
-  <li><a href="https://huggingface.co/meta-llama">Llama 3</a></li>
-  <li><a href="https://huggingface.co/meta-llama">Llama 2</a></li>
+  <li><a href="https://huggingface.co/internlm">InternLM 2 / 2.5</a></li>
+  <li><a href="https://huggingface.co/meta-llama">Llama 2 / 3</a></li>
   <li><a href="https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3">Phi-3</a></li>
   <li><a href="https://huggingface.co/THUDM/chatglm2-6b">ChatGLM2</a></li>
   <li><a href="https://huggingface.co/THUDM/chatglm3-6b">ChatGLM3</a></li>
   <li><a href="https://huggingface.co/Qwen/Qwen-7B">Qwen</a></li>
   <li><a href="https://huggingface.co/baichuan-inc/Baichuan2-7B-Base">Baichuan2</a></li>
-  <li><a href="https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1">Mixtral 8x7B</a></li>
-  <li><a href="https://huggingface.co/deepseek-ai/deepseek-moe-16b-chat">DeepSeek MoE</a></li>
+  <li><a href="https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1">Mixtral</a></li>
+  <li><a href="https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat">DeepSeek V2</a></li>
   <li><a href="https://huggingface.co/google">Gemma</a></li>
   <li>...</li>
 </ul>
@@ -203,14 +204,14 @@ XTuner 支持微调大语言模型。数据集预处理指南请查阅[文档](.
   xtuner train ${CONFIG_NAME_OR_PATH}
   ```
 
-  例如，我们可以利用 QLoRA 算法在 oasst1 数据集上微调 InternLM2-Chat-7B：
+  例如，我们可以利用 QLoRA 算法在 oasst1 数据集上微调 InternLM2.5-Chat-7B：
 
   ```shell
   # 单卡
-  xtuner train internlm2_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
+  xtuner train internlm2_5_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
   # 多卡
-  (DIST) NPROC_PER_NODE=${GPU_NUM} xtuner train internlm2_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
-  (SLURM) srun ${SRUN_ARGS} xtuner train internlm2_chat_7b_qlora_oasst1_e3 --launcher slurm --deepspeed deepspeed_zero2
+  (DIST) NPROC_PER_NODE=${GPU_NUM} xtuner train internlm2_5_chat_7b_qlora_oasst1_e3 --deepspeed deepspeed_zero2
+  (SLURM) srun ${SRUN_ARGS} xtuner train internlm2_5_chat_7b_qlora_oasst1_e3 --launcher slurm --deepspeed deepspeed_zero2
   ```
 
   - `--deepspeed` 表示使用 [DeepSpeed](https://github.com/microsoft/DeepSpeed) 🚀 来优化训练过程。XTuner 内置了多种策略，包括 ZeRO-1、ZeRO-2、ZeRO-3 等。如果用户期望关闭此功能，请直接移除此参数。
@@ -233,16 +234,10 @@ xtuner chat ${NAME_OR_PATH_TO_LLM} --adapter {NAME_OR_PATH_TO_ADAPTER} [optional
 
 例如：
 
-与 InternLM2-Chat-7B, oasst1 adapter 对话：
+与 InternLM2.5-Chat-7B 对话：
 
 ```shell
-xtuner chat internlm/internlm2-chat-7b --adapter xtuner/internlm2-chat-7b-qlora-oasst1 --prompt-template internlm2_chat
-```
-
-与 LLaVA-InternLM2-7B 对话：
-
-```shell
-xtuner chat internlm/internlm2-chat-7b --visual-encoder openai/clip-vit-large-patch14-336 --llava xtuner/llava-internlm2-7b --prompt-template internlm2_chat --image $IMAGE_PATH
+xtuner chat internlm/internlm2-chat-7b --prompt-template internlm2_chat
 ```
 
 更多示例，请查阅[文档](./docs/zh_cn/user_guides/chat.md)。
diff --git a/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_full_finetune_custom_dataset_e1.py b/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_full_finetune_custom_dataset_e1.py
new file mode 100644
index 000000000..bc8a2816a
--- /dev/null
+++ b/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_full_finetune_custom_dataset_e1.py
@@ -0,0 +1,226 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""Data format:
+[
+    {
+        "conversation": [
+            {
+                "system": "",
+                "input": "xxx",
+                "output": "xxx"
+            },
+            {
+                "input": "xxx",
+                "output": "xxx"
+            }
+        ]
+    },
+...
+]
+Please refer to https://github.com/InternLM/xtuner/blob/main/docs/en/user_guides/dataset_format.md for details.
+"""  # noqa: E501
+from datasets import load_dataset
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR
+from torch.optim import AdamW
+from torch.utils.data import BatchSampler
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import template_map_fn_factory
+from xtuner.dataset.samplers import InternRepoSampler
+from xtuner.engine import (DatasetInfoHook, EvaluateChatHook, ThroughputHook,
+                           VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2_5-7b-chat'
+use_varlen_attn = True
+
+# Data
+data_files = ['/path/to/json/file.json']
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 32768
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+# batch size per device, set to 1 if `use_varlen_attn` = True
+# To clarify, enlarging the batch size essentially enlarges the `max_length`.
+# For example, doubling the max length is tantamount to doubling the batch size
+batch_size = 1
+accumulative_counts = 1  # 1bs * 1acc * 64gpu = 64 batchsize
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+lr = 4e-5
+betas = (0.9, 0.95)
+weight_decay = 0.01
+max_norm = 1  # grad clip
+warm_up_ratio = 0.025
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    use_varlen_attn=use_varlen_attn,
+    dataset=dict(type=load_dataset, path='json', data_files=data_files),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=None,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=InternRepoSampler, shuffle=True, seed=1024),
+    batch_sampler=dict(
+        type=BatchSampler, drop_last=True, batch_size=batch_size),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type='LinearLR',
+        start_factor=1 / 40,
+        by_epoch=True,
+        begin=0,
+        end=warm_up_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=lr * 0.15,
+        by_epoch=True,
+        begin=warm_up_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(
+        type=DatasetInfoHook, tokenizer=tokenizer,
+        is_intern_repo_dataset=True),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template),
+    dict(type=ThroughputHook)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 100 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=1),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+log_processor = dict(
+    by_epoch=False,
+    window_size=1,
+    mean_pattern=r'.*(loss|time|data_time|grad_norm|tflops).*')
diff --git a/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_alpaca_e3.py b/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_alpaca_e3.py
new file mode 100644
index 000000000..7dfc92617
--- /dev/null
+++ b/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_alpaca_e3.py
@@ -0,0 +1,219 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2_5-7b-chat'
+use_varlen_attn = False
+
+# Data
+alpaca_en_path = 'tatsu-lab/alpaca'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_en,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_oasst1_e3.py b/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_oasst1_e3.py
new file mode 100644
index 000000000..98b097efb
--- /dev/null
+++ b/xtuner/configs/internlm/internlm2_5_chat_7b/internlm2_5_chat_7b_qlora_oasst1_e3.py
@@ -0,0 +1,219 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2_5-7b-chat'
+use_varlen_attn = False
+
+# Data
+data_path = 'timdettmers/openassistant-guanaco'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=data_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=oasst1_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)

From 9c28b40790f139e29625ad273e0878d3649f542e Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Tue, 9 Jul 2024 08:41:43 +0800
Subject: [PATCH 07/29] [Bugs] fix dispatch bugs when model not in
 LOWEST_TRANSFORMERS_VERSION (#802)

* fix dispatch bugs when model not in LOWEST_TRANSFORMERS_VERSION

* move rope_theta
---
 xtuner/model/modules/dispatch/__init__.py | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/xtuner/model/modules/dispatch/__init__.py b/xtuner/model/modules/dispatch/__init__.py
index 7cb159515..e81ec7a3a 100644
--- a/xtuner/model/modules/dispatch/__init__.py
+++ b/xtuner/model/modules/dispatch/__init__.py
@@ -228,14 +228,14 @@ def replace_rote(model):
     from mmengine import print_log
     print_log = log_once(print_log)
 
-    assert hasattr(model.config, 'rope_theta'), \
-        '`rope_theta` should be in the model config.'
-    rope_theta = model.config.rope_theta
-
     def traverse(module):
         for name, child in module.named_children():
             cls_name = type(child).__name__
             if cls_name in ROTE_DISPATCH_MAPPING:
+                assert hasattr(model.config, 'rope_theta'), \
+                    '`rope_theta` should be in the model config.'
+                rope_theta = model.config.rope_theta
+
                 rote = ROTE_DISPATCH_MAPPING[cls_name]
                 rote = rote.build()
                 print_log(f'replace {cls_name}', 'current')
@@ -258,10 +258,11 @@ def check(model_name):
             # a walkaround for reward model
             model_name = model_name[:-5] + 'ForCausalLM'
         msg = '{} requires transformers version at least {}, but got {}'
-        assert TRANSFORMERS_VERSION >= LOWEST_TRANSFORMERS_VERSION[
-            model_name], msg.format(model_name,
-                                    LOWEST_TRANSFORMERS_VERSION[model_name],
-                                    TRANSFORMERS_VERSION)
+        if model_name in LOWEST_TRANSFORMERS_VERSION:
+            assert TRANSFORMERS_VERSION >= LOWEST_TRANSFORMERS_VERSION[
+                model_name], msg.format(
+                    model_name, LOWEST_TRANSFORMERS_VERSION[model_name],
+                    TRANSFORMERS_VERSION)
 
     check(type(model).__name__)
     if use_varlen_attn:

From 7575f2c9137cf6ea24f7cb63ae22c1bad38bbe79 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Tue, 9 Jul 2024 08:42:09 +0800
Subject: [PATCH 08/29] [Docs] fix benchmark table (#801)

fix benchmark table
---
 docs/zh_cn/acceleration/benchmark.rst | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/zh_cn/acceleration/benchmark.rst b/docs/zh_cn/acceleration/benchmark.rst
index 6434b6d15..5a1c80804 100644
--- a/docs/zh_cn/acceleration/benchmark.rst
+++ b/docs/zh_cn/acceleration/benchmark.rst
@@ -79,7 +79,7 @@
     - 176.9
     - `llama2_7b_full_alpaca_enzh_256k_sp8.py <https://github.com/InternLM/xtuner/tree/main/xtuner/configs/llama_speed_benchmark/llama2_7b/llama2_7b_full_alpaca_enzh_256k_sp8.py>`_
   * - Llama2-7B
-    - 8
+    - 32
     - 1M
     - 133.6
     - 153.9

From 8adc8d4a5d6f5a33474f85093d8bb43c24c2b32e Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Tue, 9 Jul 2024 08:42:51 +0800
Subject: [PATCH 09/29] [Feature] support output without loss in openai_map_fn
 (#816)

support output without loss in openai_map_fn
---
 xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py b/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py
index 64ed642f6..c1798dc45 100644
--- a/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py
+++ b/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py
@@ -32,10 +32,13 @@ def openai_map_fn(example):
         elif msg['role'] == 'user':
             input += msg['content']
         elif msg['role'] == 'assistant':
+            output_with_loss = msg.get('loss', True)
+            output_with_loss = output_with_loss.lower() == 'true'
             conversation.append({
                 'system': system,
                 'input': input,
-                'output': msg['content']
+                'output': msg['content'],
+                'output_with_loss': output_with_loss
             })
             system = ''
             input = ''

From 48df4c8bd137abb215e09cd4ec03dee564f51d18 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Wed, 10 Jul 2024 14:13:01 +0800
Subject: [PATCH 10/29] [Docs] fix typos in sp docs (#821)

fix typo
---
 docs/zh_cn/acceleration/train_extreme_long_sequence.rst | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/zh_cn/acceleration/train_extreme_long_sequence.rst b/docs/zh_cn/acceleration/train_extreme_long_sequence.rst
index b59a5a83f..65b364ad8 100644
--- a/docs/zh_cn/acceleration/train_extreme_long_sequence.rst
+++ b/docs/zh_cn/acceleration/train_extreme_long_sequence.rst
@@ -56,7 +56,7 @@
     - yi-34B
     - ZeRO-3
     - 16
-    - OOM
+    - 227
 
 
 为解决长序列训练过程中的显存问题，Megatron-LM 团队和 DeepSpeed 团队分别提出了两种序列并行算法，通过对长序列进行切分的方法来降低单 GPU 上计算的序列长度。XTuner 中的序列并行设计思路参考了 DeepSpeed 的工作 `DeepSpeed Ulysses <https://arxiv.org/abs/2309.14509>`_，并加以优化， **以实现一键开启序列并行策略** 。三者的对比如下：

From b92481fb40c6df2941bd18dafb512e91514abedc Mon Sep 17 00:00:00 2001
From: Xu Song <xusong.vip@gmail.com>
Date: Thu, 11 Jul 2024 17:21:07 +0800
Subject: [PATCH 11/29] [Feature] Support the DatasetInfoHook of DPO training
 (#787)

* [Feature] Support the DatasetInfoHook of DPO training

* fix yapf check
---
 .../internlm/internlm2_chat_1_8b_dpo_full.py  |  4 +--
 ...internlm2_chat_1_8b_dpo_full_varlenattn.py |  4 +--
 ..._1_8b_dpo_full_varlenattn_jsonl_dataset.py |  4 +--
 .../internlm2_chat_7b_dpo_qlora_varlenattn.py |  4 +--
 ...llama3_8b_instruct_dpo_qlora_varlenattn.py |  4 +--
 xtuner/engine/hooks/dataset_info_hook.py      | 29 ++++++++++++-------
 6 files changed, 28 insertions(+), 21 deletions(-)

diff --git a/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full.py b/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full.py
index dd3909f72..908683fe6 100644
--- a/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full.py
+++ b/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full.py
@@ -11,7 +11,7 @@
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
                                                orpo_dpo_mix_40k_map_fn)
-from xtuner.engine.hooks import (EvaluateChatHook,
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
 from xtuner.model.dpo import DPO
@@ -141,7 +141,7 @@
 #######################################################################
 # Log the dialogue periodically during the training process, optional
 custom_hooks = [
-    # dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
     dict(
         type=EvaluateChatHook,
         tokenizer=tokenizer,
diff --git a/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn.py b/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn.py
index 3e5cdc35a..787ad68bb 100644
--- a/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn.py
+++ b/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn.py
@@ -11,7 +11,7 @@
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
                                                orpo_dpo_mix_40k_map_fn)
-from xtuner.engine.hooks import (EvaluateChatHook,
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
 from xtuner.model.dpo import DPO
@@ -151,7 +151,7 @@
 #######################################################################
 # Log the dialogue periodically during the training process, optional
 custom_hooks = [
-    # dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
     dict(
         type=EvaluateChatHook,
         tokenizer=tokenizer,
diff --git a/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn_jsonl_dataset.py b/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn_jsonl_dataset.py
index 55bb270a4..ae1a3cdca 100644
--- a/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn_jsonl_dataset.py
+++ b/xtuner/configs/dpo/internlm/internlm2_chat_1_8b_dpo_full_varlenattn_jsonl_dataset.py
@@ -10,7 +10,7 @@
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
                                                load_jsonl_dataset)
-from xtuner.engine.hooks import (EvaluateChatHook,
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
 from xtuner.model.dpo import DPO
@@ -155,7 +155,7 @@
 #######################################################################
 # Log the dialogue periodically during the training process, optional
 custom_hooks = [
-    # dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
     dict(
         type=EvaluateChatHook,
         tokenizer=tokenizer,
diff --git a/xtuner/configs/dpo/internlm/internlm2_chat_7b_dpo_qlora_varlenattn.py b/xtuner/configs/dpo/internlm/internlm2_chat_7b_dpo_qlora_varlenattn.py
index b051ea2a1..659d029b3 100644
--- a/xtuner/configs/dpo/internlm/internlm2_chat_7b_dpo_qlora_varlenattn.py
+++ b/xtuner/configs/dpo/internlm/internlm2_chat_7b_dpo_qlora_varlenattn.py
@@ -14,7 +14,7 @@
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
                                                orpo_dpo_mix_40k_map_fn)
-from xtuner.engine.hooks import (EvaluateChatHook,
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
 from xtuner.model.dpo import DPO
@@ -170,7 +170,7 @@
 #######################################################################
 # Log the dialogue periodically during the training process, optional
 custom_hooks = [
-    # dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
     dict(
         type=EvaluateChatHook,
         tokenizer=tokenizer,
diff --git a/xtuner/configs/dpo/llama/llama3_8b_instruct_dpo_qlora_varlenattn.py b/xtuner/configs/dpo/llama/llama3_8b_instruct_dpo_qlora_varlenattn.py
index 0ca90f51c..e94b88fd0 100644
--- a/xtuner/configs/dpo/llama/llama3_8b_instruct_dpo_qlora_varlenattn.py
+++ b/xtuner/configs/dpo/llama/llama3_8b_instruct_dpo_qlora_varlenattn.py
@@ -14,7 +14,7 @@
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
                                                orpo_dpo_mix_40k_map_fn)
-from xtuner.engine.hooks import (EvaluateChatHook,
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                  VarlenAttnArgsToMessageHubHook)
 from xtuner.engine.runner import TrainLoop
 from xtuner.model.dpo import DPO
@@ -170,7 +170,7 @@
 #######################################################################
 # Log the dialogue periodically during the training process, optional
 custom_hooks = [
-    # dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
     dict(
         type=EvaluateChatHook,
         tokenizer=tokenizer,
diff --git a/xtuner/engine/hooks/dataset_info_hook.py b/xtuner/engine/hooks/dataset_info_hook.py
index d997373ec..84dc9498a 100644
--- a/xtuner/engine/hooks/dataset_info_hook.py
+++ b/xtuner/engine/hooks/dataset_info_hook.py
@@ -25,19 +25,26 @@ def __init__(self, tokenizer, is_intern_repo_dataset=False):
         self.is_intern_repo_dataset = is_intern_repo_dataset
 
     def log(self, runner, dataset, mode='train'):
+
+        def _log(input_ids, log_prefix=''):
+            if self.is_intern_repo_dataset:
+                input_ids = [abs(x) for x in input_ids]
+            # Try to split list to be compatible with IMAGE token
+            input_ids = split_list(input_ids, IMAGE_TOKEN_INDEX)
+            text = log_prefix
+            for idx, ids in enumerate(input_ids):
+                text += self.tokenizer.decode(ids)
+                if idx != len(input_ids) - 1:
+                    text += DEFAULT_IMAGE_TOKEN
+            runner.logger.info(text)
+
         runner.logger.info(f'Num {mode} samples {len(dataset)}')
         runner.logger.info(f'{mode} example:')
-        input_ids = dataset[0]['input_ids']
-        if self.is_intern_repo_dataset:
-            input_ids = [abs(x) for x in input_ids]
-        # Try to split list to be compatible with IMAGE token
-        input_ids = split_list(input_ids, IMAGE_TOKEN_INDEX)
-        text = ''
-        for idx, ids in enumerate(input_ids):
-            text += self.tokenizer.decode(ids)
-            if idx != len(input_ids) - 1:
-                text += DEFAULT_IMAGE_TOKEN
-        runner.logger.info(text)
+        if 'chosen_ids' in dataset[0]:
+            _log(dataset[0]['chosen_ids'], log_prefix='chosen: ')
+            _log(dataset[0]['rejected_ids'], log_prefix='rejected: ')
+        else:
+            _log(dataset[0]['input_ids'])
 
     def before_train(self, runner) -> None:
         do_train = runner.train_loop is not None

From ff226e18f2d686dffe5cde842b78d7f9fe996693 Mon Sep 17 00:00:00 2001
From: RangiLyu <lyuchqi@gmail.com>
Date: Fri, 19 Jul 2024 10:03:40 +0800
Subject: [PATCH 12/29] [Enhance]: Fix sequence parallel memory bottleneck in
 DPO & ORPO (#830)

* [WIP]: Fix sequence parallel memory bottleneck in DPO

* loss mask before split

* refactor orpo
---
 xtuner/model/dpo.py  | 115 ++++++++++++++++++------------------
 xtuner/model/orpo.py | 135 ++++++++++++++++++++++---------------------
 2 files changed, 128 insertions(+), 122 deletions(-)

diff --git a/xtuner/model/dpo.py b/xtuner/model/dpo.py
index b46ea1c50..9a7b97a19 100644
--- a/xtuner/model/dpo.py
+++ b/xtuner/model/dpo.py
@@ -62,77 +62,66 @@ def _gather_masked_logits(self, logits, labels, mask):
 
     def get_logps(
             self,
-            all_logits,  # bs, seqlen,vocab_size
-            all_ref_logits,  # bs, seqlen,vocab_size
-            labels,  # bs, seqlen
+            policy_logps,  # bs, seqlen,vocab_size
+            ref_logps,  # bs, seqlen,vocab_size
+            loss_mask,  # bs, seqlen
     ):
-        labels = labels[:, 1:].clone()
-        all_logits = all_logits[:, :-1, :]
-        all_ref_logits = all_ref_logits[:, :-1, :]
-
-        labels[labels == -100] = 0
-        loss_mask = labels != 0
-        all_logps = self._gather_masked_logits(all_logits, labels,
-                                               loss_mask).sum(-1)
-        all_ref_logps = self._gather_masked_logits(all_ref_logits, labels,
-                                                   loss_mask).sum(-1)
+        policy_logps = policy_logps[:, :-1].sum(-1)
+        ref_logps = ref_logps[:, :-1].sum(-1)
+        loss_mask = loss_mask[:, :-1]
 
         if self.loss_type == 'ipo':  # average_log_prob
-            all_logps = all_logps / loss_mask.sum(-1)
-            all_ref_logps = all_ref_logps / loss_mask.sum(-1)
+            policy_logps = policy_logps / loss_mask.sum(-1)
+            ref_logps = ref_logps / loss_mask.sum(-1)
 
-        policy_chosen_logps = all_logps[::2]
-        policy_rejected_logps = all_logps[1::2]
-        reference_chosen_logps = all_ref_logps[::2]
-        reference_rejected_logps = all_ref_logps[1::2]
+        policy_chosen_logps = policy_logps[::2]
+        policy_rejected_logps = policy_logps[1::2]
+        reference_chosen_logps = ref_logps[::2]
+        reference_rejected_logps = ref_logps[1::2]
         return (policy_chosen_logps, policy_rejected_logps,
                 reference_chosen_logps, reference_rejected_logps)
 
-    def get_var_len_atten_logps(self, all_logits, all_ref_logits, labels,
+    def get_var_len_atten_logps(self, policy_logps, ref_logps, loss_mask,
                                 cu_seqlens, attention_mask):
         seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
         # unpack sequence
-        unpacked_logits = torch.split(all_logits, seqlens, dim=1)
-        unpacked_ref_logits = torch.split(all_ref_logits, seqlens, dim=1)
-        unpacked_labels = torch.split(labels, seqlens, dim=1)
+        unpacked_policy_logps = torch.split(policy_logps, seqlens, dim=1)
+        unpacked_ref_logps = torch.split(ref_logps, seqlens, dim=1)
+        unpacked_loss_mask = torch.split(loss_mask, seqlens, dim=1)
         if attention_mask is not None:
             # It indicate that we pad the original sequence, labels,
             # position_ids and cumulative_len for sequence parallel if the
             # attention_mask is not None.
             # We then need to remove the padded segments.
             assert False in attention_mask
-            unpacked_logits = unpacked_logits[:-1]
-            unpacked_ref_logits = unpacked_ref_logits[:-1]
-            unpacked_labels = unpacked_labels[:-1]
-            assert len(unpacked_logits) % 2 == 0
+            unpacked_policy_logps = unpacked_policy_logps[:-1]
+            unpacked_ref_logps = unpacked_ref_logps[:-1]
+            unpacked_loss_mask = unpacked_loss_mask[:-1]
+            assert len(unpacked_policy_logps) % 2 == 0
 
-        def compute_logps(_logits, _labels):
-            _labels = _labels[:, 1:].clone()
-            _logits = _logits[:, :-1, :]
-            _labels[_labels == -100] = 0
-            loss_mask = _labels != 0
-            logps = self._gather_masked_logits(_logits, _labels, loss_mask)
-            logps = logps.sum(-1)
+        def compute_logps(_logps, _mask):
+            _logps = _logps[:, :-1].sum(-1)
+            _mask = _mask[:, :-1]
             if self.loss_type == 'ipo':
-                logps /= loss_mask.sum(-1)
-            return logps
+                _logps /= _mask.sum(-1)
+            return _logps
 
         (policy_chosen_logps, policy_rejected_logps, reference_chosen_logps,
          reference_rejected_logps) = [], [], [], []
-        for i in range(len(unpacked_logits) // 2):
-            chosen = unpacked_logits[2 * i]
-            rejected = unpacked_logits[2 * i + 1]
-            chosen_ref = unpacked_ref_logits[2 * i]
-            rejected_ref = unpacked_ref_logits[2 * i + 1]
-            chosen_label = unpacked_labels[2 * i]
-            rejected_label = unpacked_labels[2 * i + 1]
-            policy_chosen_logps.append(compute_logps(chosen, chosen_label))
+        for i in range(len(unpacked_policy_logps) // 2):
+            chosen = unpacked_policy_logps[2 * i]
+            rejected = unpacked_policy_logps[2 * i + 1]
+            chosen_ref = unpacked_ref_logps[2 * i]
+            rejected_ref = unpacked_ref_logps[2 * i + 1]
+            chosen_mask = unpacked_loss_mask[2 * i]
+            rejected_mask = unpacked_loss_mask[2 * i + 1]
+            policy_chosen_logps.append(compute_logps(chosen, chosen_mask))
             policy_rejected_logps.append(
-                compute_logps(rejected, rejected_label))
+                compute_logps(rejected, rejected_mask))
             reference_chosen_logps.append(
-                compute_logps(chosen_ref, chosen_label))
+                compute_logps(chosen_ref, chosen_mask))
             reference_rejected_logps.append(
-                compute_logps(rejected_ref, rejected_label))
+                compute_logps(rejected_ref, rejected_mask))
 
         return (torch.stack(policy_chosen_logps),
                 torch.stack(policy_rejected_logps),
@@ -142,7 +131,7 @@ def compute_logps(_logits, _labels):
     @staticmethod
     def _split_for_sequence_parallel(data):
         # attention mask should not be split
-        ARGS_NEED_TO_SPLIT = ('input_ids', 'position_ids')
+        ARGS_NEED_TO_SPLIT = ('input_ids', 'position_ids', 'labels')
         sp_group = get_sequence_parallel_group()
         for key in ARGS_NEED_TO_SPLIT:
             val = data.get(key, None)
@@ -154,8 +143,14 @@ def _split_for_sequence_parallel(data):
 
     def compute_loss(self, data, data_samples=None):
         # modified from https://github.com/huggingface/trl/blob/main/trl/trainer/dpo_trainer.py  # noqa
-
-        labels = data.pop('labels')
+        # shift labels first and add a dummy label at the end, to support sequence parallel  # noqa
+        data['labels'] = torch.cat(
+            (data['labels'][:, 1:], torch.zeros_like(data['labels'][:, :1])),
+            dim=1)
+        tmp_label = data['labels'].clone()
+        tmp_label[tmp_label == 0] = -100
+        all_loss_mask = data[
+            'labels'] != -100  # loss mask of all tokens in all sp ranks  # noqa
 
         if get_sequence_parallel_world_size() > 1:
             data = self._split_for_sequence_parallel(data)
@@ -168,14 +163,22 @@ def compute_loss(self, data, data_samples=None):
             else:
                 all_ref_logits = self.ref_llm(**data).logits
 
+        labels = data['labels']
+        labels[labels == -100] = 0
+        loss_mask = labels != 0  # loss mask in a single sp rank
+        policy_logps = self._gather_masked_logits(all_logits, labels,
+                                                  loss_mask)
+        ref_logps = self._gather_masked_logits(all_ref_logits, labels,
+                                               loss_mask)
+
         if get_sequence_parallel_world_size() > 1:
-            all_logits = gather_forward_split_backward(
-                all_logits,
+            policy_logps = gather_forward_split_backward(
+                policy_logps,
                 dim=1,
                 sp_group=get_sequence_parallel_group(),
                 grad_scale='up')
-            all_ref_logits = gather_forward_split_backward(
-                all_ref_logits,
+            ref_logps = gather_forward_split_backward(
+                ref_logps,
                 dim=1,
                 sp_group=get_sequence_parallel_group(),
                 grad_scale='up')
@@ -184,7 +187,7 @@ def compute_loss(self, data, data_samples=None):
             (policy_chosen_logps, policy_rejected_logps,
              reference_chosen_logps,
              reference_rejected_logps) = self.get_logps(
-                 all_logits, all_ref_logits, labels)
+                 policy_logps, ref_logps, all_loss_mask)
         else:
             message_hub = MessageHub.get_instance('varlen_attn_args')
             rank = dist.get_rank()
@@ -192,7 +195,7 @@ def compute_loss(self, data, data_samples=None):
             (policy_chosen_logps, policy_rejected_logps,
              reference_chosen_logps,
              reference_rejected_logps) = self.get_var_len_atten_logps(
-                 all_logits, all_ref_logits, labels, cu_seqlens,
+                 policy_logps, ref_logps, all_loss_mask, cu_seqlens,
                  data['attention_mask'])
 
         pi_logratios = policy_chosen_logps - policy_rejected_logps
diff --git a/xtuner/model/orpo.py b/xtuner/model/orpo.py
index 5fb4b7d27..37264088a 100644
--- a/xtuner/model/orpo.py
+++ b/xtuner/model/orpo.py
@@ -34,17 +34,12 @@ def _gather_masked_logits(self, logits, labels, mask):
 
     def get_logps(
             self,
-            all_logits,  # bs, seqlen,vocab_size
-            average_log_prob,  # bs, seqlen,vocab_size
-            labels,  # bs, seqlen
+            all_logps,  # bs, seqlen
+            average_log_prob,
+            loss_mask,  # bs, seqlen
     ):
-        labels = labels[:, 1:].clone()
-        all_logits = all_logits[:, :-1, :]
-
-        labels[labels == -100] = 0
-        loss_mask = labels != 0
-        all_logps = self._gather_masked_logits(all_logits, labels,
-                                               loss_mask).sum(-1)
+        all_logps = all_logps[:, :-1].sum(-1)
+        loss_mask = loss_mask[:, :-1]
 
         if average_log_prob:  # average_log_prob
             all_logps = all_logps / loss_mask.sum(-1)
@@ -53,47 +48,44 @@ def get_logps(
         rejected_logps = all_logps[1::2]
         return chosen_logps, rejected_logps
 
-    def get_var_len_atten_logps(self, all_logits, average_log_prob, labels,
+    def get_var_len_atten_logps(self, all_logps, average_log_prob, loss_mask,
                                 cu_seqlens, attention_mask):
         seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
         # unpack sequence
-        unpacked_logits = torch.split(all_logits, seqlens, dim=1)
-        unpacked_labels = torch.split(labels, seqlens, dim=1)
+        unpacked_logps = torch.split(all_logps, seqlens, dim=1)
+        unpacked_loss_mask = torch.split(loss_mask, seqlens, dim=1)
         if attention_mask is not None:
             # It indicate that we pad the original sequence, labels,
             # position_ids and cumulative_len for sequence parallel if the
             # attention_mask is not None.
             # We then need to remove the padded segments.
             assert False in attention_mask
-            unpacked_logits = unpacked_logits[:-1]
-            unpacked_labels = unpacked_labels[:-1]
-            assert len(unpacked_logits) % 2 == 0
-
-        def compute_logps(_logits, _labels):
-            _labels = _labels[:, 1:].clone()
-            _logits = _logits[:, :-1, :]
-            _labels[_labels == -100] = 0
-            loss_mask = _labels != 0
-            logps = self._gather_masked_logits(_logits, _labels, loss_mask)
-            logps = logps.sum(-1)
+            unpacked_logps = unpacked_logps[:-1]
+            unpacked_loss_mask = unpacked_loss_mask[:-1]
+            assert len(unpacked_logps) % 2 == 0
+
+        def compute_logps(_logps, _mask):
+            _logps = _logps[:, :-1].sum(-1)
+            _mask = _mask[:, :-1]
             if average_log_prob:
-                logps /= loss_mask.sum(-1)
-            return logps
+                _logps /= _mask.sum(-1)
+            return _logps
 
         chosen_logps, rejected_logps = [], []
-        for i in range(len(unpacked_logits) // 2):
-            chosen = unpacked_logits[2 * i]
-            rejected = unpacked_logits[2 * i + 1]
-            chosen_label = unpacked_labels[2 * i]
-            rejected_label = unpacked_labels[2 * i + 1]
-            chosen_logps.append(compute_logps(chosen, chosen_label))
-            rejected_logps.append(compute_logps(rejected, rejected_label))
+        for i in range(len(unpacked_logps) // 2):
+            chosen = unpacked_logps[2 * i]
+            rejected = unpacked_logps[2 * i + 1]
+            chosen_mask = unpacked_loss_mask[2 * i]
+            rejected_mask = unpacked_loss_mask[2 * i + 1]
+            chosen_logps.append(compute_logps(chosen, chosen_mask))
+            rejected_logps.append(compute_logps(rejected, rejected_mask))
 
         return (torch.stack(chosen_logps), torch.stack(rejected_logps))
 
     def cross_entropy_loss(self, logits, labels):
         logits = logits[..., :-1, :].contiguous()
-        labels = labels[..., 1:].contiguous()
+        # labels are already shifted, now we need to remove the last dummy label  # noqa
+        labels = labels[..., :-1].contiguous()
         # Flatten the tokens
         loss_fct = nn.CrossEntropyLoss()
         logits = logits.view(-1, logits.shape[-1])
@@ -126,7 +118,8 @@ def odds_ratio_loss(
     @staticmethod
     def _split_for_sequence_parallel(data):
         # attention mask should not be split
-        ARGS_NEED_TO_SPLIT = ('input_ids', 'position_ids')
+        ARGS_NEED_TO_SPLIT = ('input_ids', 'position_ids', 'labels',
+                              'chosen_rejected_tag')
         sp_group = get_sequence_parallel_group()
         for key in ARGS_NEED_TO_SPLIT:
             val = data.get(key, None)
@@ -137,53 +130,63 @@ def _split_for_sequence_parallel(data):
         return data
 
     def compute_loss(self, data, data_samples=None):
-        labels_ori = data.pop('labels')
+        # shift labels first and add a dummy label at the end, to support sequence parallel  # noqa
+        data['labels'] = torch.cat(
+            (data['labels'][:, 1:], torch.zeros_like(data['labels'][:, :1])),
+            dim=1)
+        tmp_label = data['labels'].clone()
+        tmp_label[tmp_label == 0] = -100
+        # loss mask of all tokens in all sp ranks
+        all_loss_mask = data['labels'] != -100
+
+        if self.use_varlen_attn:
+            # create a chosen rejected tag for varlen_attn ce loss
+            message_hub = MessageHub.get_instance('varlen_attn_args')
+            rank = dist.get_rank()
+            cu_seqlens = message_hub.get_info(f'cumulative_len_rank_{rank}')
+            seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
+
+            chosen_rejected_tag = torch.ones_like(data['labels'])
+            unpacked_tag = list(
+                torch.split(chosen_rejected_tag, seqlens, dim=1))
+            # import pdb; pdb.set_trace()
+            for i in range(len(unpacked_tag) // 2):
+                # import pdb; pdb.set_trace()
+                unpacked_tag[2 * i + 1] *= 0
+            chosen_rejected_tag = torch.cat(unpacked_tag, dim=1)
+            data['chosen_rejected_tag'] = chosen_rejected_tag
 
         if get_sequence_parallel_world_size() > 1:
             data = self._split_for_sequence_parallel(data)
-
+        chosen_rejected_tag = data.pop('chosen_rejected_tag', None)
         all_logits = self.llm(**data).logits
+
+        labels = data['labels'].clone()
+        labels[labels == -100] = 0
+        loss_mask = labels != 0  # loss mask in a single sp rank
+        all_logps = self._gather_masked_logits(all_logits, labels, loss_mask)
         if get_sequence_parallel_world_size() > 1:
-            all_logits = gather_forward_split_backward(
-                all_logits,
+            all_logps = gather_forward_split_backward(
+                all_logps,
                 dim=1,
                 sp_group=get_sequence_parallel_group(),
                 grad_scale='up')
 
         if not self.use_varlen_attn:
             chosen_nll_loss = self.cross_entropy_loss(all_logits[::2],
-                                                      labels_ori.clone()[::2])
+                                                      data['labels'][::2])
             chosen_logps, rejected_logps = self.get_logps(
-                all_logits, True, labels_ori)
+                all_logps, True, all_loss_mask)
         else:
-            message_hub = MessageHub.get_instance('varlen_attn_args')
-            rank = dist.get_rank()
-            cu_seqlens = message_hub.get_info(f'cumulative_len_rank_{rank}')
-            seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
-
-            attention_mask = data['attention_mask']
-            if attention_mask is not None:
-                # It indicate that we pad the original sequence, labels,
-                # position_ids and cumulative_len for sequence parallel if the
-                # attention_mask is not None.
-                # We then need to remove the padded segments.
-                logits = torch.split(all_logits, seqlens, dim=1)[:-1]
-                assert len(logits) % 2 == 0
-                chosen_logits = logits[::2]
-                labels = torch.split(labels_ori.clone(), seqlens, dim=1)[:-1]
-                assert len(labels) % 2 == 0
-                chosen_labels = labels[::2]
-            else:
-                chosen_logits = torch.split(all_logits, seqlens, dim=1)[::2]
-                chosen_labels = torch.split(
-                    labels_ori.clone(), seqlens, dim=1)[::2]
-
-            chosen_logits = torch.cat(chosen_logits, dim=1)
-            chosen_labels = torch.cat(chosen_labels, dim=1)
+            chosen_idxs = chosen_rejected_tag == 1
+            chosen_logits = all_logits[chosen_idxs]
+            chosen_labels = data['labels'][chosen_idxs]
             chosen_nll_loss = self.cross_entropy_loss(chosen_logits,
                                                       chosen_labels)
+
             chosen_logps, rejected_logps = self.get_var_len_atten_logps(
-                all_logits, True, labels_ori, cu_seqlens, attention_mask)
+                all_logps, True, all_loss_mask, cu_seqlens,
+                data['attention_mask'])
         (losses, chosen_rewards, rejected_rewards, log_odds_ratio,
          log_odds_chosen) = self.odds_ratio_loss(chosen_logps, rejected_logps)
         losses = losses.mean()

From 381d1c866d106505a164011a421df5a742aa33d7 Mon Sep 17 00:00:00 2001
From: bychen7 <55865490+bychen7@users.noreply.github.com>
Date: Fri, 19 Jul 2024 10:08:15 +0800
Subject: [PATCH 13/29] [Fix] Fix typo (#795)

Update quick_start.md

Fix typo
---
 docs/zh_cn/reward_model/quick_start.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/zh_cn/reward_model/quick_start.md b/docs/zh_cn/reward_model/quick_start.md
index 3762a4e8c..736624cef 100644
--- a/docs/zh_cn/reward_model/quick_start.md
+++ b/docs/zh_cn/reward_model/quick_start.md
@@ -76,7 +76,7 @@ XTuner 已经集成好了将模型转换为 HuggingFace 格式的工具，我们
 mkdir work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy/iter_15230_hf
 
 # 转换格式
-xtuner convert pth_to_hf internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py.py \
+xtuner convert pth_to_hf internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py \
                             work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py/iter_15230.pth \
                             work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py/iter_15230_hf
 ```

From ba7afc79b1f338b8485d6f51763c078f877116da Mon Sep 17 00:00:00 2001
From: Xu Song <xusong.vip@gmail.com>
Date: Fri, 19 Jul 2024 13:32:02 +0800
Subject: [PATCH 14/29] [Fix] fix initialization of ref_llm for full param dpo
 training with zero-3 (#778)

* Fix initialization of ref_llm

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update sft.py

* Update dpo.py

* Update dpo.py

* Update dpo.py
---
 xtuner/model/dpo.py | 45 ++++++++++++++++++++++++---------------------
 xtuner/model/sft.py | 21 +++++++++++++++------
 2 files changed, 39 insertions(+), 27 deletions(-)

diff --git a/xtuner/model/dpo.py b/xtuner/model/dpo.py
index 9a7b97a19..9384ddb34 100644
--- a/xtuner/model/dpo.py
+++ b/xtuner/model/dpo.py
@@ -16,6 +16,26 @@
 from .sft import SupervisedFinetune
 
 
+def disable_grad(model):
+    # freeze parameters
+    parameter_names = [n for n, _ in model.named_parameters()]
+    for param_name in parameter_names:
+        param = model.get_parameter(param_name)
+        param.requires_grad = False
+    return model.eval()
+
+
+def create_reference_model(model):
+    if is_deepspeed_zero3_enabled():
+        raise ValueError('DeepSpeed ZeRO-3 is enabled and is not compatible '
+                         'with `create_reference_model()`. Please instantiate '
+                         'your reference model directly with '
+                         '`AutoCausalLM.from_pretrained()`.')
+    ref_model = deepcopy(model)
+    ref_model = disable_grad(ref_model)
+    return ref_model
+
+
 class DPO(SupervisedFinetune):
     """A general class of DPO and its variants."""
 
@@ -27,32 +47,15 @@ def __init__(self,
                  label_smoothing=0.0,
                  **kwargs):
         super().__init__(llm, **kwargs)
-        self.ref_llm = ref_llm
         self.loss_type = loss_type
         self.label_smoothing = label_smoothing
         self.beta = beta
 
-        if not self.use_lora:
-            self.ref_llm = self.create_reference_model(ref_llm, **kwargs)
-
-    def create_reference_model(self, ref_llm=None, **kwargs):
-        ref_model = None
-        if ref_llm is None:
-            if is_deepspeed_zero3_enabled():
-                raise ValueError(
-                    'DeepSpeed ZeRO-3 is enabled and is not compatible '
-                    'with `deepcopy(self.llm)`. Please instantiate '
-                    'your reference model by modifying key `model.ref_llm` '
-                    'in your config with `AutoCausalLM.from_pretrained()`.')
-            ref_model = deepcopy(self.llm)
+        if ref_llm is not None:
+            ref_llm = self._build_llm_from_cfg(ref_llm, kwargs.get("use_varlen_attn"), kwargs.get("max_position_embeddings"))
+            self.ref_llm = disable_grad(ref_llm)
         else:
-            ref_model = SupervisedFinetune(ref_llm, **kwargs).llm
-        # freeze parameters
-        parameter_names = [n for n, _ in ref_model.named_parameters()]
-        for param_name in parameter_names:
-            param = ref_model.get_parameter(param_name)
-            param.requires_grad = False
-        return ref_model.eval()
+            self.ref_llm = None if self.use_lora else create_reference_model(self.llm)
 
     def _gather_masked_logits(self, logits, labels, mask):
         logits = torch.gather(
diff --git a/xtuner/model/sft.py b/xtuner/model/sft.py
index d030c6c20..9c3fa38c9 100644
--- a/xtuner/model/sft.py
+++ b/xtuner/model/sft.py
@@ -79,10 +79,8 @@ def __init__(self,
                  tokenizer=None,
                  max_position_embeddings=None):
         super().__init__()
-        with LoadWoInit():
-            if isinstance(llm, dict):
-                llm = self._dispatch_lm_model_cfg(llm, max_position_embeddings)
-            self.llm = self._build_from_cfg_or_module(llm)
+
+        self.llm = self._build_llm_from_cfg(llm, use_varlen_attn, max_position_embeddings)
 
         if tokenizer is not None:
             if isinstance(tokenizer, dict):
@@ -90,8 +88,6 @@ def __init__(self,
             smart_tokenizer_and_embedding_resize(tokenizer, self.llm)
 
         self.llm.config.use_cache = False
-        dispatch_modules(self.llm, use_varlen_attn=use_varlen_attn)
-
         if use_activation_checkpointing:
             # For backward compatibility
             if hasattr(self.llm, 'enable_input_require_grads'):
@@ -119,6 +115,19 @@ def __init__(self,
         # the sequence.
         self.use_varlen_attn = use_varlen_attn
 
+
+    def _build_llm_from_cfg(self, llm_cfg, use_varlen_attn, max_position_embeddings):
+        # For forward
+        with LoadWoInit():
+            if isinstance(llm_cfg, dict):
+                llm = self._dispatch_lm_model_cfg(llm_cfg, max_position_embeddings)
+            llm = self._build_from_cfg_or_module(llm)
+
+        llm.config.use_cache = False
+        dispatch_modules(llm, use_varlen_attn=use_varlen_attn)
+        return llm
+
+
     def gradient_checkpointing_enable(self):
         self.activation_checkpointing_enable()
 

From 30133d57fc205488c2a7e1ea110ee46e93cb6801 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 19 Jul 2024 17:09:17 +0800
Subject: [PATCH 15/29] [Bugs] Fix attn mask (#852)

* [WIP]: Fix sequence parallel memory bottleneck in DPO

* loss mask before split

* refactor orpo

* fix attention_mask in preference_collate_fn

---------

Co-authored-by: RangiLyu <lyuchqi@gmail.com>
---
 xtuner/dataset/collate_fns/preference_collate_fn.py | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/xtuner/dataset/collate_fns/preference_collate_fn.py b/xtuner/dataset/collate_fns/preference_collate_fn.py
index 8a6060410..ca21613bb 100644
--- a/xtuner/dataset/collate_fns/preference_collate_fn.py
+++ b/xtuner/dataset/collate_fns/preference_collate_fn.py
@@ -58,7 +58,7 @@ def preference_collate_fn(instances: Sequence[Dict],
         labels = torch.stack(labels)
 
     if use_varlen_attn:
-        attention_mask = None
+        attention_mask = torch.ones_like(input_ids).bool()
         position_ids = torch.stack(position_ids, dim=0)
     else:
         # Some tokenizers have the same eos token and pad token, so input_ids
@@ -74,8 +74,10 @@ def preference_collate_fn(instances: Sequence[Dict],
         input_ids = pad_for_sequence_parallel(input_ids, pad_index)
         labels = pad_for_sequence_parallel(labels, IGNORE_INDEX)
         position_ids = pad_for_sequence_parallel(position_ids, 0)
-        if attention_mask is not None:
-            attention_mask = pad_for_sequence_parallel(attention_mask, 0)
+        # We use attention_mask to distinguish `input_ids` from
+        # (sequence parallel) pad tokens in `get_var_len_atten_logps` method of
+        # class `DPO` and `ORPO`
+        attention_mask = pad_for_sequence_parallel(attention_mask, 0)
         if use_varlen_attn:
             (cumulative_len, attention_mask
              ) = pad_cumulative_len_for_sequence_parallel(cumulative_len)

From d58c1dd48fd5405fcfee0a409a854e215ec5f022 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 19 Jul 2024 17:09:46 +0800
Subject: [PATCH 16/29] fix lint (#854)

---
 xtuner/model/dpo.py |  7 +++++--
 xtuner/model/sft.py | 11 ++++++-----
 2 files changed, 11 insertions(+), 7 deletions(-)

diff --git a/xtuner/model/dpo.py b/xtuner/model/dpo.py
index 9384ddb34..faaa43402 100644
--- a/xtuner/model/dpo.py
+++ b/xtuner/model/dpo.py
@@ -52,10 +52,13 @@ def __init__(self,
         self.beta = beta
 
         if ref_llm is not None:
-            ref_llm = self._build_llm_from_cfg(ref_llm, kwargs.get("use_varlen_attn"), kwargs.get("max_position_embeddings"))
+            ref_llm = self.build_llm_from_cfg(
+                ref_llm, kwargs.get('use_varlen_attn', False),
+                kwargs.get('max_position_embeddings', None))
             self.ref_llm = disable_grad(ref_llm)
         else:
-            self.ref_llm = None if self.use_lora else create_reference_model(self.llm)
+            self.ref_llm = None if self.use_lora else create_reference_model(
+                self.llm)
 
     def _gather_masked_logits(self, logits, labels, mask):
         logits = torch.gather(
diff --git a/xtuner/model/sft.py b/xtuner/model/sft.py
index 9c3fa38c9..522950489 100644
--- a/xtuner/model/sft.py
+++ b/xtuner/model/sft.py
@@ -80,7 +80,8 @@ def __init__(self,
                  max_position_embeddings=None):
         super().__init__()
 
-        self.llm = self._build_llm_from_cfg(llm, use_varlen_attn, max_position_embeddings)
+        self.llm = self.build_llm_from_cfg(llm, use_varlen_attn,
+                                           max_position_embeddings)
 
         if tokenizer is not None:
             if isinstance(tokenizer, dict):
@@ -115,19 +116,19 @@ def __init__(self,
         # the sequence.
         self.use_varlen_attn = use_varlen_attn
 
-
-    def _build_llm_from_cfg(self, llm_cfg, use_varlen_attn, max_position_embeddings):
+    def build_llm_from_cfg(self, llm_cfg, use_varlen_attn,
+                           max_position_embeddings):
         # For forward
         with LoadWoInit():
             if isinstance(llm_cfg, dict):
-                llm = self._dispatch_lm_model_cfg(llm_cfg, max_position_embeddings)
+                llm = self._dispatch_lm_model_cfg(llm_cfg,
+                                                  max_position_embeddings)
             llm = self._build_from_cfg_or_module(llm)
 
         llm.config.use_cache = False
         dispatch_modules(llm, use_varlen_attn=use_varlen_attn)
         return llm
 
-
     def gradient_checkpointing_enable(self):
         self.activation_checkpointing_enable()
 

From 16e2f8f1abc272f2b31ec900f6c94731e20a49d6 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 19 Jul 2024 17:10:37 +0800
Subject: [PATCH 17/29] [Bugs] Fix dispatch attn bug (#829)

* fix collate bug

* fix dispatch attn bugs

* fix rotary_seq_len bug
---
 xtuner/dataset/collate_fns/default_collate_fn.py | 4 ++--
 xtuner/model/modules/dispatch/cohere.py          | 7 ++++++-
 xtuner/model/modules/dispatch/deepseek_v2.py     | 5 +++++
 xtuner/model/modules/dispatch/internlm2.py       | 5 +++++
 xtuner/model/modules/dispatch/mistral.py         | 8 +++++++-
 xtuner/model/modules/dispatch/phi3.py            | 8 +++++++-
 xtuner/model/modules/dispatch/qwen2.py           | 8 +++++++-
 7 files changed, 39 insertions(+), 6 deletions(-)

diff --git a/xtuner/dataset/collate_fns/default_collate_fn.py b/xtuner/dataset/collate_fns/default_collate_fn.py
index f4d5f9197..0ca9264f0 100644
--- a/xtuner/dataset/collate_fns/default_collate_fn.py
+++ b/xtuner/dataset/collate_fns/default_collate_fn.py
@@ -56,8 +56,8 @@ def default_collate_fn(instances: Sequence[Dict],
         # Some tokenizers have the same eos token and pad token, so input_ids
         # cannot be masked directly based on the pad token id.
         attention_mask = torch.zeros_like(input_ids).bool()
-        for i in ori_length:
-            attention_mask[:i] = True
+        for i, length in enumerate(ori_length):
+            attention_mask[i, :length] = True
 
         bs, seq_len = input_ids.shape
         position_ids = torch.arange(seq_len).unsqueeze(0).long().repeat(bs, 1)
diff --git a/xtuner/model/modules/dispatch/cohere.py b/xtuner/model/modules/dispatch/cohere.py
index edeb771e3..d3529f570 100644
--- a/xtuner/model/modules/dispatch/cohere.py
+++ b/xtuner/model/modules/dispatch/cohere.py
@@ -105,17 +105,22 @@ def cohere_attn_forward(
         query_states, key_states, value_states = \
             pre_process_for_sequence_parallel_attn(
                 query_states, key_states, value_states)
+        # self.num_heads is used in self._upad_input method
+        # num_heads has been changed because of sequence parallel
+        ori_num_head = self.num_heads
+        self.num_heads = query_states.shape[-2]
 
     attn_output = self._flash_attention_forward(
         query_states,
         key_states,
         value_states,
         attention_mask,
-        q_len,
+        query_states.shape[1],
         dropout=dropout_rate)
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
+        self.num_heads = ori_num_head
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
     attn_output = self.o_proj(attn_output)
diff --git a/xtuner/model/modules/dispatch/deepseek_v2.py b/xtuner/model/modules/dispatch/deepseek_v2.py
index dcdb677a3..667d2227c 100644
--- a/xtuner/model/modules/dispatch/deepseek_v2.py
+++ b/xtuner/model/modules/dispatch/deepseek_v2.py
@@ -128,6 +128,10 @@ def deepseek_attn_forward(
         query_states, key_states, value_states = \
             pre_process_for_sequence_parallel_attn(
                 query_states, key_states, value_states)
+        # self.num_heads is used in self._upad_input method
+        # num_heads has been changed because of sequence parallel
+        ori_num_head = self.num_heads
+        self.num_heads = query_states.shape[-2]
 
     attn_output = self._flash_attention_forward(
         query_states,
@@ -141,6 +145,7 @@ def deepseek_attn_forward(
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
+        self.num_heads = ori_num_head
 
     if self.q_head_dim != self.v_head_dim:
         attn_output = attn_output[:, :, :, :self.v_head_dim]
diff --git a/xtuner/model/modules/dispatch/internlm2.py b/xtuner/model/modules/dispatch/internlm2.py
index 5b855d4ab..7c601f0dc 100644
--- a/xtuner/model/modules/dispatch/internlm2.py
+++ b/xtuner/model/modules/dispatch/internlm2.py
@@ -149,6 +149,10 @@ def internlm2_attn_forward(
         query_states, key_states, value_states = \
             pre_process_for_sequence_parallel_attn(
                 query_states, key_states, value_states)
+        # self.num_heads is used in self._upad_input method
+        # num_heads has been changed because of sequence parallel
+        ori_num_head = self.num_heads
+        self.num_heads = query_states.shape[-2]
 
     dropout_rate = 0.0
     attn_output = self._flash_attention_forward(
@@ -161,6 +165,7 @@ def internlm2_attn_forward(
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
+        self.num_heads = ori_num_head
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
     attn_output = self.wo(attn_output)
diff --git a/xtuner/model/modules/dispatch/mistral.py b/xtuner/model/modules/dispatch/mistral.py
index 49dfdc108..d08b0f00e 100644
--- a/xtuner/model/modules/dispatch/mistral.py
+++ b/xtuner/model/modules/dispatch/mistral.py
@@ -214,6 +214,11 @@ def mistral_attn_forward(
         query_states, key_states, value_states = \
             pre_process_for_sequence_parallel_attn(
                 query_states, key_states, value_states)
+        # num_heads has been changed because of sequence parallel
+        # `self.num_heads`` is not used in self._flash_attention_forward
+        # in mistral/mixtral, we are doing this to avoid some unnecessary risk
+        ori_num_head = self.num_heads
+        self.num_heads = query_states.shape[-2]
 
     attn_output = self._flash_attention_forward(
         query_states,
@@ -227,6 +232,7 @@ def mistral_attn_forward(
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
+        self.num_heads = ori_num_head
 
     attn_output = attn_output.reshape(bsz, q_len,
                                       self.hidden_size).contiguous()
@@ -311,7 +317,7 @@ def mistral_varlen_attn_forward(
         value_states = value_states.transpose(1, 2)
         # Because the input can be padded, the absolute sequence length
         # depends on the max position id.
-        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item() + 1)
+        rotary_seq_len = max(kv_seq_len, position_ids.max().item() + 1)
         cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
         query_states, key_states = apply_rotary_pos_emb(
             query_states, key_states, cos, sin, position_ids)
diff --git a/xtuner/model/modules/dispatch/phi3.py b/xtuner/model/modules/dispatch/phi3.py
index 4003c9d62..97ebc8d33 100644
--- a/xtuner/model/modules/dispatch/phi3.py
+++ b/xtuner/model/modules/dispatch/phi3.py
@@ -233,6 +233,11 @@ def phi3_attn_forward(
             pre_process_for_sequence_parallel_attn(
                 query_states, key_states, value_states,
                 scatter_dim=2, gather_dim=1)
+        # num_heads has been changed because of sequence parallel
+        # `self.num_heads`` is not used in self._flash_attention_forward
+        # in mistral/mixtral, we are doing this to avoid some unnecessary risk
+        ori_num_head = self.num_heads
+        self.num_heads = query_states.shape[-2]
 
     attn_output = self._flash_attention_forward(
         query_states,
@@ -248,6 +253,7 @@ def phi3_attn_forward(
         # (b, s, nd // sp_world_size, dim) -> (b, s // sp_world_size, nd, dim)
         attn_output = post_process_for_sequence_parallel_attn(
             attn_output, scatter_dim=1, gather_dim=2)
+        self.num_heads = ori_num_head
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
     attn_output = self.o_proj(attn_output)
@@ -333,7 +339,7 @@ def phi3_varlen_attn_forward(
                                                        self.layer_idx)
 
     assert position_ids is not None
-    rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+    rotary_seq_len = max(kv_seq_len, position_ids.max().item() + 1)
     cos, sin = self.rotary_emb(
         value_states, position_ids, seq_len=rotary_seq_len)
 
diff --git a/xtuner/model/modules/dispatch/qwen2.py b/xtuner/model/modules/dispatch/qwen2.py
index d89dbf947..1c8c5a8d0 100644
--- a/xtuner/model/modules/dispatch/qwen2.py
+++ b/xtuner/model/modules/dispatch/qwen2.py
@@ -151,6 +151,11 @@ def qwen2_attn_forward(
         query_states, key_states, value_states = \
             pre_process_for_sequence_parallel_attn(
                 query_states, key_states, value_states)
+        # num_heads has been changed because of sequence parallel
+        # `self.num_heads`` is not used in self._flash_attention_forward
+        # in mistral/mixtral, we are doing this to avoid some unnecessary risk
+        ori_num_head = self.num_heads
+        self.num_heads = query_states.shape[-2]
 
     attn_output = self._flash_attention_forward(
         query_states,
@@ -164,6 +169,7 @@ def qwen2_attn_forward(
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
+        self.num_heads = ori_num_head
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
     attn_output = self.o_proj(attn_output)
@@ -227,7 +233,7 @@ def qwen2_varlen_attn_forward(
                                                        self.layer_idx)
 
     assert position_ids is not None
-    rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item() + 1)
+    rotary_seq_len = max(kv_seq_len, position_ids.max().item() + 1)
     cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
 
     query_states, key_states = apply_rotary_pos_emb(query_states, key_states,

From 3617c9868fb3a853c06488ef99029ed3243fdd24 Mon Sep 17 00:00:00 2001
From: RangiLyu <lyuchqi@gmail.com>
Date: Fri, 19 Jul 2024 17:14:35 +0800
Subject: [PATCH 18/29] [Docs]: update readme and DPO en docs (#853)

* [Docs]: update readme and DPO en docs

* update link
---
 README.md                                     |   4 +
 README_zh-CN.md                               |   4 +
 docs/en/dpo/modify_settings.md                |  83 +++++++++++++
 docs/en/dpo/overview.md                       |  27 +++++
 docs/en/dpo/quick_start.md                    |  71 +++++++++++
 docs/en/index.rst                             |  17 +++
 docs/en/reward_model/modify_settings.md       | 100 ++++++++++++++++
 docs/en/reward_model/overview.md              |  43 +++++++
 docs/en/reward_model/preference_data.md       | 110 ++++++++++++++++++
 docs/en/reward_model/quick_start.md           |  85 ++++++++++++++
 docs/zh_cn/dpo/modify_settings.md             |   2 +-
 docs/zh_cn/dpo/overview.md                    |   2 +
 .../reward_model/images/sequence_parallel.png | Bin 0 -> 38382 bytes
 docs/zh_cn/reward_model/overview.md           |  14 +++
 14 files changed, 561 insertions(+), 1 deletion(-)
 create mode 100644 docs/en/dpo/modify_settings.md
 create mode 100644 docs/en/dpo/overview.md
 create mode 100644 docs/en/dpo/quick_start.md
 create mode 100644 docs/en/reward_model/modify_settings.md
 create mode 100644 docs/en/reward_model/overview.md
 create mode 100644 docs/en/reward_model/preference_data.md
 create mode 100644 docs/en/reward_model/quick_start.md
 create mode 100644 docs/zh_cn/reward_model/images/sequence_parallel.png

diff --git a/README.md b/README.md
index 4e729226c..2a8eb2879 100644
--- a/README.md
+++ b/README.md
@@ -39,6 +39,7 @@ English | [简体中文](README_zh-CN.md)
 
 ## 🎉 News
 
+- **\[2024/07\]** Support [DPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/dpo), [ORPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/orpo) and [Reward Model](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/reward_model) training with packed data and sequence parallel! See [documents](https://xtuner.readthedocs.io/en/latest/dpo/overview.html) for more details.
 - **\[2024/07\]** Support [InternLM 2.5](xtuner/configs/internlm/internlm2_5_chat_7b/) models!
 - **\[2024/06\]** Support [DeepSeek V2](xtuner/configs/deepseek/deepseek_v2_chat/) models! **2x faster!**
 - **\[2024/04\]** [LLaVA-Phi-3-mini](https://huggingface.co/xtuner/llava-phi-3-mini-hf) is released! Click [here](xtuner/configs/llava/phi3_mini_4k_instruct_clip_vit_large_p14_336) for details!
@@ -144,6 +145,9 @@ XTuner is an efficient, flexible and full-featured toolkit for fine-tuning large
   <li><a href="http://arxiv.org/abs/2305.14314">QLoRA</a></li>
   <li><a href="http://arxiv.org/abs/2106.09685">LoRA</a></li>
   <li>Full parameter fine-tune</li>
+  <li><a href="https://arxiv.org/abs/2305.18290">DPO</a></li>
+  <li><a href="https://arxiv.org/abs/2403.07691">ORPO</a></li>
+  <li>Reward Model</a></li>
 </ul>
 </td>
 </tr>
diff --git a/README_zh-CN.md b/README_zh-CN.md
index 16c1a2af2..58076210f 100644
--- a/README_zh-CN.md
+++ b/README_zh-CN.md
@@ -39,6 +39,7 @@
 
 ## 🎉 更新
 
+- **\[2024/07\]** 支持训练 [DPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/dpo)， [ORPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/orpo) 还有 [Reward Model](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/reward_model) ! 并且能够支持打包数据以及序列并行功能！ 请参考 [文档](https://xtuner.readthedocs.io/zh-cn/latest/dpo/overview.html) 了解更多信息。
 - **\[2024/07\]** 支持 [InternLM 2.5](xtuner/configs/internlm/internlm2_5_chat_7b/) 模型!
 - **\[2024/06\]** 支持 [DeepSeek V2](xtuner/configs/deepseek/deepseek_v2_chat/) models! **训练速度提升一倍！**
 - **\[2024/04\]** 多模态大模型 [LLaVA-Phi-3-mini](https://huggingface.co/xtuner/llava-phi-3-mini-hf) 发布！快速开始请查阅此[文档](xtuner/configs/llava/phi3_mini_4k_instruct_clip_vit_large_p14_336)！
@@ -144,6 +145,9 @@ XTuner 是一个高效、灵活、全能的轻量化大模型微调工具库。
   <li><a href="http://arxiv.org/abs/2305.14314">QLoRA</a></li>
   <li><a href="http://arxiv.org/abs/2106.09685">LoRA</a></li>
   <li>全量参数微调</li>
+  <li><a href="https://arxiv.org/abs/2305.18290">DPO</a></li>
+  <li><a href="https://arxiv.org/abs/2403.07691">ORPO</a></li>
+  <li>Reward Model</a></li>
 </ul>
 </td>
 </tr>
diff --git a/docs/en/dpo/modify_settings.md b/docs/en/dpo/modify_settings.md
new file mode 100644
index 000000000..d78cc40e6
--- /dev/null
+++ b/docs/en/dpo/modify_settings.md
@@ -0,0 +1,83 @@
+## Modify DPO Training Configuration
+
+This section introduces config parameters related to DPO (Direct Preference Optimization) training. For more details on XTuner config files, please refer to [Modifying Training Configuration](https://xtuner.readthedocs.io/zh-cn/latest/training/modify_settings.html).
+
+### Loss Function
+
+In DPO training, you can choose different types of loss functions according to your needs. XTuner provides various loss function options, such as `sigmoid`, `hinge`, `ipo`, etc. You can select the desired loss function type by setting the `dpo_loss_type` parameter.
+
+Additionally, you can control the temperature coefficient in the loss function by adjusting the `loss_beta` parameter. The `label_smoothing` parameter can be used for smoothing labels.
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+dpo_loss_type = 'sigmoid'  # One of ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'sppo_hard', 'nca_pair', 'robust']
+loss_beta = 0.1
+label_smoothing = 0.0
+```
+
+### Modifying the Model
+
+Users can modify `pretrained_model_name_or_path` to change the pretrained model.
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'
+```
+
+### Training Data
+
+In DPO training, you can specify the maximum number of tokens for a single sample sequence using the `max_length` parameter. XTuner will automatically truncate or pad the data.
+
+```python
+# Data
+max_length = 2048
+```
+
+In the configuration file, we use the `train_dataset` field to specify the training dataset. You can specify the dataset loading method using the `dataset` field and the dataset mapping function using the `dataset_map_fn` field.
+
+```python
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(type=load_dataset, path='mlabonne/orpo-dpo-mix-40k'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=True,
+    is_reward=False,
+    reward_token_id=-1,
+    num_proc=32,
+    use_varlen_attn=use_varlen_attn,
+    max_packed_length=max_packed_length,
+    shuffle_before_pack=True,
+)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(
+        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))
+```
+
+In the above configuration, we use `load_dataset` to load the `mlabonne/orpo-dpo-mix-40k` dataset from Hugging Face and use `orpo_dpo_mix_40k_map_fn` as the dataset mapping function.
+
+For more information on handling datasets and writing dataset mapping functions, please refer to the [Preference Dataset Section](../reward_model/preference_data.md).
+
+### Accelerating Training
+
+When training with preference data, we recommend enabling the [Variable-Length Attention Mechanism](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/varlen_flash_attn.html) to avoid memory waste caused by length differences between chosen and rejected samples within a single preference. You can enable the variable-length attention mechanism by setting `use_varlen_attn=True`.
+
+XTuner also supports many training acceleration methods. For details on how to use them, please refer to the [Acceleration Strategies Section](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/hyper_parameters.html).
diff --git a/docs/en/dpo/overview.md b/docs/en/dpo/overview.md
new file mode 100644
index 000000000..0c20946e3
--- /dev/null
+++ b/docs/en/dpo/overview.md
@@ -0,0 +1,27 @@
+## Introduction to DPO
+
+### Overview
+
+DPO (Direct Preference Optimization) is a method used in large language model training for directly optimizing human preferences. Unlike traditional reinforcement learning methods, DPO directly uses human preference data to optimize the model, thereby improving the quality of generated content to better align with human preferences. DPO also eliminates the need to train a Reward Model and a Critic Model, avoiding the complexity of reinforcement learning algorithms, reducing training overhead, and enhancing training efficiency.
+
+Many algorithms have made certain improvements to DPO's loss function. In XTuner, besides DPO, we have also implemented loss functions from papers such as [Identity Preference Optimization (IPO)](https://huggingface.co/papers/2310.12036). To use these algorithms, please refer to the [Modify DPO Settings](./modify_settings.md) section. We also provide some [example configurations](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/dpo) for reference.
+
+In addition to DPO, there are alignment algorithms like [ORPO](https://arxiv.org/abs/2403.07691) that do not require a reference model. ORPO uses the concept of odds ratio to optimize the model by penalizing rejected samples during the training process, thereby adapting more effectively to the chosen samples. ORPO eliminates the dependence on a reference model, making the training process more simplified and efficient. The training method for ORPO in XTuner is very similar to DPO, and we provide some [example configurations](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/orpo). Users can refer to the DPO tutorial to modify the configuration.
+
+### Features of DPO Training in XTuner
+
+DPO training in XTuner offers the following significant advantages:
+
+1. **Latest Algorithms**: In addition to supporting standard DPO, XTuner also supports improved DPO algorithms or memory efficient algorithms like ORPO that do not rely on reference models.
+
+2. **Reducing Memory Waste**: Due to the length differences in chosen and rejected data in preference datasets, padding tokens during data concatenation can cause memory waste. In XTuner, by utilizing the variable-length attention feature from Flash Attention2, preference pairs are packed into the same sequence during training, significantly reducing memory waste caused by padding tokens. This not only improves memory efficiency but also allows for training larger models or handling more data under the same hardware conditions.
+
+   ![img](../../zh_cn/reward_model/images/var_len_atten.png)
+
+3. **Efficient Training**:  Leveraging XTuner's QLoRA training capabilities, the reference model can be converted into a policy model with the LoRA adapter removed, eliminating the memory overhead of the reference model weights and significantly reducing DPO training costs.
+
+4. **Long Text Training**: With XTuner's sequence parallel functionality, long text data can be trained efficiently.
+
+### Getting Started
+
+Refer to the [Quick Start Guide](./quick_start.md) to understand the basic concepts. For more information on configuring training parameters, please see the [Modify DPO Settings](./modify_settings.md) section.
diff --git a/docs/en/dpo/quick_start.md b/docs/en/dpo/quick_start.md
new file mode 100644
index 000000000..19fffbf8b
--- /dev/null
+++ b/docs/en/dpo/quick_start.md
@@ -0,0 +1,71 @@
+## Quick Start with DPO
+
+In this section, we will introduce how to use XTuner to train a 1.8B DPO (Direct Preference Optimization) model to help you get started quickly.
+
+### Preparing Pretrained Model Weights
+
+We use the model [InternLM2-chat-1.8b-sft](https://huggingface.co/internlm/internlm2-chat-1_8b-sft), as the initial model for DPO training to align human preferences.
+
+Set `pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'` in the training configuration file, and the model files will be automatically downloaded when training starts. If you need to download the model weights manually, please refer to the section [Preparing Pretrained Model Weights](https://xtuner.readthedocs.io/zh-cn/latest/preparation/pretrained_model.html), which provides detailed instructions on how to download model weights from Huggingface or Modelscope. Here are the links to the models on HuggingFace and ModelScope:
+
+- HuggingFace link: https://huggingface.co/internlm/internlm2-chat-1_8b-sft
+- ModelScope link: https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-chat-1_8b-sft/summary
+
+### Preparing Training Data
+
+In this tutorial, we use the [mlabonne/orpo-dpo-mix-40k](https://huggingface.co/datasets/mlabonne/orpo-dpo-mix-40k) dataset from Huggingface as an example.
+
+```python
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='mlabonne/orpo-dpo-mix-40k'),
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=True,
+    is_reward=False,
+)
+```
+
+Using the above configuration in the configuration file will automatically download and process this dataset. If you want to use other open-source datasets from Huggingface or custom datasets, please refer to the [Preference Dataset](../reward_model/preference_data.md) section.
+
+### Preparing Configuration File
+
+XTuner provides several ready-to-use configuration files, which can be viewed using `xtuner list-cfg`. Execute the following command to copy a configuration file to the current directory.
+
+```bash
+xtuner copy-cfg internlm2_chat_1_8b_dpo_full .
+```
+
+Open the copied configuration file. If you choose to download the model and dataset automatically, no modifications are needed. If you want to specify paths to your pre-downloaded model and dataset, modify the `pretrained_model_name_or_path` and the `path` parameter in `dataset` under `train_dataset`.
+
+For more training parameter configurations, please refer to the section [Modifying DPO Training Configuration](./modify_settings.md) section.
+
+### Starting the Training
+
+After completing the above steps, you can start the training task using the following commands.
+
+```bash
+# Single machine, single GPU
+xtuner train ./internlm2_chat_1_8b_dpo_full_copy.py
+# Single machine, multiple GPUs
+NPROC_PER_NODE=${GPU_NUM} xtuner train ./internlm2_chat_1_8b_dpo_full_copy.py
+# Slurm cluster
+srun ${SRUN_ARGS} xtuner train ./internlm2_chat_1_8b_dpo_full_copy.py --launcher slurm
+```
+
+### Model Conversion
+
+XTuner provides integrated tools to convert models to HuggingFace format. Simply execute the following commands:
+
+```bash
+# Create a directory for HuggingFace format parameters
+mkdir work_dirs/internlm2_chat_1_8b_dpo_full_copy/iter_15230_hf
+
+# Convert format
+xtuner convert pth_to_hf internlm2_chat_1_8b_dpo_full_copy.py \
+                            work_dirs/internlm2_chat_1_8b_dpo_full_copy/iter_15230.pth \
+                            work_dirs/internlm2_chat_1_8b_dpo_full_copy/iter_15230_hf
+```
+
+This will convert the XTuner's ckpt to the HuggingFace format.
diff --git a/docs/en/index.rst b/docs/en/index.rst
index c702e0a04..c4c18d31a 100644
--- a/docs/en/index.rst
+++ b/docs/en/index.rst
@@ -56,6 +56,23 @@ Documentation
    training/open_source_dataset.rst
    training/visualization.rst
 
+.. toctree::
+   :maxdepth: 2
+   :caption: DPO
+
+   dpo/overview.md
+   dpo/quick_start.md
+   dpo/modify_settings.md
+
+.. toctree::
+   :maxdepth: 2
+   :caption: Reward Model
+
+   reward_model/overview.md
+   reward_model/quick_start.md
+   reward_model/modify_settings.md
+   reward_model/preference_data.md
+
 .. toctree::
    :maxdepth: 2
    :caption: Acceleration
diff --git a/docs/en/reward_model/modify_settings.md b/docs/en/reward_model/modify_settings.md
new file mode 100644
index 000000000..4f41ca300
--- /dev/null
+++ b/docs/en/reward_model/modify_settings.md
@@ -0,0 +1,100 @@
+## Modify Reward Model Training Configuration
+
+This section introduces the config related to Reward Model training. For more details on XTuner config files, please refer to [Modify Settings](https://xtuner.readthedocs.io/zh-cn/latest/training/modify_settings.html).
+
+### Loss Function
+
+XTuner uses the [Bradley–Terry Model](https://en.wikipedia.org/wiki/Bradley%E2%80%93Terry_model) for preference modeling in the Reward Model. You can specify `loss_type="ranking"` to use ranking loss. XTuner also implements the focal loss function proposed in InternLM2, which adjusts the weights of difficult and easy samples to avoid overfitting. You can set `loss_type="focal"` to use this loss function. For a detailed explanation of this loss function, please refer to the [InternLM2 Technical Report](https://arxiv.org/abs/2403.17297).
+
+Additionally, to maintain stable reward model output scores, we have added a constraint term in the loss. You can specify `penalty_type='log_barrier'` or `penalty_type='L2'` to enable log barrier or L2 constraints, respectively.
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+loss_type = 'focal'  # 'ranking' or 'focal'
+penalty_type = 'log_barrier'  # 'log_barrier' or 'L2'
+```
+
+### Modifying the Model
+
+Users can modify `pretrained_model_name_or_path` to change the pretrained model.
+
+Note that XTuner calculates reward scores by appending a special token at the end of the data. Therefore, when switching models with different vocabularies, the ID of this special token also needs to be modified accordingly. We usually use an unused token at the end of the vocabulary as the reward token.
+
+For example, in InternLM2, we use `[UNUSED_TOKEN_130]` as the reward token:
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'
+reward_token_id = 92527  # use [UNUSED_TOKEN_130] as reward token
+```
+
+If the user switches to the llama3 model, we can use `<|reserved_special_token_0|>` as the reward token:
+
+```python
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'meta-llama/Meta-Llama-3-8B-Instruct'
+reward_token_id = 128002  # use <|reserved_special_token_0|> as reward token
+```
+
+### Training Data
+
+In Reward Model training, you can specify the maximum number of tokens for a single sample sequence using `max_length`. XTuner will automatically truncate or pad the data.
+
+```python
+# Data
+max_length = 2048
+```
+
+In the configuration file, we use the `train_dataset` field to specify the training dataset. You can specify the dataset loading method using the `dataset` field and the dataset mapping function using the `dataset_map_fn` field.
+
+```python
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='argilla/ultrafeedback-binarized-preferences-cleaned'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=False,
+    is_reward=True,
+    reward_token_id=reward_token_id,
+    num_proc=32,
+    use_varlen_attn=use_varlen_attn,
+    max_packed_length=max_packed_length,
+    shuffle_before_pack=True,
+)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(
+        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))
+```
+
+In the above configuration, we use `load_dataset` to load the `argilla/ultrafeedback-binarized-preferences-cleaned` dataset from Hugging Face, using `orpo_dpo_mix_40k_map_fn` as the dataset mapping function (this is because `orpo_dpo_mix_40k` and `ultrafeedback-binarized-preferences-cleaned` have the same format, so the same mapping function is used).
+
+For more information on handling datasets and writing dataset mapping functions, please refer to the [Preference Data Section](./preference_data.md).
+
+### Accelerating Training
+
+When training with preference data, we recommend enabling the [Variable-Length Attention Mechanism](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/varlen_flash_attn.html) to avoid memory waste caused by length differences between chosen and rejected samples within a single preference. You can enable the variable-length attention mechanism by setting `use_varlen_attn=True`.
+
+XTuner also supports many training acceleration methods. For details on how to use them, please refer to the [Acceleration Strategies Section](https://xtuner.readthedocs.io/zh-cn/latest/acceleration/hyper_parameters.html).
diff --git a/docs/en/reward_model/overview.md b/docs/en/reward_model/overview.md
new file mode 100644
index 000000000..eb210140c
--- /dev/null
+++ b/docs/en/reward_model/overview.md
@@ -0,0 +1,43 @@
+## Introduction to Reward Model
+
+### Overview
+
+The Reward Model is a crucial component in the reinforcement learning process. Its primary task is to predict reward values based on given inputs, guiding the direction of the learning algorithm. In RLHF (Reinforcement Learning from Human Feedback), the Reward Model acts as a proxy for human preferences, helping the reinforcement learning algorithm optimize strategies more effectively.
+
+In large language model training, the Reward Model typically refers to the Preference Model. By providing good and bad (chosen & rejected) responses to the same prompts during training, it fits human preferences and predicts a reward value during inference to guide the optimization of the Actor model in the RLHF process.
+
+Applications of the Reward Model include but are not limited to:
+
+- **RLHF Training**: During RLHF training such as the Proximal Policy Optimization (PPO) algorithm, the Reward Model provides reward signals, improve the quality of generated content, and align it more closely with human preferences.
+- **BoN Sampling**: In the Best-of-N (BoN) sampling process, users can use the Reward Model to score multiple responses to the same prompt and select the highest-scoring generated result, thereby enhancing the model's output.
+- **Data Construction**: The Reward Model can be used to evaluate and filter training data or replace manual annotation to construct DPO training data.
+
+### Features of Reward Model Training in XTuner
+
+The Reward Model training in XTuner offers the following significant advantages:
+
+1. **Latest Training Techniques**: XTuner integrates the Reward Model training loss function from InternLM2, which stabilizes the numerical range of reward scores and reduces overfitting on simple samples (see [InternLM2 Technical Report](https://arxiv.org/abs/2403.17297) for details).
+
+2. **Reducing Memory Waste**: Due to the length differences in chosen and rejected data in preference datasets, padding tokens during data concatenation can cause memory waste. In XTuner, by utilizing the variable-length attention feature from Flash Attention2, preference pairs are packed into the same sequence during training, significantly reducing memory waste caused by padding tokens. This not only improves memory efficiency but also allows for training larger models or handling more data under the same hardware conditions.
+
+![img](../../zh_cn/reward_model/images/var_len_atten.png)
+
+3. **Efficient Training**: Leveraging XTuner's QLoRA training capabilities, we can perform full parameter training only on the Reward Model's Value Head, while using QLoRA fine-tuning on the language model itself, substantially reducing the memory overhead of model training.
+
+4. **Long Text Training**: With XTuner's sequence parallel functionality, long text data can be trained efficiently.
+
+![img](../../zh_cn/reward_model/images/sequence_parallel.png)
+
+### Getting Started
+
+Refer to the [Quick Start Guide](./quick_start.md) to understand the basic concepts. For more information on configuring training parameters, please see the [Modifying Reward Model Settings](./modify_settings.md) section.
+
+### Open-source Models
+
+We use XTuner to train the  InternLM2 Reward Models from the InternLM2 Technical Report, welcome to download and use:
+
+| Model                     | Transformers(HF)                                                                 | ModelScope(HF)                                                                                             | OpenXLab(HF)                                                                                                                                                | RewardBench Score |
+| ------------------------- | -------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------- |
+| **InternLM2-1.8B-Reward** | [🤗internlm2-1_8b-reward](https://huggingface.co/internlm/internlm2-1_8b-reward) | [internlm2-1_8b-reward](https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-1_8b-reward/summary) | [![Open in OpenXLab](https://cdn-static.openxlab.org.cn/header/openxlab_models.svg)](https://openxlab.org.cn/models/detail/OpenLMLab/internlm2-1_8b-reward) | 80.6              |
+| **InternLM2-7B-Reward**   | [🤗internlm2-7b-reward](https://huggingface.co/internlm/internlm2-7b-reward)     | [internlm2-7b-reward](https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-7b-reward/summary)     | [![Open in OpenXLab](https://cdn-static.openxlab.org.cn/header/openxlab_models.svg)](https://openxlab.org.cn/models/detail/OpenLMLab/internlm2-7b-reward)   | 86.6              |
+| **InternLM2-20B-Reward**  | [🤗internlm2-20b-reward](https://huggingface.co/internlm/internlm2-20b-reward)   | [internlm2-20b-reward](https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-20b-reward/summary)   | [![Open in OpenXLab](https://cdn-static.openxlab.org.cn/header/openxlab_models.svg)](https://openxlab.org.cn/models/detail/OpenLMLab/internlm2-20b-reward)  | 89.5              |
diff --git a/docs/en/reward_model/preference_data.md b/docs/en/reward_model/preference_data.md
new file mode 100644
index 000000000..2f304e627
--- /dev/null
+++ b/docs/en/reward_model/preference_data.md
@@ -0,0 +1,110 @@
+## Preference Dataset
+
+### Overview
+
+XTuner's Reward Model, along with DPO, ORPO, and other algorithms that training on preference data, adopts the same data format. Each training sample in the preference dataset needs to contain the following three fields: `prompt`, `chosen`, and `rejected`. The values for each field follow the [OpenAI chat message](https://platform.openai.com/docs/api-reference/chat/create) format. A specific example is as follows:
+
+```json
+{
+  "prompt": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": "Who won the world series in 2020?"
+    },
+    {
+      "role": "assistant",
+      "content": "The Los Angeles Dodgers won the World Series in 2020."
+    },
+    {
+      "role": "user",
+      "content": "Where was it played?"
+    }
+  ],
+  "chosen": [
+    {
+      "role": "assistant",
+      "content": "The 2020 World Series was played at Globe Life Field in Arlington, Texas."
+    }
+  ],
+  "rejected": [
+    {
+      "role": "assistant",
+      "content": "I don't know."
+    }
+  ]
+}
+```
+
+When conducting Reward Model training or DPO training, XTuner processes the preference dataset into different training labels based on the type of training task.
+
+![img](../../zh_cn/reward_model/images/preference_data.png)
+
+As shown in the above image, for Reward Model training, we follow the ChatGPT training method by adding a special `<|reward|>` token at the end of the conversation data and calculating the loss only on the logits output by this token. For DPO series algorithm training, we mask the tokens in the prompt part and calculate the loss only on the chosen and rejected responses. In the configuration file, we control the dataset type through the `is_reward` and `is_dpo` fields in the dataset.
+
+### Training with Custom Data
+
+After understanding the data format of a single sample, you only need to save the converted data in [JSON Lines](https://jsonlines.org/) format. Each line in the file is a sample.
+
+```json
+{"prompt": [{"content": "How can I learn to better regulate my emotions as a grown adult?", "role": "user"}], "chosen": [{"content": "You could do emotional regulation exercises with me, and try to notice and practice ways that you could better control your emotions in real life.", "role": "assistant"}], "rejected": [{"content": "Do you mean how can you learn to be more emotionally mature?", "role": "assistant"}]}
+{"prompt": [{"content": "What are some interesting things I can add to a grilled cheese sandwich?", "role": "user"}, {"content": "A tomato, a slice of pickle, some chopped pickles, some chopped green onions.  Also, some mayonnaise if you like.", "role": "assistant"}, {"content": "Sounds great. Anything else?", "role": "user"}], "chosen": [{"content": "If you’d like something a little different, I recommend grating some horseradish on top of the sandwich.", "role": "assistant"}], "rejected": [{"content": "Maybe some nice fresh basil?", "role": "assistant"}]}
+{"prompt": [{"content": "How do I bob for apples?", "role": "user"}], "chosen": [{"content": "Ah!  Do you mean the game of “applesauce,” or do you mean the idea of “to bob for apples”?", "role": "assistant"}], "rejected": [{"content": "Sorry, I don’t know that term.", "role": "assistant"}]}
+......
+```
+
+After preparing the custom dataset, you need to fill in the path to your saved data in the `data_files` field in the configuration file. You can load multiple JSONL files simultaneously for training.
+
+```python
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_jsonl_dataset,
+        data_files=[
+            '/your/jsonl/path/here.jsonl',
+            '/your/another/jsonl/path/here.jsonl'
+        ]),
+)
+```
+
+### Training with Open Source Datasets
+
+Similar to configuring SFT data in XTuner, when using open-source datasets from Hugging Face, you only need to define a mapping function `map_fn` to process the dataset format into XTuner's data format.
+
+Taking `Intel/orca_dpo_pairs` as an example, this dataset has `system`, `question`, `chosen`, and `rejected` fields, with each field's value in text format instead of the [OpenAI chat message](https://platform.openai.com/docs/api-reference/chat/create) format. Therefore, we need to define a mapping function for this dataset:
+
+```python
+def intel_orca_dpo_map_fn(example):
+    prompt = [{
+        'role': 'system',
+        'content': example['system']
+    }, {
+        'role': 'user',
+        'content': example['question']
+    }]
+    chosen = [{'role': 'assistant', 'content': example['chosen']}]
+    rejected = [{'role': 'assistant', 'content': example['rejected']}]
+    return {'prompt': prompt, 'chosen': chosen, 'rejected': rejected}
+```
+
+As shown in the code, `intel_orca_dpo_map_fn` processes the four fields in the original data, converting them into `prompt`, `chosen`, and `rejected` fields, and ensures each field follows the [OpenAI chat message](https://platform.openai.com/docs/api-reference/chat/create) format, maintaining uniformity in subsequent data processing flows.
+
+After defining the mapping function, you need to import it in the configuration file and configure it in the `dataset_map_fn` field.
+
+```python
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='Intel/orca_dpo_pairs'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=intel_orca_dpo_map_fn,
+)
+```
diff --git a/docs/en/reward_model/quick_start.md b/docs/en/reward_model/quick_start.md
new file mode 100644
index 000000000..5c802be2f
--- /dev/null
+++ b/docs/en/reward_model/quick_start.md
@@ -0,0 +1,85 @@
+## Quick Start Guide for Reward Model
+
+In this section, we will introduce how to use XTuner to train a 1.8B Reward Model, helping you get started quickly.
+
+### Preparing Pretrained Model Weights
+
+According to the paper [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155), we use a language model fine-tuned with SFT as the initialization model for the Reward Model. Here, we use [InternLM2-chat-1.8b-sft](https://huggingface.co/internlm/internlm2-chat-1_8b-sft) as the initialization model.
+
+Set `pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'` in the training configuration file, and the model files will be automatically downloaded when training starts. If you need to download the model weights manually, please refer to the section [Preparing Pretrained Model Weights](https://xtuner.readthedocs.io/zh-cn/latest/preparation/pretrained_model.html), which provides detailed instructions on how to download model weights from Huggingface or Modelscope. Here are the links to the models on HuggingFace and ModelScope:
+
+- HuggingFace link: https://huggingface.co/internlm/internlm2-chat-1_8b-sft
+- ModelScope link: https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-chat-1_8b-sft/summary
+
+### Preparing Training Data
+
+In this tutorial, we use the [UltraFeedback](https://arxiv.org/abs/2310.01377) dataset as an example. For convenience, we use the preprocessed [argilla/ultrafeedback-binarized-preferences-cleaned](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences-cleaned) dataset from Huggingface.
+
+```python
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(
+        type=load_dataset,
+        path='argilla/ultrafeedback-binarized-preferences-cleaned'),
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=False,
+    is_reward=True,
+)
+```
+
+Using the above configuration in the configuration file will automatically download and process this dataset. If you want to use other open-source datasets from Huggingface or custom datasets, please refer to the [Preference Dataset](./preference_data.md) section.
+
+### Preparing Configuration Files
+
+XTuner provides several ready-to-use configuration files, which can be viewed using `xtuner list-cfg`. Execute the following command to copy a configuration file to the current directory.
+
+```bash
+xtuner copy-cfg internlm2_chat_1_8b_reward_full_ultrafeedback .
+```
+
+Open the copied configuration file. If you choose to download the model and dataset automatically, no modifications are needed. If you want to specify paths to your pre-downloaded model and dataset, modify the `pretrained_model_name_or_path` and the `path` parameter in `dataset` under `train_dataset`.
+
+For more training parameter configurations, please refer to the section [Modifying Reward Training Configuration](./modify_settings.md).
+
+### Starting the Training
+
+After completing the above steps, you can start the training task using the following commands.
+
+```bash
+# Single node single GPU
+xtuner train ./internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py
+# Single node multiple GPUs
+NPROC_PER_NODE=${GPU_NUM} xtuner train ./internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py
+# Slurm cluster
+srun ${SRUN_ARGS} xtuner train ./internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py --launcher slurm
+```
+
+The correct training log should look like the following (running on a single A800 GPU):
+
+```
+06/06 16:12:11 - mmengine - INFO - Iter(train) [   10/15230]  lr: 3.9580e-07  eta: 2:59:41  time: 0.7084  data_time: 0.0044  memory: 18021  loss: 0.6270  acc: 0.0000  chosen_score_mean: 0.0000  rejected_score_mean: 0.0000  num_samples: 4.0000  num_tokens: 969.0000
+06/06 16:12:17 - mmengine - INFO - Iter(train) [   20/15230]  lr: 8.3536e-07  eta: 2:45:25  time: 0.5968  data_time: 0.0034  memory: 42180  loss: 0.6270  acc: 0.5000  chosen_score_mean: 0.0013  rejected_score_mean: 0.0010  num_samples: 4.0000  num_tokens: 1405.0000
+06/06 16:12:22 - mmengine - INFO - Iter(train) [   30/15230]  lr: 1.2749e-06  eta: 2:37:18  time: 0.5578  data_time: 0.0024  memory: 32121  loss: 0.6270  acc: 0.7500  chosen_score_mean: 0.0016  rejected_score_mean: 0.0011  num_samples: 4.0000  num_tokens: 932.0000
+06/06 16:12:28 - mmengine - INFO - Iter(train) [   40/15230]  lr: 1.7145e-06  eta: 2:36:05  time: 0.6033  data_time: 0.0025  memory: 42186  loss: 0.6270  acc: 0.7500  chosen_score_mean: 0.0027  rejected_score_mean: 0.0016  num_samples: 4.0000  num_tokens: 994.0000
+06/06 16:12:35 - mmengine - INFO - Iter(train) [   50/15230]  lr: 2.1540e-06  eta: 2:41:03  time: 0.7166  data_time: 0.0027  memory: 42186  loss: 0.6278  acc: 0.5000  chosen_score_mean: 0.0031  rejected_score_mean: 0.0032  num_samples: 4.0000  num_tokens: 2049.0000
+06/06 16:12:40 - mmengine - INFO - Iter(train) [   60/15230]  lr: 2.5936e-06  eta: 2:33:37  time: 0.4627  data_time: 0.0023  memory: 30238  loss: 0.6262  acc: 1.0000  chosen_score_mean: 0.0057  rejected_score_mean: 0.0030  num_samples: 4.0000  num_tokens: 992.0000
+06/06 16:12:46 - mmengine - INFO - Iter(train) [   70/15230]  lr: 3.0331e-06  eta: 2:33:18  time: 0.6018  data_time: 0.0025  memory: 42186  loss: 0.6247  acc: 0.7500  chosen_score_mean: 0.0117  rejected_score_mean: 0.0055  num_samples: 4.0000  num_tokens: 815.0000
+```
+
+### Model Conversion
+
+XTuner provides integrated tools to convert models to HuggingFace format. Simply execute the following commands:
+
+```bash
+# Create a directory to store HF format parameters
+mkdir work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy/iter_15230_hf
+
+# Convert the format
+xtuner convert pth_to_hf internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py \
+                            work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py/iter_15230.pth \
+                            work_dirs/internlm2_chat_1_8b_reward_full_ultrafeedback_copy.py/iter_15230_hf
+```
+
+This will convert the XTuner's ckpt to the HuggingFace format.
+
+Note: Since the Reward Model type is not integrated into the official transformers library, only the Reward Models trained with InternLM2 will be converted to the `InternLM2ForRewardModel` type. Other models will default to the `SequenceClassification` type (for example, LLaMa3 will be converted to the `LlamaForSequenceClassification` type).
diff --git a/docs/zh_cn/dpo/modify_settings.md b/docs/zh_cn/dpo/modify_settings.md
index 7b4672792..2365be25c 100644
--- a/docs/zh_cn/dpo/modify_settings.md
+++ b/docs/zh_cn/dpo/modify_settings.md
@@ -32,7 +32,7 @@ pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b-sft'
 
 ### 训练数据
 
-在 Reward Model 训练中，你可以通过 `max_length` 来指定单个样本序列的最大 token 数，XTuner 会自动对数据进行截断或是填充。
+在 DPO 训练中，你可以通过 `max_length` 来指定单个样本序列的最大 token 数，XTuner 会自动对数据进行截断或是填充。
 
 ```python
 # Data
diff --git a/docs/zh_cn/dpo/overview.md b/docs/zh_cn/dpo/overview.md
index d1bfc4379..d3c3a7aad 100644
--- a/docs/zh_cn/dpo/overview.md
+++ b/docs/zh_cn/dpo/overview.md
@@ -20,6 +20,8 @@ XTuner 中的 DPO 训练具备以下显著优势：
 
 3. **高效训练**：借助 XTuner 的 QLoRA 训练功能，参考模型能够被转化为移除LoRA适配器的语言模型，从而省去了参考模型权重的显存占用，大幅降低了 DPO 的训练开销。
 
+4. **长文本训练**: 借助 XTuner 的序列并行功能，能够对长文本数据进行训练。
+
 ### 开始训练
 
 请参阅[快速上手](./quick_start.md)来了解最基本的概念，若希望了解更多训练参数配置相关的内容，请参考[修改DPO配置](./modify_settings.md)章节。
diff --git a/docs/zh_cn/reward_model/images/sequence_parallel.png b/docs/zh_cn/reward_model/images/sequence_parallel.png
new file mode 100644
index 0000000000000000000000000000000000000000..53f86c81aa66d94d5cb523f6fa20e2c27e2f5bc1
GIT binary patch
literal 38382
zcmcG0byQqSx-SF^9>E$9?vh4=hu|LE-5Phd0Kwe}1b27$U;)xdaEIU$2p;@xPIBg+
zxo76yd2g-v2a841Yxl0&Rp0ls>X5gxV(2LNC@?TE=n~?>3NSE_p2EPu7C(LjyuuCg
zwT6Kqg^>^zRC0ygPt&!1KD~Hx@d?*sHLUGNug=d;JZq?j;mo6Ga$#4ELhAjjmT#Vy
zWF2O<p86I`4{cLE7p3eWCm_ffost)Sf&UecZxg-jqc7*nF4wdP5sV38maHi0z6kMe
zqHL_06oHTX?Wak*+Z~)K-CP*UxQ@8%R0SfCoK*#Ghto$c%i6i`k&$&?!#sw8C53?}
zj)Xa~HPJ(We+Tmf=F#idkD%{Jf#1M?eS;`rQjRKfJ28H}^9P&%JKTr&z&wtJ@io{D
zE(7%vcAZ_qKYoOH=!r?`gkY<W&AIYOOAWp<96cNLPzu`lbP7Eon04h79i$VXvGTDV
zNXYR+04f_?7)3hkds-ak|6i{XG1*aUn^l9cDbz({*$i!t-&cgHVU$^+HC7I)2BGfN
zP(m!<p|TkwuXVk<ewN>Cplx9*gIq4@m$_d)tha4MQyG>s^w>UuGf};27JS_Yi;E+V
z&QvOwm|F$QG#XAB{uu5+DtR~5sjQ2Aw<ZR@=|%Dv6KZT85DI2bY_&-)6Qb6hTJc8;
zv(y~&?3rLy|8nj?x`JeG)yi?fr@6PUwPJgIn8cD=wad?rB`?c9Gg73Xlu)30YVABz
zOcR@%>~yeFs1*Xjw51TV#&}G`t0@?ZU77<94X2by3BGC=I7_!rSzJM;O-T2X#k!`v
z!?o3iom%;HWh8-5u<{71_9>WUrTu+Tx!N?omO+oU%oF4IvVre=C}N_q4p4=G=s2oi
zb5q+_I9&UgLp1I5v=Vmr_^~vGgh&D68O!QI&*04jl37N>)k-25a`vHb_U&hCQYg%M
z<>h0k{CFjvV29^$5rTg36h{->D@(*a>x#gV@8gbo9HT=b_gpxFnoKev&68RLrm>R6
z!dBm_nj!+-*4KnVUl`9;_mh%psRa?!QOe5l@H|98J>R#qwx!u1n<6igN8C>@)sQCV
zLt#M*DZ!%w3;SEW2xcZzfw7R{=b|YwA@1gnvlm`Wzsgm@l}j8u==x|fVM5TA-Gb+r
zspS=$`?KPN+UR~@$Zr%kq#*RwrEDmSt-k)$>ZdD9<lpn+`scj#+UAw_zGuQzm3U%Y
zdQnZW2P+s&q0V%l@KiW~=a{pijA444Lc$=*@CEdV7PfeJwd~;7i|GWVU6A;DqwpgP
z0>QV#{1VOE`o{MTlSH{{A|HPb7`HoMBiAwf*MRZ<958$N`rFQy?hS%3&XnS@Sc}kS
z@H?+)>CVvR55x@^Vs2X!pH*DS3-sW3D2xP`?141Jv%}A{;+xPs@I*9T%G;m|zKvXg
z3=gQ0aV3<QVvVBZkyGf7;GrN!zRw;YlWmV9gV9n;o@}h-S4-(bY<vw)WT8N7WY#aH
zpQ^)8avie}x<%nItr4)LSg~=QqkHW44Q`r(7GysPufB$cCYsZx({IkR{1h8(x4f(6
zBU&*cpLRGB{S7Ia+Qi{C9*P_1OQmw|XBMKjAT5Vi?yS#dhG%!RvMhSpg(TteX|luZ
zh!d~DbjaCrL?2>Ncz(~k%G#$ZJu_hDHQt+iW^9EQ9G!pf8K7N0WIF$ZIJWMlM&8Vw
z={SU+TI5N+O4Ds~@W62^Z&VB3;Su+A6`M#bg@p3_Pa@y_%na!-nk5=KA9A;qTen;Z
zeW8z;Qb_5v-gqY$8Au;@A<MLr^FSs(RS$+;*>9-cyh~#+<n6|<G4{yrP9QvO!MBt~
zkcfY3gkU(ojiL2Lg-4;=+muKs23~^oobj&kiy$HLE7ftx6QwksnXaw5d<t6ZGb^L9
z?Lv*EX<Xa*tw6(&kRR4dDj}ObqjjwXV)=|0zPN>^;<54JKO1I1p%jyA`6%3%vQOZc
z7`||8v0V2jr(-0|xmKkJ#!|@d42B?Hl?9c5{yw~Ggcd0!3hsfFq>(+Y{(7*mapL5v
zO`ztev4*1%oBR1SXm9_=HRjcA8B^0`y4_ixEjwJ@P|net?YeCCZc-B*PpQd_s?6__
z>@2Rub~Q@&je6%JZ(bNY(GHtL%O4(&uC;i5XY4U)Lrv>JB(=#V)QW%3HoKl-f~C22
zifJr1AC}`Fv;$XGobet(tVJ^?-Hx#{;kj@gIa@k26YJFX6BpE@R+GYFj3QzrktY+*
zuySDWR4;Uq(a30NcUkbur!LrnN*>vRH;8Pf;^p(k2209?m|Qd8>s1HGpJMcXYa=cX
zC8G@#h)zw3?%2rVWqGW)U?g=Q^#nJ1s(sxS^yLn8meqx&`3dEPJ`cX$_qOoio|w!Q
z-@GA-YQNQbBm>Dj&NwXYkIrmELAk12tzLs8?5y4k-P2zSxv~N;PECBe^8#<nxfAKA
z*_0>46_)cpzl<%e=iwoDNOPStxMCVXBkmKi=yR@U3Dt8J7M74Ctc}AuoHZUY7nQVC
zelCo##YG%)iZB{|Y2QL3Qx<x?o^@JirUs60{|s@Ff{3C-tiRYMim~hpD^gN$%TrKn
zau1XQ_rT(6i@w$TM5Z#2HE`)6F9%ZtH>ra&g_g%4K3LC3bL!3&N;+sZPo?Whx+1Lj
z{MCKkA=6v*PgoaJl7AgMZ9eF^zA7+<$L;OGwxeC~PfrtV51x`+@%Y7bMHfVp4O+E7
zl^}E^_ea0`EY4ONK{@yxB3Z=C^igdC1HIos2-Xx`gcIR;Z2Z3HTI(g;n^SwXg)d{+
z93yYb?bJ1;mp49ibvLR(GhH5k!aziK#(I9|MA5{Gw^K<wSn4VDvL#zmB~BYs9bZPC
zNVJ<y@Xart>N>UI*U*nys7-Z2)8u;=<@bJzM9_Mfn2$abF^!4h;g6RPm7psvami<B
zOf2co>cQ}4CptJJ8sJwG4GejH1oE8KrjHZwc8TQk89Y_9%%9{hPcoL~I>O+pnC;kN
zr{0odkD+bMGw{%hOZFlm=(D6Nq{KVac=lMxXUPTYNFmRU)~w;j+x#dVksVyD3N;!1
zc&#Vi_)&bLnR<g+G<I<v^<4B@aZ!m{6L?WANb0-KSV|Q*7T<B4VvYY`Yfz${2Q*|B
z__Bg-xt?i2_jWtzTs$W0v{$r6^9^`C!ydkkNcr<yz6%Y8mv~P{*-Q*%!S-~1cg#66
zR$0QLu`i!aAjxtmd#xNlk&DGNr?1-9OkaPr7RyF#uk>VsyCeD#j(LhBZvs3>r0=LX
z_>y3-*3`|?g8yLhMU7EEqJFcHS-D>SJJgaNp{rk^cViZP^_KHW=kxov^$E18wZ{n}
zEo5A<(G^1-ZJYUbnG!|oCa3T70zD%qn<R>{Ikuj%i77u@%OWBaqQw*#U~_$M`J7+n
zPMU)$P%4(<Ih)Wk1Xap`2^J$R5@|`ucKpJl!U|Kuf!k3gj`;SOXg4WHARF=e5VGH<
zfAJKtU1x70yMJv{R8RoFAYO1!`NnPm1dCQMel}iI84Led3s$LNT)0>Z{FY59b`(C~
z<zV}8A`xdoh<ovRmuj`D(*0-IxWsv6L9n#<w{}z)M7cAUm>Wqcp@<>9bIh-3v80@I
z9sThsoVz1$O%Sp&T(%QJ&*2{}#^fF+D5n#2y`qATg6}4m<ny;X>)FbGQbhOUn=siN
zu5B6F$<PzhBoHeXJ~Dew-7AN>4?ffe>`aSZ$yuOS8p|XxH)~<~!3}ny!821=53(5k
z4xxkj$KetR-ZX8z6WV<7?LQzc7LY)SK8#E|MR>-jRP(b&j@b|Pj~HVdd1EY62N5#r
z<DZGWVk^lEbzr~tgff+fQaldjNw(S?oFAS9k47aGih){Fi7ZmltVu&&rX0Q=NBvj_
z=}By@MTCf^FQf1PWmQ!T=%(|^YcEr7!Pw5(A6`3;3<TTG1SFQz8n)q5nvsR8aFwik
zwLM)8j-y0&2jfGY?Y(|xI8Pjo7SOgw(<Gzz4C$>8V$B$}RyazU?6vvucv5Vie<WS#
z?0S?|WIr>?=Va%=f*wdsL~*5udv?{BTJF$5JgtR=H}%_4>g>L}nwKJAil%NHcIhmP
z11&x{FS~)Y{87jE_9<4j8~xKG=y|$Z7(|X_Um?f}5_GTT7RgdxOuv!a+%%0>bc|T6
z>c+*7AxI9DZu*>{LN~Rss+t2r7Ut?LwR3H*$>Fd^-1dnu5mzXlmPmRbqp)d8x>>Tm
zgNmMR+YrX2n9QsS9W5ux8_rkn58>|+;29J_?_#)+IuJWi+M<#&*5s^<ns4a9p0vcS
zJx#Q`pE~xIWpAVy&^wp=)`%MMg28lu-al)x-V;-I*WQ|a14V`1n(jp{w`Bt44ztbO
zdO@jl9y&P+p;1bJj?sPIV+}hZ#_R|k=I_0<%YA;_O2UR2l1pcGQy$GLzdeteLYpzM
zs&K9K-lTA`)9%G1#0At3NmuOX<#|(RZtK$4c&W;4IiVNC!Y3Y{JZ^~4Y!~<~L}#Cp
z?doz9PkKVi(-+Y0LMSIBFfpje6mdaDrk`w>tr3ll9CT!rcv1D1%6I6Dttry@Bt3rX
zHkv|l3&Zp}A~t4fUlz|(Y*bV1Sv<J(PFc3)eRwu#_;p57Mb>Bgxq8Y*Lhn6LQppP@
z)@pyjAqOHc$7gLiNLOxQ^#rS_8_TUtZd|_p7=f`C*+mdNf}nad!RI^LzC-r?s;p2N
zCHO(oB;ytxf5aqKbyw9`v_GKnyFZa4dU<xri;8RExdpzenarq?&lqt@xQx_0?B$)W
zDQbrz^FMu-nxjE?oWPfZJkb_2RH4)IOAaLw%fT#^@!f{J`68>3HH&WTleAoiU#6VL
zR?B3X*w+|ONpx9`g+Hew8(p%nIyAOoQd`?{LXo2SZc{u4YkuY2B`8Ni-Ea_cr$@oD
zQkV3+hmK_KSyiP87Sp5A@C`!7#*6$Artd+5;96d3KJbdQIX8}7ewVT|IR4Cd8eyF^
z%D@#><ShgfO#~dDF*$3)?S?-07(Y#r%VR?qn%!!TTdXGj)tFf`F~gw9(|67bVnZpx
z8#2W?6cMf7Zn?Ears76EJ$jS=GuphNFg#gu;6#*CV&BPROWc^9yumArB1ExBNm5C4
zGRw|=LzeATZKeqp?Zd4R(EoIKxV8Sj-`cttLvZ!NCWrz*V4om?oqSm+6`5xU`Pugl
zia?GmJLcJd(FDW~1}X9fk7WiZg4P9@o0<xTa1K1Rc6a!Mxz=7WVBX7#jC~D-fKY3(
z2xGJl!1Dod&}s@wpiR|m47%GvIE%AW<Z~Ybc-yMbGmKzdIO+T(6Qn_$pANF+WG4!H
zHNHZZ51J)uYTpKKWI!oD*+l%9#&&R4H8;IVTYZr78i7vEGCnt<1jJtcy5imD1)9in
z0Q*5A83pZ;-XcZ3tWr9#A2s2^7IiA2Q+-9tY*;h+795HN3keiS7vdsF36f7`9Ytfr
zR2)HTOihUt`DkW4kRrH__$aH+8}kaT%})M})Wnkwj*BC4N1F9ny&4;oo=^+X-P|}j
z^7y;(6<Hqq&B6(==z89^N{f)tkx@OS)(5l>KX##VjU9i%ota8bF#}wobi`V($8B7A
z8KdTX!lS-8qmGEo{bWdK^(bs><2><LsP!_~!1F-kyGrmZcf+d#!9qumS<b$^Ybn}M
z>-m6~lnqLm!l^w(xH{uy`pJ>i8=Jtiu=2QA2x5ZyrA`JQzqI?Cql1IBP&?9<a;TXo
z6*0tMU0ZtLQ~WuqWA79g#XQiFalc%MD^HaN4mZa)0!>0;VO4MfDLw5zirX$b4TB9*
z?6=_@U<gM{AQG#5ApSBKCoR#XdGZO95_~2I#|>k{5<)C$)%W_ZYq(yiF+Wx!FH5v#
zdrGnZ_)CR5iR=QiG4vg+=PV2^L{#4>UyFUkI@BkOOZ}PFG2)6+8mA{K&d&Ny^?vg$
z&g%w>MuyHjaU2-crz`=?5A2&J{7ZZZP;8EDSgED_>9KjfeKV)24ktBBTn2^H`sb+d
zx8!xz9XDYOz^>3e%Ku>ur<Dl`i1}bY`L2cbj&Tzn&{1gqOGo)bEvbTa;T6l%M+fu^
zkS#0|!y~mm#Ib-f&|D0IEh79M+6@lHTWP~e_MBw$M*fU<zA7V}f%NZC=@DUYS3Z%w
zGq!x9oe5H-I0{E^T;rlcx9AK|9)5>vtN*m~B}XvRC>leoYaimm5n)0gJD}H?{#IrP
zgmEE=I1;>ua0jlDlHK^W`U&6_4Ycwa5~qydOZSRa*Il^#kmQOj8!Lo6riQko_QN(g
z*<~M=u!KZ}2Nej=Scay5Yb<a~rP**=VO^7~ogIXivliV4LWkCi0zcG-K(e<DHt&2k
z7ML;6t-S6gnpK0m8JDCC^Ne8){kI<kjbrxTf(C%Iwqd3XZ0k@X7Uf5>TpA7tRwm8I
z>;o;H<41{P5D*DmEX|}^c}2BxuYYL2Mpx0mVR;x}I3<JbZ_x&r#ZU}eMnu(?(HW=J
z4{f*T++C3GKI;iC+Nw6K)S){+&sHPvs7BF<Oo5*#V7b+zzHH=17+f4!219K};FCKp
zPvp)qd6!s!G)i*p-uldX?u<7GvwGc5iSZpaeBZf!3JHt^Ljo~^KNK{xu#eYad%-f5
zDx`G-ILiWb5q)FE`mLNIS$j}NA-giI%BhD!LdUTrQQty6KheQu;QB(Jvi>+?%XYxc
z3>Sa?iiGIq$M$@V&kVIuYJraH)<9PWiHuwU)j{}$&kxC#(Z&K7Mo3`!4?(cw>o2^M
z0Yx9m%eH}Mn^26eYD7(b*iJ2h+g8s#9C==sK#@xc(bcr;ger0SO-Fm!O^hg4rt7<{
zy(v9aX{^FeqLBFhX0098+x==->B{-Xry54nMrZxoqNn%M%M=A)8-mJ-eOA9yYYf|F
zf=<GS*bMofkC$(Mnqcd|ae{BGL{W*5+IRSRJS=`B%QFo+FJFZNko0DQY72Qb*6aYF
z0DP3peM2x%+Nqb<@F`-l5eqsqzs$mKr;Ig@%g<HSa6h&ZU~02)<$9$XN8(BBwzEGj
z>5mQRyD>n5CXe*g(gX7etq>ss=F@L-MKXXeVc3=qv;aBL!6N5T-x^S=4l}KPFCu|v
zL-k>CtkprEpm$#hzv6hh)F-oxavBWWxX{t`AvBoNCaTTdyFY09o|`=Xxo)^4x<K6F
zDhxE41O$xiL$b37Zhrn{T_HKJQz|vI209V@776U98sX)i$#Ku7t&ZaKjrlv%OiY6P
z2xyz);OOhHZAyCs=JJ705F5Ui#2N=1T;&jqfmppx&Sx^1>S{k0Awfr16>tS0fhmCl
z{XmJ39e8#Cuq4)9TS{)OdG5$rtk?!=%pWDM%2P0@Vx;WePG{(!==4brqV#cVg_`X9
zc&wQ|Cs|7U(8NhDD=TxVed*Ow$<K^8njv4gA_zt?VBN}Ql7`xuYcm7;#PY0Bw6*?a
zEYHW+vM*Ve>HM^|lH{VzGodIJ>h+1nsz>SRuH`mx^j!&7rQLZD(bQ2_4sQmBi-Xfs
zBS(gtPdswP2ykpUdNKYcSUm>ABHGC*YZJ{nzzxqnD(_uUV9Pm^zFp1A@bNtPnkTpH
zzEi45jh6{RK>q!H*W#*Iz{rriA}7&A<dyOwmWgDIj7$y4TEk4u96zY9onq!B;#Ktv
z(&k82ebrnA#z@c0Ddm05t;JX{eYm6DQs82BQgqcFqwCy$Qfps)chU4iV`PAYm}}4X
zEPtXGE+s2p|G0`tTISrE%LQL++Rrlpx|sBYQlb^dUPaLp*E}_<Zmy6${V}(6FNSSA
zZz2|3w|$kKcOcssr9?pmA#%FU($B-~%dIH%hV+)J&3&&8q_9huI!nG?+oD5&d&zXK
z1Er6uzz7+*`~akpGm!4-Hf3PLTLyeaIx0k=)tH41){eAq{?B*aMd_;NhuMA_D$dGj
z{1iZELfsnAEd^?I^n)H(DgF&aVB7DYgg+%>8!`~)-9ESp`s%Vkh2h|sCXjBEIv89f
z8(DNB>X2e!Whr8-&NIHu@A3{cEDaEx|5BWtV71_wQYeD$r|jM`!-vE%X|j{xiM+|K
zl|f#_aIT|%DUV-C5dsMHYLbt7i77_I4UD84bylQC;VTlI<k3`ji&u(rz9ShleWWHY
zBW#%?A|BFt;z39uGKNjDxDu5lJ><Zu#U=TpBM@fg>x$ZT56}bq-u%EhYKOfLajn}I
zn{4bUc~j|PeL1=QG_q9Dgn*)($I?>q^`fCKTh1J1c>m2FMAJcOInY>M&s+-!n*CuG
zumsp1^iaK+9x%sB<W?bDDR32iOsUy6p<~|7kirUIgZdoh^JhU^CTY5y)K=CD6Lpz6
zu})_R!a;%fs%G$?v+a=^D}R81?%GKfVrtn*G$L=NW?>g{WX-8dFV~Z)Zms$HU^Mu8
zDq0xCbwOv!*S%|}s8BtfpnEYtH&feKnx^{PzM$LbJm=s!tL|X}uAJA3z|)hno3gUb
z<&ApIrHi(Uft_ni6PZST?2R`;QeKCdAai^w5&~OJE=s9JI4!l_$Bj7tC5G>OxQ=V4
z0~dpq7Mp!mSQpWLc){J!Z+Y86AbTeTQ*GR*7RloUqUe7tO;+AXwb!t}8gOY?xpP((
z&o0;L!IT%=tjyJM^%xZ!aTq@*xA#F?73cgFvA@@q<Hx@j@yo`Z_oIVgitqS6%`pl3
zrQMTEY|h%QCN@4oI&|^=Q+HxIZqrU@eID7#k-8)Y+cuA&qu4Be>ptutRjN(h{$r?)
zm~s)M&*ttX?iMb}1&y~nHp2a0mqYjJlu=$<x8qEH<l4%^-3LA;3sve@1C;SDhPKvi
z{LxcF5mgZx7a=7k$>xDY8+>h$r>7HkNXlDBhxWGU&YI{$oT$e%wFV&+2IRD~&(D;S
zLQQJV1h{{$jx~tgSIT|nkZb!f9Ov&NkdiJdvj950UrBJzdV0Q-O2YKJ?w%;6ZcbK5
z0(EB0+!nog2j4|l$5^U$-^CiE=OyG7NWHHaX+5TYAi5^j4n5eQBCI@<=8Oc2RKB6W
zL<ybY?39yR{vf%Gq2pS=fi}%lD!~5j{wj4{>FRuRa*et57QQqf9TgRNp9#Vz+)Rh0
zn>%Q0ju8N;M;X2RX+_zm&4bXMI0rEPMp0xs#Mz(~Z%()hdf9X;5=#eM{y*qe6W~oG
z6iA;ZRlfEZ?_|DPUFky;spZxh$a-fA?*EF_Hn*BEpFgV9KiTd&TvQp6?ILV%Dssn6
z359TVqxb4|aFQ>z@}2Kq5|QZK?}aDJda{0HzmaYaf@3rE&U|O`#nfO9dr{FUMK_Ea
z*j)&k@JY$rPV+SUub3KSl(+UkMQ@)E)c0jtN&1%VkGrhjR0y;3k>?9EKTUf&4m7^o
z9R;bO5ALJkWMcs<z{cb&_$TnkUM}1aH%~lhZTxC;i>~oqHi%;HAkL$`0>LfC4s*Cp
z3tP6NXBTw6y!3OnC+Nb&<Nz8Rr}*4U)q9z#bOaUg)D#Qt*=V?*+Lxw6VC5V^%pTg2
zD|@5DSSJpTW?p#ew}nNB2vd{Q;P>V4l@%r*Bol2S@XET`_w|RyQWzjkR&G`}n*DFm
zNYomx4>_h*Vkz>{f!&fGK{B+^#>2-4B0}u|R*`uZ43tI<0`#1!WCwZN{LNQ&<+#99
z*Q0)mO|lmDQ0PLfD}0i2$Snfa7?~g;zPjw^WY)eqElV2acJhNwk!P`IBSh6GWZrEo
zS@v7Igol~-q(W4MNrL@&X~P(7VcoQRNy5}d{~tD_qeuVLkbp6Z<Mtq=WNj98J6#RN
ztXytI+()MS+|AhTHI>*5FEM{eYqkIE*l3?L1}w(Em78CH^8x?<rRV`5+8$90Xbdok
zRf3NHR8+@;iCke}Nh^6DV_i?dLTN<%-tPgJ|H<zHgf?c3^ShTk$yaOr%?ti-UIj|n
ze+K-W8Yj3OLEnzET+X~}8K71FL<H<E(at(NEh4y^p_q&|j(vWh3<Bk6$K>51!HO=|
zMVOF~W`;oB!DN<h_eXwq^HDT*^tdPhc|St=EAsvm^MAJfnv@j5k#(eY+BA0La2U}x
zJ9UA#|Gjt(EhP)iV6zB-L5@};y1+9~g8fAov~kmC=a-|CP&P8_?qgJ!oV44H(rNP7
zrh|B{QeT7rth@mQ_+eDP>6g<OP-^csM+9aCez^8`uF=gFUfdg>0AS}?@9CK)$yw`i
zeZhrTCWw0&_#rh(-~BWpJ*{#InB9_27^+meWyT9a#|6E?EwOsewAV)=pNL(AjQ-35
zIE(6h<^6*mf&e&X&gG(NnPtURX>DYTH?kUg*q{xu-1Bk#(Vb%xb*G;ZZI$l6nnsOu
z@6E=SmeL4by6>X@{)T*ioU_AG5WERy{T>Pcp%V*vF<{JklZ|9NAkC82#a@w$vdEf>
zm<C4?jZf}OlHcB?7CFLKytw#Qm{Fp~*G*Y~Z>|;F>eKhXV}3oxQ$JYb6!|A%T)0m9
z=8=kf?yy>F#wC4}=Uh91uN7jyr8$>__qDr``Y6r5*i<-+3DZ``Wi=pkC5Q)b#H#yt
zN!kPul-SXxR5f(ydwwgX$!GSoH?X<S=|UY@YCJnS?Vu9)$$R{DwjcqKzR>gwv-4>9
zrL=TQHmc(IjmE@NU=gXImn(gi-$5&g?TX+y6k)B7b4r({Y;iS_Qj34^)MtsHJh=7p
zUu5%+=zO48NcdV!URQzNDwb29+AI-!oZUoXR@^&@N>jJv?cjS;EqBAFunN3<{kQq6
zYmD_p4{zaIFSSF?rbu_J=IPF249`Kb+A1zFh38bjnQu7LZ#1(rzHsKVzf;XlPJcR1
zvKTxZyp!6{f9F_~w;|03bl{{Epb!6Bu>XdRm3T$fPK%vHnk?yCX1PH3W?OhdJa)KK
zV|<(TMMv9FO$g>8^vgo(Q|bb$tMZ$6ZM6(swCxcZwI6@4dG1=|I+~yXB3V-E56L#z
zg5@Ve)Z*;xc8Uy|L)9g&a&o-W4ZaC0gB(+<0Cbmu@Dd0lHhH+XyT9bd@dj4b<U>KM
z{QP$s@vwhrQ^wr$?|zJq2(-KHd@ZnPbDS!Ajt12D_rKo%t=glwBA|`H$`XBOs`4D#
z9=u?yFRKFmk=@UZjcUE6mXZw#oPRoT%OQLI$}IK>E$6+I!<q#j6F9W<6Tf4lhHS<1
zc#E^#;oPR;SrB>0zUP+NhSKe!!)jf6oB*%=6x^On5ggnvL=F&<eWH17veDPGnhBR$
zML#$Kmk33zmWMX;dJ_PDYQO&NoA}(K<c@&@{-WmDqv)JOW3$n4V@21YJJH%P%CvTS
z3fDWgwT_&v3;~V>_k-QknhWp{zYo=NCSct1{MRGf(O)noJ1AbT)R&MXa)~bSI@&m8
zfY{2sC&VU9F?zd_@R*}2vb?e&F->(uT*P_)1EP;*ac4h+vxNVUft#k^J`58>;a~B;
zQ6zh@W+l&iBhlC%VF=jJB9f`3{VjU=3rxkeW0&SeWi&;rHJ1SS03mP6W>;qinj&d?
zl?EB(9yRLz3AdbLln}1<#5`Q|a$~okq1}7{BiZ5o1r^7rPf2vuq>z~f7Lo37+ZcWD
zYB0Aw<D!jqCJfJe2^iL{DqHzj;Rru9OkdGToUjRPJ6D$Q#UR({F}euAx4!FCct&#$
z^Uef|-0+hVG$|AL4rhz(`AM1n&tW`;Q0bRnEei-99rZS`U5w*v`7<?t0u4_x#%dGR
z4QG7rxkcwt%_w!!fRfUNmXq!v-dCGjfs>+!QxTI6_+F)n!Pr8D3>>=vMiKb5QKa3p
z(<>&u1KMA4wZAbdp|ib~!>`Fjq9xlg$lh3f$z^_U9xL|BT$_}2w`QD85&GE&hy*B6
z4v~b1@nKZ|Vj9~UQ!D3IlDn(?IbCPAB(%o#K=Z%)r}+*&i(e_uCDNAcfHA)+FmD?(
zo=0M$^d~TzkF+Htmv~-iHMmq%hGn(~im#5{d!QXq6=7-L`tUS{ifxU)%y>-z`->O>
zX`zSI(jPh;m3l$o?J##@k;$xipp`9>n7xbu)unDO0bjXMPs4zy>f6GXRhl@MPLB}S
zg<@md%5R#N%^miL_UzP|gYq}UgZB2t7O4%|m*|<(uR49mHuo+yHGMmVtZZ5jJr3i>
zFG#+mxHHpte!xh&iWCG-uu1cCtXSJiCfXta%GJ{c<*Fa%-hSP*0=)&Zt;OlS)36z|
za*g;?b@=w%=;+jzT$zW{_B2a@Tm=5F4FfRocaNAl14Lkk)Nr-GzcF-~aM)ffyGE=|
zJ>4*Vs0BD)){Zv|9)>pcXMBhrExWJ$O}5`9yA`7<f&p+0Fc>=J6T1-PtOQf>Ird;D
za{6wypb~A!e)EqnOLu{-N;F7iw^6Of<$Jpk7s9MW9sCPp%K(9(*ZV~vfOe5@L4O5c
z)jac;<<eg|A~#aoB8wHq!k2F@sjrcU5Aj^JC<)H)IxO2ycQmt;+r9U4(l6%EPwyVH
z3E?LG>aHm_N~vxZLx0>3HhQ@<AS5)<=0sb#$uVod^EoXDPP*N+5Ni1^s&pSi{Z_N%
zLuSM!2Zi}2^Rhu!w^110uBYF}1ood-Llv1@Pu5qKqw)(BpS$#33%KTKL*9E`ZT3_A
z9kv4t-MYDtC`NVm8S`~oDqIzs#z7K;;Ue5I)@P&7>I034!d24&f!Rg(5Pn<6Bzs!3
z|E!m4sX^`lL1=~LLKt8?dvl>nz-}%ww~cVM!sc*MrAYrsAiyZ+bh@6u5}YN2HBYEX
z8lvzu9FQy5o~w$mdacGi_qZlO*Ks}XD~l?)XTR*)ib)#gFJn^^<TLc{{LK=e083CE
zdSD46B|sVm#g!xHyXf{LM#Tb0Ks~nl_o0#ZSoQL!k)Iz)#K)S{eq8<Enx)?FW_eJ=
z@uv>Q=qDw8+H8eILYOq4Uwm>Fu2dLeq-i+Hk@s5Ve23@X_D9i}?KeLFZ6sasuL5%o
zzH0XTfE8wzlM00d{)xl!6nUXlsL`UA{m4ZfsT}yAt0}SMlUafqpW*#Hfk5idA4^}p
zw1MpG&9e^&59b5U+`8V~h5HFU`h`>`<?6VI8NWZI6Dtr^>F{R^OiR`$2`so=IOmX6
ztyAr<U3m6O)bnFEG|_Q8NX{!aC?sM5268H3wx?Ej0=MUUzu1Sb6`MN&o_pU{b7c6T
z#igKtI>&GocganG@j>uHMQ6()#c^N)&>vyJ)Y&6IrFRz8I_B89^YlSXbl3y!{rmo@
z|3e#sSWNyLzJ|WI!~c&g)D-0<7kMQ7B07V%Y%G@H{5tk=4zdGh*2_#Q0^Y2W!nyJV
zb8qtJ)gMi!xxD-X*XA5Bw(cuX$1f3WMPy)Xi_92YTS=MJ!JAi|xQ-AjIHK~z@x6=F
z^2PC%`b(wT)z8d4%WfK6%J<h=u4P;6m{A|kI^7j_XDSO#q>&Hk{0b|%cy8LQIEur8
z6doi+@eWX}ACh+m2RNUC(kwx~9R0&)h;3a2`xuBV{=OzinzB#g&F4c}gMwa-2J8rV
zX6MQPOZkX77?n*3su;9)nn2!lGD2`q#CLyFwL0aXy8N-N<KsEn+b#2fNPsKsFc3h+
zMd0RN)%yvb`EjeRR4=G=60eB+Ef79F+{6pbHFsTpOZIh;yoPQVoOyF2CyMdvyE4|5
z5&G7FAY;>4_O9vuqqO+FjSs<LRKb!i!rBL+TTb2-6ot!Cq=RGSwkO5HMJu8}!jbp5
zecap@)4{UyR-a$ToCAQ$K)UBogJeh;-18YDkS6ESKW66Lw2thP5>M9JS=nf6<ZM4D
zXOuSu{!HYL@vZWna2*&(_!XY@y;Lafmq>W}B~kCxU`AW&*~Y>`u8M@TPMBDWwk4>`
z@;O_JuBk!$B!wW4pQ>Wf5P$i{%=J)l6|nPTiOYmWJX<9oZzy%W8Wdmy+>v&C>;qA<
z`iJ>4{c;RV;-h4qLbg^@@U@Dm&zUK^A@)~939OIU@bar@a)58PsV(wkBXTsJBBW?S
zSTcU^NlY4=d|zXbz3uuSykN9+oItTW+N>ViN@1*6$P;+=J(K|=;^sTFi7I)0mmzSj
zy)5e;=St*io{{v)qPQ&#{JYMMwAWS4=^NRgkT_y*Y#G{|+ren!_q~>U{+5q3MGK|j
zo#c(SXee)-a<f6-3Kym$Qu4(CgKH4t;!SDF`t&<$5Vtl*Xy&dbt9nc|DHXgbx`YzX
ztyheAMN|dFXAL<cx_J%}n`2b}n9q8`+vry*c-7&XEo^NsdLg^z4~+*e;}Me~f3pj0
z5=QV5?aIr|9FJ+3>EqFGCA^5tV*7wN2dChm@o=L|Uy4D?^cPA|!TXum&?jO-?J5-l
z^Q$i|B%(Q%3sIyp!cqtZDFZo&(r|FME=(UB=-n5~Pi}Aw3(y)%tPwD3;DLoo1M{mj
z&X6{@^FTAFnr@M9FFQsz;^Rpyc%C}O0|_~>D;EyecCB(Wyx9c?rAU)NcV+qu)mMXR
z7XxagBYTBekNws$`(jmZ3Kieiw`kn9FCNbnNG6e1Tzx=rQRXOg{^>y;6UU!U6^tUr
z@}b0Zy-x-q>gU29#E*1l(+w_wBpV*BQJ|+lYWTtZ<NAmDx7RX`W_3bnCh#O<98DtJ
znR^sch>|#u9f|I0i6x=ZL@qIbWL*eSFdrRKR)mm=#4*2oW<MECVSJ*K7z@Y0g&#<U
zoc+;+K+8Am3L`?eVmgLBGr4)60?o@u6a>U8iJ+_Ro3;0{W(jP4AS9GVj{ygV000=%
zi{Lhwb2sXdkL?0^cZ?q#+<eaOewrGL?CEyge4EG%Ybv2U?GSYI!cCEXlUt|GzEaNq
z9xM#xZF^xK{6QZsAKtdCgOZ4-T=zTI!@Uajaud@->N~rOjph95`z_U|o0YyhuM`)l
zwI0&uhqx**{y*ZX5AGxPgHqsvNu2C$?$unhIuwE4pew-Pu_Ze-zrTMwA`pj=V_kiB
zt$TMGqG@xgM`?uq+a)C9_-`&D-_LKGUk=Y@?xz*UWzcnhoauHveN&zFSac=t)SUuP
zm))uno7esD{QD(xNRIv?POT<#ilmnDn#)h}HM$b%m-aKz3X{>-qxLCf4(?I?)e&oh
zXy52ugvS@s9+wKi366DUY#hD(WiOKbv%UCi+P|W`_4?SR;C`>Dh?=1N=TT0BU&SKP
zWkw4UCTZ77`^AQ0)biC~&WHGR609;wms!(_%%)9S=_PDdb{u5O8hK?5<H#fm%5F_O
z^8_$Uym(fK@xY7(v#P|bW{&vNXp_v1ptLr+ksN*x_U?BkdI^P;5rAMqr_y=GXlI<g
zG`FCn^4n+h_5^&!m?Sp2-pIyAjleSF4;sUZc^``3Cu|S0GvR4tmPak$-X7-roW~7N
z6JH)zUD<fw`h4BnYcIOL!<FOXKUx+8H+x4~Za()#44us?qc&O(n@6aHzl~1KYV1r#
zOn9|zGRSqGoDJfbmQA-t*$6u9PqC{1&J^kA9*C$xn~sGxt^j{z>xQU>*-vYc*M~hj
zU^E@nzFjwIcg}unUoS4DHX<THXbVn=%&h*}^-G)>%6|$^KN7~GvX<g-veM<h4sEHp
z<NxkSw)oSZ_CSDYuOrFwn@!N2Qe~hoe=-UeO<$a@CD|7JY;W9g_?jA6Q;!?BSSQ-k
zMY7zqa>+M3=VilS44&Qe%&ouoXgJ}}v%3!4<V2PyYo&01)EmH-;HPF}PI_`!>yV~@
zh|kXw$+2Ifn&W)(SvGJf(EHI)g%X6kNtoct_JOT>TCts(wWYH&$9cm}exR8;xP7zz
zex&SP^JYtcontr70=8U>M^MeJVMIc?@jY6Ky0jH?iNFhzelH7g4^y%dP?F*&XGADx
zG)2|)faPPVas&Wc3DUQrigeeWzU7sUgq`kwJ2(%q&3h9mY5FPw&AQf&IW#7%^u&m6
zwim29Y5-k@Qq*`@D->5o7&y~KsJfX>35>1`nqGAIU$13^d1ojLh)?_2RO`!pJ^JZ)
z7TzTC-M;)SUk>8;c47<T8pT0-o2+}Mv+$wVSNr-FeDbW=b3L!z6`nx=*JXeh7h^oO
zx9V*l?DU)WY=;(hlyC%iE;gUl*PaPW@a`D71PW=cGm}f`jy$%`WvI6)36zR?&RAYL
zb~|!Ud!_L;LLj;H{tms}d$-9nKNY|Ia^Q;D`%3ueF5Ji#E>D+QGxi$noGpuCH`5(I
zf2+!V;d=$9X<|hQLE38z$B{TLGg;(=ZP(Dre#9mOSE?k35@=TRN$u70I}3eLI#i-^
zVQN#&e%91CkXOU5?{s_ZCTH6|jp8Qe3&1hPzcptjVv>j@_B`A8;@h^<N8;J-b=cR}
z7h<^RB6erGIeM6%;cfplCJx=-PO(L3)=qIcyX!kgv4##4roK#I1FKRQn58g0RkGK^
z%M-hoC_0FE9pG}$kht<w^j!uPcQQeFI57w(JOa|FDP1A~Fe{%z;rSligYb2XAygKi
z=PSb(S<(i63_Wh7O)N8>Rc;lVO+uruO0X&I4uD_lz1``%7xt_2UhKK&3*>Gy1=nOm
zLJ4=Jc`}k6*}RX&sjY91T)T#C?by~AEltg085B27;mI!L-xgBRfw;CeRkP~Cs!zbO
zgC1SR1V!uah|iKpbz{bDM#B?G45`*%<K|Pj*4Cl|e7N<U1#6>;_c%uzF>;?%`<=>m
zRL!zyj;lxX0PoM63~-R{NAIhUfr2UOd)0CHFL2X-fm~h>5vDH1hyVv-Cf233s&R^W
z!qgLTs!uJhiTO0422-zYy9aw{jPHDpc~Trml6=O2Fu2CxEa!!M7!kDFxa#J^M<Sx~
z9T|f5)jM)}ytugC9%7@H7xza4UsiF)1vWY~I-FbKrcxEwVqSyUUVY&k3n*$DF58j`
zlM3IjpXIpZC7`J@>Cv#7z>iwMhz=!Mp>boNO%YMvK^q9;49@Y6tiBXok5GFJF2$u>
z<`A3Pjy<^TNaraVS>!(ID^+2sJ^+81?Geat(QP>+SYB^(SmuCY-*6EzImy2ME2S>z
zrKDoY*?r?0rUCBh(WBvw)pXSVVpH&Q<$XNwC;35>NJ$Sn^o@P&F~m+cC>|<zvASwS
z^MG3~`S>@DqgnM`G(h6{^jFCNuy_&d%-DtC$n27H_HG~=6NSq<z1L!+rF9b`6>WpQ
zyM%5~ag=h6DYu*X{oxV`n-6JNJ#}9W>_CecK=(s?$OHXXQfJSvG1N@-Tkenw6iGB;
z7HV+AG6DoW*<})IhOKkS)uG#wuG5u~<W2JQIAleOJ!r%4-VMNq^3i94dOs)Iwygjv
zjyL0OHfi~rl$h{Z`+aklR}E-d$GLIp8yLDmS1)wXxl7b9-hJf|QDBpiFD+r7{%HKd
z5y4KK#)rxjpiQvsod_wuBF{!<-tna8e63^(wn#Vm<@%E5&&4RtykTuQwlEA{g<`ev
z&`pEI%fac}POQqa*jJ_S)uc@H#-j_Zz$B&rZIX7_$nBC1r!JisWq`Cu-Q08e-Dd-p
zB`Nlf)LU0z3O4w#^oybdj5^Ztg!H1k7N5?5-YGym`b!{5!=IkK5%R2U@NqWUt(?_k
z3d-77h4yqEVzC>grfPu1=E`YPVlsmtSX9Ns{oaUbIwR}L$P>eM?zwu+rPDq-7WT%{
z*rL3lAdKnRsAjny&o)&Z-ZS5O;V;uC4fT(FlIE9OmoJv63l0U6wVMev)6)UE!ioIL
z=7DvIQ>=c^18lMIhb_j9_%n%Iz{6_B>p@yUBDUb_F1kNQ|55Srx;HDUY$QYDVt2Hn
z*Tbx|mO<~2IF^)JJK{nT%Ov(55Uj!`dH_y;v4~qW7=b8>)tj}dZHJt-ttA4FnzF+x
z)gaz+mm2yoy}ZWkh+hFKsj0sNtiS|GczgDC>=0#~=0FE^>w8tAhl}og0()E&IQosv
z-G8P(Y%X7NmBSi8+Uq8|mB|D(SHeR_VOdX<aPv)?>$91G5fyKKd$X4}{_&m+UpyNB
zcm~B<B*@8U=*u6f&kZ!gr`l)bDAKoIxmy14erQuFqn$EzXRm>O*R8`!?l20lQz(I1
z`fo-Id*e&Uj;JPkw-Q*I@o+~ZZb`A9NGYwqW2lRL#)_R`nE|gfjVux+%Xh)GR(e9w
zux^70NEa9VCq~Fx`cmD$9?75j%~ok}((N$RS55Sp2cZ#Yk2t!Ylv-$5vX7lKw2i6R
zFMGm=jifPOLcMjrZY(m>_Y>rcYF`LqQYi?K)^>Jy6nEwUpQ8Tb>RiBCf#{8{Puxez
zW#VS{n}CK4z_#@@`NFLIul$OCSjvM2bf2Gp4L7e9s=Ik6lsZwuha)3Nh|X=UW*XI-
zv}KNqe{s$kE~kavdm11oyZ_yX&2@E_pn*UAZh9RlLrjPJy7sFc0^nD?{Vl<40uoG#
z@k10UdvaA01A#Y{D>46<wnKUTs&JCu?K|VT+KlKJ4_TiCh5O>U`7n=O8C~E<(?8b4
z1{x=p&+SsUE09kr*>dr8y{!Nna3A%GPyYa45@P^kfrU$9%?L45b<{WiyVYD?Vw(F#
zhGHHc3`v!A<e612ADz*XLD=(7Y^uTl@}B>52#V>yME2hysA4<H=Cp<Lw{Oe>0o4av
z>-88~tkcCnLZF3Uy2*mS$QjI;zM7qKIj;gclf=OYFy-I>eI6>NpnxCd4+l~W=!6~(
zA0M`9F1{8J$PP;@Q|ObiJFsSZB>7MBPYu02wlrmP*cij$8-?>>&l;63xzfrod@U|0
z<T?IGUzG8|Qd)X?cw~UC>y|{F9w3u582+}oQ579NYD}X7<$E5%?gdnh{7@obV>%m%
zG)c}Y?R~iF781DsxtRy-le4et1+px!!6<!?Afkc{n_3eAMO?CFdqAa4`0m$e1Wn$9
z>$kV&GyD%_8`r-N6ZIQlzMf31a(iuv|F}8n^AQ^n@Ud(>TbSfDJv2pPC?V3g2F}r>
zssA!r(Hm1g#!@Z-%g6x!UpvdNxGG$5{~*b~(oCtX1w4&#566K$RpCRuW_^E|25@B@
zT`e10FifRn2><!Ten+DKIp*Vk$uWQJg1_>rK$Sg6-)XBE;c?qMy`^<Sa7(MWuDR$J
zc<!Ndf3<;Iz@$CVrjAm8IEIi?^&VEQ(+gQqyQ&<x{V`Y|v38Yo(mzhH+dlVGym%|>
zqWw-|&D{Ci`x}Xi?wytUt&jKFXH4t;<pQ?`Hd(Ykva3R!e<`gG*;S{D72lnLsR97c
zF3i|opBdJ;92m7hhnMbe!0If!03*5IkjrR16U9_{GxcZi)Zm|)A>y<@v-R+e_426?
zAN%6O4+P!&GlCZs1fhEHI+7J)en-mVftp=87uL(^8PTus&#eVp5~6{_kh4FD?mjmT
z(#3@nI-it)fb+#eX7$56e8h1LQt`P!y_9*Pd=P9BvMQ`cBoWW^DJ6_9X!$bhAtk_|
zZ}#T_17IIGzO#sFe89$wNj^Ncx^gmnvp;&HNZPPHzaM<QU*L0vK}Ni2W6p5Dneril
z=2q*T?bh}T@p~XZ*@Fy^%Y%!m#Pw1mitXQ5YpzrMgPj^vNJ#g|x;IG<F5!!WzGH)9
z6xiQab|XHUEde0{K137SgPl+3@|&G%^zl*nXoNS?Bby{ulVs3XnpzUEN=W{dpkixs
z2$)vcx6-Wnjh(Gr{ajqWx4tqJ_~7|1hMA<~eslVsx~=^n_3F^)XkzX8y%$@~-qz>*
zjK^$mzhGGLy&%_JV9KQ24q8C#Md)L5p~q<cU_UI74+(_U#TBH!xm<i=2<4b}#uXFf
z6M`$ldLS6Iey)wYaOFX3@pTGoq5DRbWvr)tza8|Xn!h^e8@gS8b<iC?lJL4lx4K<S
zE*GT9<r4U?^xpOx-Co4X+<YI+K$&70MN)i$)a!~BknTVM_|?W2XKEVD3EqEZ0l-q$
zk2lXqE&PgcDVEb>_dumCr0a$2k0jfK!y7I3?*xI+D2izjF%HJKLSB_#JM6Sy@C2f{
zmO)0GHGq0(m&~@fyO-EK^|WI9ZqCgotcf4Owxe=-mkB6V6g7`>oK4RR*_mW?OifJ<
zeNwH>lM~t3EDo;k7p^*beeM(ubNHKZR)7PD@@JHnjeii<aI9HDMa@qqOOZgV?tEBS
z1<X|N{(JZ>%-p2pzSBzrzYq)9(Qvm>$-_5eER1|um~n=kXDso@96*(g&iM+G1za|(
zr7!~{>OI~MlFmn85-Vv!9tbN>%6;Z<!g^`(KND6<DZQmAifOPp?^rmNg|UQshq4Ww
zj6$({pRoktd1YOCVKR!90U;kPGMi9$3BVD9%!<CMu>c%VP}m%fvD`=xZtVb_%R<YK
zCWJa|f7lihc4JJjvj^&mh+i1UayGd&WGuJzKJ>Yh8=H=`zUcxUnaz(g#SuO*urZ->
zQ$7F_3%xL<@F$=#NzL5-lt=-zo|zDCE+D_CXnmnONh)#+ieJDBz%NoeG?p0b@hnky
zk$YF+jI0Q-=s?wh2jL?YcwzvHet2TsS|PH-^^t}g0i!Onzi4ag*V4%LO14`N$7!0f
z!T%d?ox>h>HlXO9$19!!;VJdW<b7-7Zx8*3@mUD-IOT;zh(Dnsw6bOWz_HLYUhOr`
z@WBlWqa~86@gv}bz?0K|5fgr!+Qe%eYHXqIjD5HhvUSB#`AlmT=w{+9Itv-rH*yqs
zd($#E%Z7#H9vtB7otZCL!?h<N5@Kt(0$$&yNTP7}(Qw6tmY~jwR0A)Ky+uj056*@u
z$dMLr=mN<?_HV-DxHjpa(XoQ<WChEaAbOs5>1GeW*I%Ekwgx2Gv4?<h%B9TwgGx7j
zd8L^pFzkI2@$I`H)yc|<&oH64-PH~?Rfa~!jjqgd7ECQv{LND{w`BKvhV<PcS6if%
zN=fBW=`?y<)_Js4{BxdZ7Az5YwxL55>cp<V{uA4fBjTtus#y~^tfC?jLs-)jTXj-7
zsbqwI=B)2Y{&&twGMtJx+xwN^=J*pr5hN(BSX787HzYdFgizbw%Z+!xIIEF-xA?64
zl4D--FXR@$4Y_rVC@G!2T#vZ5F79+k(x$eo(_9Tz=<#MTjb7G-_(;2SIPY<`4>)XX
zw#I*n!ARl;X6nzBSYGx)&(GI1u^F-Mlcohf^(14d3|7}-v<w^GBkCTc5QM2u%s%VL
zdSJKY5~wy~5X6+$dhp8pCkG2PlLW9U>$l}T2PX!_y~pA4D)enQ)6bHigD#kWbv61|
ziX~l@9vb8_mmJa6s$|rbPj-OcH(5Z$4Ur5@mPPrNQUNlGkV<}iV%MEYTRjMRW6>QF
zLxk3*-6?HroQm61S6x#Z|8UW5{J(@0#th0HP9_URJEl>8)D~sjV#PjVXQ|h2?%Hf}
z>6pJq{r1(OkBKc$y{SSiC}Ivvfesf~z;*J(CRd9~CmpQpk91z^(IR7oS^bVT{{Z66
z9m~T3$>GFyFO`QEXzV2bSgl&+X?nV>+Z*km^kmu-`gue@(q-m2jEw@rVM26bSB1O6
zuThfvx{(V(OReYO6yBft<n4J)>KwkTri5ECaN>PqSLmSUserE4#p&&JOZ&+Odcf8q
z9}#nT`y^=?QT7BLMwPg?oA4op;fnWe`D@?1>xR?dTUNI;e`wdMxf3<%kB(LGt)wSP
zJ=)z&LHuY=fhZ67j9h^Pw<nL|ogq7HXI_5Q${*h>g0SP|0S9-myhM9I9rtwEghyxy
z%&Ky<q|=1jIxX=OzQ|o(=Kd~_2)#UfC41wVo;>)a<7=zed-bButbpD6{v9H&wNiX~
z+kZf38QJx?wRZRYO$yDDJuN#I*W(#hF#sA${y;+`Ved9RkLVCMleoL=-TC$dM0B^~
zZ$alH9ejF5RJ?~nLVeE?D_}1urJZ`45do_i=^-l(!~-JTF@S)=GE9X^SRz;pIAFkV
zwXprO>7sARa?fLb^2NyM;{G6+8f`h`w}K!0*K<B!`YEa2tE9g$o6fuvRTz^kXiLeb
zgE)7$R?V4rZ+QJB2<<P+qH#^)wxK;eNpKiq!0qkTn9~9D1l7vF<?<$tv_2&J5Djs9
z5GmtB#fe>5O8DL=#J(&N2lumETy*s{ulbZzVE~7Nex?fC8{K!V5y3^3809S8Cqa|=
zy`-?lr)|6@PQ2&Smq>g~wg=JKgphi^C<3)>XD+lIs9k`CZ~SakyR|GuzZjICc=4H>
zNa3GpZg~4;SjV%fr^XHDbtV)yPPHew*qdz-TL}f=MAFp#_DYc0GVyIU|BHFvGvE{k
z#@#K%otxdxwDsO!buOTfugwAoiL(NyAO6O#5f>dJc>F|_q`nz<I5?0d#`F*){SlY8
zx%o#_K}7X!{8iK@+~p+f=m7>D^hx(HO3PN|h<yJ`<x;OWx!3e{rJs}juXwNJ|HykC
zP}#|#ZxdNXy?aNgpyp8l<6q7Cgrlh?oWpoK#H8OhR4EC)?$A-!7aKY#e)wEKspz9F
z4>h>@L`{8uE`Z<qV?RQZMDE_=*nfnaIn7JMtnv{wSvUIc=X8EYtmkL~biTheSq<jH
zFt1}ycc8P(0lwB@K|SW)DTNi=J|3LD`thi6M3hTqyI6-?3kbvvfvjonuG&CHo8PNc
zM-G4fxY9N4ar_n*Q@(0?%tx3BYO2n-CSfy-4A>C^P9QCI^b$T|Dt+`%$%oC|-?{4_
zbNMe>Z5Y4r4^3)gL8CQ0E3KWl@E$?{cK<etjRq)_$bSEQGv@rxLKt&u68?h|(@TYF
zqaL%Z5r6Tz#+8y#9f{8U3Cp9sC6cOb8oT<Pp^X(X6QlFxu;acNIM3A}Q}=Q8{C~<4
zTK+vt_|HVSRy$#VW8?hLc}BHHnsTT?G`vp$-DAeSRuol3TT`dDcj|>rQL4mymy_eA
z&hWB<pdLfxcD6?Qlz$u9p7{GmqJDdmQq`AnIMVqC`DMzpD*o5Ui@IU{RobH3<ja|_
zAy>oyY45G0;>i0w;Rq5Wkl>95O$hD|!5xCT2X}V}ArM@HTL>1cacDfaySux)!xl-N
zyEBt{?=ySPefR91vvcy#^r5?|x~uB<{b(fvmu}a_iL1-l?uN#cTh6W3a!k@5XMSSN
zojY~SsO{k=;a(T*T5@~N_dmYuJOsJOTD9JL-F*<++RNaZ^QP6``vXmsGOU1{R4@H+
z%*8Kj6`&vgDd_WOG&>$&^f@8g+Y%_<0`i*57*GMmh~)-owoqehsI>i?7p+zR1JLr|
zKcwG37|#A0{QMIg3*2s35Pw;u|JExGfYyH-dnp<$e*eQxPN{Q&{i~$uCs6#I*Bt73
zFvs$}-ewo`lW#e<r^!q%e?6cWm;k@&{czGG7YvQ|E2CPq4v<>aY039FPnNKmbQ&6*
zzwnuK<5|*epx<tx_v7FFTd6rqD(VSo&^vie?}!U(wyIo~xMU>S3XmO-S%91GFP9p;
zG!us8FVn=<S76=wM+D{XNt;bvu*9%h7v$5jbl`JG(uQr>(<M}&ouHa@=hqI+EhF@1
zR9(Y1yWnxYDI@Kf_D<U5Cek^}e{#<j8qc!y*!b`>xB)B`t7Bj9^DPZI)`AZ7<%-VL
zfq|N8c!pZUG=_ntAA#nl=7%hD8&5R(cUxq5lp0)o#if{rA)mQo06uwW$T4%U!W%Q$
zoUN%0VM{vry#AJqEvk;o?&KW7lINjmDkh@HZ?u#e6LnFaC<|0X1loUyo7AFu9<F5^
zL}Gwtq*k&$Ti|tRcsD$vq6@{ea<KToKed__dH;4?CjVcWW#d;8UzW)3U8C8`KSRKs
zp?^n4h67gZzmSo+Em4SO+S9d!hQZ>`NDtB@XFy2iI=Ea$97-Bo#E|gL-Kl`6S$LMw
z#fZ=Dmm!IFQLi!$n~#NAz5>}y;7-q@VPecCha8{UltJzG8yPmdAZBM%ZjW0+{!w0+
z^LxP}#Dx2m<s22#<|TvKY?rdAN&~>64Fug-T3(wiHn}n|oEbECWc$}q)LUYXu0cP(
zWPr!u{Eg(x#~<?>QxvlKgxW4jwyGu=wsn07w|NoWCRqfWU*JV9t^@Ts7{w11Q$Ic(
z$T*l?E3fX(Y;1I_E1&x~Cd(rLbtMf%usyH*S`q_+)u=7oz7`<dilG653dnK(J>7@`
z^Yicb<@KzABSX4KG+j<<WnvaC2@9)u7qb0jLc?~ZQu&t3AP>;*rsI+U{q7(6#uoof
zKt2Wg7e*S)JpN{_86V)HC^P+hxc^anIE}+o@U;T#E50{>?+2r~(<Wb>h=>976=$eF
zMn{rh#{7Vlx~bB^Q3od<5jklHk#5yed%K#rI7gZ2E;g=zzVX@&E*=QR@Ui?^sT!R}
zNlR93ok)6<6OJ}r7L+TE&zw^#sgKeinP>j#1v0|xr|d^I*$qs48jvr<4<L4Lf~4<*
zidboC!__d;m;ixGFuH*9$sJcxRsm5#YQn)c()I?lu<w9VO`3jerHt0vI6*WI4HX^n
zrsx^e&%{j?4KFIE>ZA|Mm~vP_vFoSh0D;%5b730N<i*}(6h;AG&3XpZo1!vO(fSf$
zl&l$?gb9J7JfeGB^dE(-V2Y8R$WA_2o7qSWPL)pvI06p+c5<|I7<Tf-Vx|hlh2&x4
zOyEbBO{7!+epo5`9(UhN2t7`<YW{dP(AspQ&%{dOxDY+jQUEOO=dX<_X(^n$y>j(E
z67VCd<|{($Cel2m){_GoNPrs=+8<5iKckgJ)GSI}v`mz8J(Cp|X{dgDnhYeOfHpg@
zb)5mh{D_OJs-E$jD6l=_A7Bj)fli9UGyrb=T6F&%+^}Q#e)d=4Oh+mOo9XA$3!K6a
z|3xlSR{<08?zgr6YPGYc?^rgvy6$t#e}ox`rw$|5Fux02ODL#(6n#Kv+OsVoMc*G2
z|F1d#=j^vge-9_V|EF*Quo?f8W&mo+pBYc4i)`kj(de{$z3nx&dYph0CLmV+D^_U^
z{!6S9M>;{eSx8ovY%`l0y~pSuMT2<Aeno@+C>6~E%~?#l?PxYHG*3BLV0A&F+ty+F
zmG=0NgodO`*w0!tF*SH)HhJQa)w?p(#o~)B-=`$kF+|nj)$E&ffxsHdo+izACoCP=
z$2+OHZCXxOaLNc=m>BVQyg#LU&XH9>Mq~jFD+l{NlmdjZ(fwbp8MG1id^;5=5ibt-
z@ANa<Qc4cw;%Qczi)~xK1ri#GaymGIdw$Ek0oPh$@jx>KPC&hfrBIDC06330xNC05
zKy*pxy6L9pac;=4K(oM`hrgNyvL^r4Vo?5snr>jLT*l2x-6A3WUqY96IXQH3vGiQ{
zQ9Y3V3W~j57qSzQm;;iUS>jYwhMT+$0RosGVB@_`68|IZeLn<_Ku~6+EL<&8=jedN
z92%|Xg^FWJP4CiwlbiXowOnE~2esI@WZry-Ke(CCvatXA+{_{jqvuCAAm!eC!yjb<
zf4|81KKyOD`@c!k`XxgdH6zvB&*=DouC?_m3<HGl{eJbO_^#6<{JL2`k)zc7FT6S`
z_(y{!X~BzvBm2J?EUQ?=7vuqrPXD*D1PWN?PiWRcvsi}R%MQZj`v~^jZ^~Ln0BdA3
z&u-oI)bZur^=3H1^mx7Wk%e(0)(?8d_x)e!nXTQ1$3xQj(kx7Fl|>7uTWtn=(;Y4e
zNws$uhQ->$lB1q?T`AnwTfz8wAM<p6`4yV~>Q{i(nohi!soS`7F#(*K2z9qxEPNu&
zXbn$lK5K?TE#+oDt@t3kH&e&OLuk#oJ@dhVf>1Ofr}%*5I@~%+fegg9y{g0@&n6l@
zWR^7K${8&?D@*~43Usi3RncXAt}KkQ5>mMOwR6GI18BSY0uooo{_`G?fMK=VodH*=
z6kzd`lV)~1SQ55`)~e<><N*w%Z$31*?{#J2*Pj6P#q4dRT6U^~;=bFvDUP`>WV_d)
zYY&l6|MhPigkQ+O`GTR8l*@uMMpm0bShS3#H7zS9#dd9ZeOPpa{o#SMNkQzS7=ud-
zuI7$Uh$gBUU1=LA=TorMhK~b8S8Y;TFc==&YZ(Kvuzo`F{R}CkN0$w3f;8FV>N&T=
zi`}!68<IV@I|~fny;CTkY9gV_n_emx&bD=_St6G0Y{G*_ktGah{<gR{NXg{`6OdHy
z?-Y*;K-Z^DWCiWE1v6pd5nn*Mog2=GDJJPIu<AU|l{mG#UQ#7BR;$#LstH?2Pltw{
zn3MWY^$<6t8L+3MACl9jlrzFAb@!>HqtJA)>f1@qVI?|kos1*&1<P~Uuf;_S#N0ij
zqQR$fUOc_D=lA5sj|sLVIYMR=x3)<rO?rljQgL}swa!6BBE@k<s?_D6+%7RWkkXG>
zTNzE5Gt(F>=~OarQR|T#*P5C{=oKCElDvr#kC#(5?%`OPwV0~E^@XXZd#zP<U>CDH
zo%t)2@{*OZkHwm+;|>UM=5}!PA0{+mlLS?&2c?*rA*-y?($b$SC2?ITOmYIEx!8nw
zm}L*0An}!Jx2ByKFn6tfQ+|=7ANl&&78`pc7I(KAX3!KD;zzvG;pS?-$42}0_kJ<k
zvCw1Qkth{ct|#rAU$=U1;AfS&rL##Zse~OE4cMClM9f~awj?t>&*-IXJ7L0J7Yfby
zvqrK}uWt(LI!m}R71c$!LsfPXn2u7G)o$J`^`LAeCEr?JezzZ}Y;P>lA9Qt$I>vk#
zU+W;LOUw<j8jN#sEnpR8<u+|B>8@j4n-iL6N_M6>n+T-k*L#%>nt48_IA|4CUkSIx
z6pjw=e~-$jhhKUuu1qU2gxr$$8Ow?bK0^fu4(vA-Yknxu2o|f|q2DUVkjNrlF!0Sg
zD8${G@C)a7RAPB--q0~Aur^{Ae~L^3F$e^*@Jvog_T#9*K}H=*#2JD*OI3nXppNA4
zDm;n(z3nA1i|U2RjZiljwXy3FC}j*)Y~@{ID{7C;nrpon%Y|6`5#b8#xJ2<XH^lS?
zY#xIJpZh;ehcx!W4=&m#9CmX@>q#ekQw^;aI;US;-a%9%1s7eF2NY4-QMN>t##ea+
zmyHQ_CF7DiFOtd8mmjq#+i$Gfu2iHHs5)^HZ|8jPpyD2Q#@uGz=p;nN)862q(jno!
zl(Wotet@XH9<<1j{Ys6ct;PG4uXAwoolQ?29TFpdBZaxNtgrFppmOQ_ds7(!-NzvC
zTjG2z@kpHMYsZN%k>p=C)X4_&dp&hTTN#e89Tnrb-a%MH66#zHujiAECL+_0mslve
ze5Fbyj;~Fx+qa4&`$Dh`-|_3r4kJ0*%nrZ#UZxPb7SGke8&qhB;J=oTi0AJt7tURf
z)X}iv-8k&6(lZa~*NN{Y&E9FWVlQPFq<kTV>F5!gS?(?G^<bV49*PZ&pTPY%ErG8l
z(y3qbOVcR(N7TAL^;i}`nF}<9@5*Xd2Cd9t^y)LM^y)R*xa`lUu0S+cE2f`rZ9Whj
zCkIL2_Vak8)0yO8(?!pC)R~;uMv}ueO=wn_r+E?4QW6(&H*H{sy0kUjHeqDGZ>9s(
zq}C14DpAjTBQ=?+P(3e$qN7{3-AOU6LPxH_%meE59B6g%2Buq8{=7uYARLK~!mHYj
zfq{<|8%}PQmt_Z;{6&%q#9LiW^*Q(-?B`|s3p1rV`9~Pe<%t(QFTgvF2m6qOEIE}Y
zPb|QZT6A>#a!!E)<|8|ZBGUr*{6+h4w7RSaQ>FOzK;UuT0GZ38bnT!ZKQD>4CAm3I
zmRY-3mz=g0ZawoJ<CbDUpC0#w&A&Ypz{uQWA4_I|13#>ruwcVkghhv)&Qwn{Ud6d<
zaN*K?hkp5xsKJ|&x|<Zbr(cuyn<kN{o%M86eHr(+IX1V=8_{fSlB3|Zm;7^Ws<R-E
z9&Ve351W`Jhj^V<{xn#O{1W^tdeFPC*V5q}QPdaSw{esRBDcAiSOug|b^5YTCnv{&
zC~lA*_d<M|cY^#pKY3BwL&@z}=EI5D`h?BLi~FVx0|Jf-HA4DpK)_LcABWumKRseq
zz(jyBuiF%ZClN9`NlpNH$}}KQlxA6O`MMVd$GCVLBX$5S|6xS5;X`#<V%7uz!5N5f
z4thcZ&dpGP{VWi<)txa?*$o&CGT8%_IixS!+GIocB#C*;7Gy|6L*idYt`!*my7cl>
z=~H8{9lO&5D~n0Ve|Z7CR!C?uO$Lek5y_Di`efbUayGT}dNCT&D>q9pP<_1{J;dDN
znr>a<Uza)=lzZu^1!IMvs$j`%u8Si^nIcN&EK+`?h@wijevd%@1P#EmFTw>F6z4-$
zUvcHNK*n>qV;n86M)qZ}DWvk4pSYH9WSe9;HGdcjQZKQjT~`)gcSPWj%6dpg2WEV=
zD;xL(687Z&%XvfOZ;<yLuSdp3Iner``d}VLEvxXY$t!J|7EztFjkl<hVEgX8P%)dh
z1Qw=wdIP!D5@WhAgsoBd4~x@fA5oLI!2?A|VIp5a6js8vpYca!)y=d5eL65Pbf2Ns
zwvZzl?I(C|(ic_=lj-lE^d&4dY94KKgC<+aS~7NubKbIpz(;eXiSU!PTKMcoQA_gu
zvzLPW0fvrES1D6GSG#H18=~l7ZH3CP1r#kH!yW6}*#<wY&?|}xbZp5oJpyOUPtNt$
zY<MmtQFnz)vk(}0W`ct;c+ScgEVa(-iIWnv*h)^7*SfZwUdd`=(<<J_I`s3jpm)1s
zw{q;CLc(%@);Q=fO)VAi?=H(jJy(;UBGuutfv0uFs9s66-k985cV8B^8QR$Mi~N^+
ziA31ICVD>i^>wZB3iEYAm0;De_XydW!>uha2|afBb0CrB0fU~II|X0Uf`Yt@!9dsE
zs7|P~+i5vVPm?;ABZlbU)0y>2PcRVU#Q-v}^e;@3I~(OC$5erai7*od4fSh}4X7PR
zw={vDo$3j>rQjz!cZzQ1`;Xf$tLJh2;y!bGYELxr%KBPwDT;vm-NbtEf~6{t<@m}C
zw2^NffDYfi8Rm@g$&Lq?0qs77BRNn?iE~vlDdG{8ByoMpwbx|p;gLc$<D26a>Ecf>
zIRhqkP%X!(x&n7jDyIxr(^?*na!%VR8WdMalR{W6fWugr9>UWixPzQYU%7ScW-IYH
zmoLq=r4_aJdpFo>)z*$Fm3LJGejvqz-w)_RZWG_K%5@{Pnps6t=>RC`XsDwQv2&2R
zn-i&8aw5GlNx#;B2%)V-`hzu+h>bzDxl3a#L@RRHXT*)Dfz+tA&VdZWl@_CRt?F1-
z0|(faM`oL~CnG0D?U?h)axQ_=NZA0O8g1d;_tuTbzSGHx*9r#0u_EXc4I_F(h8t8p
zw{`o|VD&~>;ajAUMsIt|@7XgYrW=^u0Xz1fqgQfU-8=CZjOzL{B35wmMQiFC^Zg3Z
z!J3!5!^yVN)zr0iw&zpho<en1yPiwenF62gtLsc)0Dvh_hNcpmv#OK^5Wli5omM^^
zNu$Axg7X5Y17ydN2jX}LJMIip?rK)L37)HzCwF0Ey6qX}JMc-$*>jLXIep0KW#{JP
zdT*U0K}bo(?g>FPWM>QK4%aX(b&9J;gQhnG7<9S|RPLoXdGT<cY!eTr%_mlb3fqq$
z<6&c8+dyMq`dOklrh5G#xpX<)-}+h5P3UBx4IJ;(#i2J`fc8H`B|e4_&1<eMp|**2
z7M90LX2+woxAmIYxuz5kdw|PI32<%x<+2K}_ED8MR&+d6^u0l14|E`7e4IvU=p)qh
zK9EpEyzPU__fVZ>ER8|0r|DZN3$429iTx}JV?qSyki)jr0)miG9$Z7=R^MKai&0Ij
zi#+Y`PC}jR&^;plf=L3v_q8Nn4HVAF&WKqFKI+AO92E(yuWG92ptpQ;pIeHWj9#U;
zn+0T_QEWG%k4uR$>o-{OFPT+f4BF?v)q?|rdW`m?lJ+>Br^NcT2tB$<2F^xBN+7aJ
zzJ<a_Op0NTqOX$^BDVwdn26X{zoORJ61q*6-=Al_&|LSg%f2j+$n2f-aNX(zt{u0f
z8d0+Qw-Uqtd@JUSd~Rg0B4BM#k5@+<n|7!d@nABcY*ILrxuD*Q*O+Ch?)%Jm;W>8R
zk=dMmz9-d(y|s}-z-8%kliAGsFrF|o^{|LiLg5t51{A$zi($4v#}lrSH=PW`2y0+!
z$#~$`pLhC$FA<b@ViUy@IGlT4L@vaOuHJbY_!)n5%lodelE?^1_Jj(=fM_HF6&8dJ
zcorY~lf)N2a%2?V0NO&sHt*X)Ngj!`!UK*-F9DPuWiO8SwBy3jhR47YUeNiqQ!VlO
zgM8Y*g(1azK|TNV$ImoO7JlCE5p@EZ5&jFW7O%$}9dVBv;uv2UnMjQDYOb1^*qH1K
zm!HgBS>j_?+o+uCvL^y-cp&1>@8OYSl<nZXmtH>jy(u~5R_{_j@9Vk6id^{U8UAcN
zjWrV+8~UZkVh@i{m#4#7_Df6r@t4j_TvBNFK1~q+8?P(+)hZK>$G{l$_<jtsKG!yL
zcP_0umeSv-b(EIib_zRYW@EFz5$2^05Ee^@9J!6`U*(H{jGL4R?-Phc0q77KJ#g+Q
z(>z}ekU&IJAF8knuCJpxO*kpq9GsHqxH+4tF(ZL$x)3S_2CV84^d|5U^4?5B%JZqM
zqLF<^iUwN5z5#|Q+*{so1<q~DVp<=PWPhDFC<U|PBt$Y6zaZGWN_LHq7MclN)@Aj;
z4f<ew(XZ_bo;PoQ?g2B}pY}jNeaLZ7qUNG)+;TLOOU>mfMjE+#Hg5f5ta-t^M!P?B
zPNm-BysvF39L@D%$&3U4@3?Z8gy3{I%*m@`0u<297V;EUdTQNl`Ypjcy4M(@$CXZ`
z(sJdPMLVfzKVmxeNIIQcPIwNw)s2Aq!F#HdQakuPgPUkGyh<)>9`4bKsNp?i8+8(=
zVTrX!d2Yegq3G55fD5)j`9)GtRuFgt#~anM7T<NDGCU`1ms5`Z@{r)`N=q84^t5KZ
z#X7K`#nvJ3ZV3&^r7R-LQ5X3Z6$$F0-Lrg}jScIUO(u{K;zT(vQR7ot#^?pgTu&|h
z8fIK>6d8i~Irtw@;oTzw_76^_^pC;9OP@<omJY`vE?pKh=VHt+btm7FOb<qlp>(>$
z=}l3+b2#)6Gb`}!{hFs?oQ18cZkw2umDN;17E+7Oq;L;c(g<+KojKKerfu(-^bx#0
z^_tC_MNRoq*Kn;D3}8@q1F)tgb<n{oTq5!NJqSU@{!exgJ4bh!4TgFJsz08n9U_Jx
z!5aBKF|5MRL>jWoC^D*iIQG@A8AexqTLp-*Alj;HrFmH#j|X6_y=Xh(K&*7DudI(I
zDc|Su@)2&<@F#^9+LqdC(^ZcsmXeZ!srOTeX@%;6Nb^{m9~DfPj*&1gMm~I*lwENa
z5czSbgp9LmJJ9yl97=O3?#w2o_6Y-{Z$$h6-hnykQ`n1yghEYJl?C{xtN`1~D_;K*
zx{2}TEvJ}20KwnJmxa#<K!PX8q)E1OhNq}YZdE(&*ikGD+Od(lijsq|Dv4>Ygyt|3
z^Caywdoqph-h41EhYStLd5@`l;u>i8i{|8(RYhY4tCc2&Z^I>?G{Nlzaoz&R+XWu=
zSnXK%tGL*3@A~REHAR%=)57OxFoWaR3vU=mY6ee8>|--2f(5&tLNEkLn8Oy~fk!|D
zj+u8dTq%cenSV9m+QnYz@{6osP>+Jpwi7yJ-sq8o-#KT2we)uE2Z~?bU#A_PgLd5L
z6gAi{)h~9|axw7xo>afKvI&%C{MciiNSiFvg=|x(p-4Y84|&bo8EQDsXNWc@DmYQa
zMvY&JDI2Z$GSc)k*!l|yMbp?df?i@My7enGta>033QoI~54Q1*YW(#4YJC?>lEg#F
zHF|#w`&$HY`sCEL`jWwI!loAM>EjuWCD=0o;Izpgs8>;8w)?98bxmtYSToez#i5y%
z7zh3+(*Y8BlnL_)obSRpwk?XJ`;kjcgmZ3rTckT7;MjRb0!fMhEfc@n1g6agY3Lob
zs5z$ZP-;@R!Na&J$Waqx8D2U(5Ji`L(QbV*ehk8zL0e2}_C9@tS3XLZxbME&SVJDr
znf2{a>*OR0L%G-Ig-zz@g+=J4elMemKBtCPS{%nQsMlZ4-nkuqx;-EcmXRCrE80Fi
zQfS&sl3>H?J*_gwd5Y$x1M)usN!2?IbKsv6l+!|!glt4?F-Y~%Vj9C>I?$;wTt1C2
zs#+4y$c^#I7Sm+ff&o?+s(Dp-@WB|i?n;gJw6>V<%${UO2w9C}+tScjK`;wmfO!tK
ziWdH@!pXC|$1xu?%bDCCOJNV`__KK~QhriL%oL{+A}PP<Xybp_hVeMArJ#o-P%=cy
zI7V;UvLk$^ABw^GqI=vTE8DO6f;0(cXyMJG`K@_#qP;It7tvE3GyIKG(-jI0!jw&`
z>K^4cC@QQn+NF+uo=68gZ*~N+6c-{QWT{7OYG1c8Vc40k#7r(iiCWtBTuX_s;#)91
z$1>holpZ>29kWsIQlP><plCv$dP^t96O9%W{@j`{Yv~|+uO7#e9nC$tQmxLwif5aJ
zw6(<T?C~9ek<fUpSHJ$VW9I6#SrPqByN?~j7u1MXq3pQ9{#i4{(f#^iqg}XdmsFMP
zc)V)J9C_&J)ZUC0N0z9Jj~Rxm)TyYIN`^Gk+8FGVi@Z|^&Ww;%V!0<1hRdrqzmynX
z>DLJ?3}I9?J`d_MXn=9R`NnA@Xad6n3ZGm=D#B#}q&^(VEJ1s=b>>w$G(i1Ui?S+w
z@^IqQ!NQy>!SEg~_sMO#JZ*&L;z39MSy8m3uU1K@I;7Cgl&X%hiIjTrt~}}aDFGwM
z1~?lJjl{nuFyWV??_U)`R4IMU5WBtxNj7j7MD;e%CB8!<!+A(p(4+if`;>1zZl=yE
zacR+g*CjW9;N^9g!OTc3HNWNFb4*3M={?<IyJZYsId)0`-w-iM|Ef(|ZT1@9lx}Jx
z7z=ieBkEDZ-oXv4NXuOg$JGqzTM8?!dA!Y+unuq2X_NHoOv}4+Xo@dFun*A3d$DYA
zN{l?hc>{#o6(m6kT%&F#njFN7b2=}?@fzf>)Ux(V<0BRM?DB~`fC;G>&FY4RiFnYY
zk-_d=Foo5GUPqf<oWD=_^&%#o?^Z>Xn<h^1mR;OO5AT_wUB;DlJFFe7XR8_dn}v@8
zvpD7_MGm!=zY?T|TYW1S(&--1xjtU|{DluJrY?+y8hF+ue4)wc;ZHW#vp}bR80_C4
zWPl~@>m&SX+$59UZV~;CTq6(_9TM0ReH5F5DKb`_7X|Y=COtqh1RSTerRQOq+A!K9
zBn1}fSuc%nHV{sWqkhq$&Q&tg#KQVf_*+~mb5)bSVDDmGHY<)<dWWvsp<#=St_ED4
z(i^*cD8x{gr|-99MByCZA-6H{4MpMYZ$F}YlwX-%ND=B}NboVBTt2r}&~-zOD3ndP
zTS(zmW@n}D6(S1v+gFP?iKk7nu&Aq~)f31XT7V`~+Up(ze7()4MD&HtL=qkbCpwQa
zn}S|KFG2b~ZrMSXe<@71wczAuZ9I|$=VrT$*@`5STS*(&Bl>en<n=Ft=3j=K;W431
zko7&QfB2zZ(Q>>``c^G9-g@+--p2#0u(=hg?Fx)uO7e9_`U^4-*(J`N{W7d<zT6ZV
z4OPx}4RBd+-e^2Ra1XdYmeVTA4F8>vbd>xt!;hu<gO)KuuTd27qSarZj`MENKe2^v
z6MPhc{a94H(HoX@62WU{>)2U!yOmQym==u2Kk6PrBoU);*XLHbkJCaq-ur$1hTHSg
z+}y%Zl3gr3EuMvPVZAB-68aYqVcR5|{qniWgO8O-fu#v286n)f`AXjm>|*^Xhqg!i
z7Ywkqww|<H;byoTsibC;j36Sm`BMhPk`LG9^lLZ5`+-LU&;b}6{Rn)QYVa{B<2uY+
zA$Epw%-0F3=gIEaQLQ}eMK+^ZZn&l4Fl?pBcl&+=j!kL1x?`rpeArACc1gb_RuAlI
zBxy=_vA5f$5g>MxiozDP0*%m(Cpw;D7f!$e5Pz;oSz9RHohJYO(V4JwH8yjeB?Yf}
z$^;8vKcYD7)&$?PiEH2P>&WkRaZquVaWi2e!ET-b$N-538_WP!(clqVGEGX=toC{l
z5tGyvYOpYYW(x@o_4*I~hNhq3LJn*sSTFp+R8VUp0Ky6k)=%Lf(Mk+<Y(YDPO>{?g
zE?47~mF0xObeUaWD6_m9N|^QV!N6s)pdG7GlhnY%Y8R_h!Q(VxPDOfvh*vZ);z+FB
zGqyo5QZ>5EYLsegSB2;yCp|$->9#lY$xxBwKo9Go8h0kKk$75@s@Q(aaEId?xOh4S
z%bb=Kvr>9CH{rO-8=2}XG)NcN<U;Dg+`Jeh?D%mn!Gt-2&0r!EmZCJN!AB$NqesF^
zv^5A}on>eLtE56QbHjOml^LL)fjP>|kCI?0YQdRQV65}Sfr!#W2tM3UQ1~OT{wSg`
zEDleWtrP=9<Wq2O%Y}W!h=XV7RN^Nc$n0n>#K7t?zBrw~4wj6OjV2N+g(zavM&`4r
zy&}aepOzR_neWQdovX)3oR6kNlaLsaqTTk6YTudqoPa7}Y{d_EAAK<JgMYv9j!-Yj
z+1@uFAo_O2Eafhk$_rZf)=-=<croH|8`IUSkth@x&UFy4V7S74j)ML1J~;UmaAX)*
zsgIi`&%ER}1J^D^J=e(f%|#r$uvC?T-kGIAPP*u{er;rV^Odlrq<Z0|Zzsk5Y+(3!
zIfpz;4D)pp>1mZ&JYg@1MkqQYivlIX<}o!X`;q2j^`o#?r0rHamWH#(_4`sWEA3ec
z$fZ1TV?u!`SCerSG}m+Cy*&kqDz14)`lmYb9c36=kqao8&;P`k0(VvHhhwr<J10?P
zP<g`KGc^TC&vOMeaAxKyIG&GBa5ycL+#+N5a@*z&VJC>}QhxZnl*9as8B_TKGiFg~
zb;7YhIiSvwZmQY<P~NWY`aLn4)3#uFGOUxeWS1+sqmhJ^BD2yxp|M-rkuRWPRp_3u
zhcZX<7?~zjqxYd>`P{A#fhj!_GL`4so22RDsvSQ+r{LzM^6}2rX+;5Qhmijg0Zzg7
zK4F)TC8FohrE?SY9?ctucO&regpa?jW>ZRx*yZ2XzGf?uLKX@89E4J0Ig{>*tkO3g
z4*W$UMKdBX*vHOj^)0=+pkt=N%Y6JZ3?e+nw13K5VKDvVt;D!TOxQ;=@Ata^MI8@=
zsBckvJE(t$*4z$esc|p^v6ILQ(&00cY332ny2W%PH_&{~TJ5uya5?aMwGb?f9jAL6
zY?~VD7QyNbxPrxL0B*|FeZ|lP_e{76cqK_my6)0iK?CP05c&P#<B*{eJj_<?$u$5X
zzv($q^P5E*CCAYQo?1>8D3yi7;r0Aw``R-&`^$|ZpC<}=)nshE%;Sw(lWjA+`IS_d
zT{f!aTzk28s*VSH7A*DdKo^@;cfm&qD0qqhC*}?PlM^FJF&U{b8GR0*U{w&<Dr=}J
zyng11j8Q3^yWF+8ApA}V1>sMY%8T+`<KQVrg!uZ90K>nhrjW6IQB&?Y3P-v3HMQBg
znBQA`YcRNaAuv95+0zpaLd_b%GA9kg?s9Wi`OZ+`k<(JX-!C|V4iaFAHTpY&p?7kN
zBG)u6^gm;%#(;-m{5?y>Qr`#PW_^RC3ORW)$KoRsUDC*M_mikn?~CHo+!*R*-M3x|
z<=5iDCGceJ4GDUhbaR+rTX(3^@ctm^Gtem)juOc3zbIYCj?V;i$`L9QKq-;T=h40O
z_vPa@Hx{*+^T$-h<p}cS3+T>_y<9#`3S5ogjvoLnr#VL=Zo21z!BCf61s;ca%0LEK
z))(GLNQ`om1UQ*CA;ZN$ceQ+V;3A4eWQY}oZu+!R62QrQwalr@PtQK8n47J#S^0!K
z$7&e?2RHybRbhUk$*2Rw`-3X!tVq@-8KM_*uD-d&IQ9T!)Qq86{B^zQ^*gym`D@24
z!{C$SzR=`~Mi03~AV2(U=<S6vM&@hRvHO^`d`2LUPkDCqdEs)sD-9f2@QNo?%WXwK
z>Bc9QKthW2C1oI<$(6FNxVmybBHhy{;Kr1j7)b`?hto~&^TYlRB8vEc%;_IghliN5
zT=)55fd?Hw$d*5d0AH`v8VG^M3g77g(*j_&pcD}DvFk@9xi?!7K7T{+yDocgwm=bO
z)O#XFa&NZyUwt)Na7-ZKtbuz<WJ*|0uMz2xUP~@g&YPz=e{a_FsDC`Fs~B@=K^{1s
zcYkv{;teMRLo|BmEshi^0i|_k`?v1?xg8zZ{b#6_w{)tvCDqbg&PU(5J90OyzusFE
zy09L>Psh-WDFCGogbF~Z13cKj307~bD{Umh@wNC~t*h=~gLirb@EKb?@E>G2We=N^
z)^dI4ECoA7aDEVVML`OA-`zpB8f8FqS$5<hpj^b)xtm^}xEIakg(}3%8i?)I@2%<+
z6JqtUe7%Qd`Cu#uJWr>++3TQj&Cq<`dNnv-9jB1wiPe4NfV@Q=&qyC~k^>3EE|PD$
zkvs2bBb>PK3s*Po*f8qf1ulfyQe}2Ax{&x2eY!j0|EX9^zWsw@@n8(aT5_(I;b4_O
zuU>`c)NEr%i<@iP#Sp;3sGpDzKk)K+7~WJZ@DPLwM`)ca){I+J69#^9bp6n%@tpX^
zV-15VS<eokqJ0HYnC07yAko|VmZaC(jMZsU$Z*NmWg%Lz#$f-9oX%M=*>%TPJp!ud
znVeLg9w1GQaLix&Ll%M2!gvF1w&#*+g5Ab-(ZdZGnX=~ymLVeh-{N{-X1fJ+XHoSu
zpdZ;{WH*b`OMx_oSYTexx^SH~+7c^B&p2nkJ6_=t=n51TPlnWb%=D*(w&^e7c3kuK
z$|g>enZIoIY0l@4UK*D|@E_{+=ol%+!=R6JIJU(t0J+#w@yqwk48De_70j7rS{oUP
ze92f6?WC0yeio`|@cqOnF+x#S{Z>_0i(pv`Vm5?1<V&Q(p{6PRYD+)FydrEuC}y7`
zcE`S7<#QRujy!*mX~@y>vEBJ-b5X_D%p5?k);rcCCJp_R8|S)PpJ}GEZy5v$W;K?4
zb(}~{J>9Jhe2d=I<av#x^(}xI9%~5YIiq)RS#|J4DZ>D@&P_C@cN$P+P1qO*{*>Ls
zs9{X0YVp176;Y8F7N+Z{z-RU*+_5pEFF0J~M6GP?bA^5X|GFGhIR1CL9G3R#@bj{0
zdFusf$g0u>c7iNIc8X;$Mjzl<TfPX+fBwT$pdl^$$+!ye6iom06p(`pOFoo3&17<x
zgszMak>!jofuGj}u{v`!DflCJ<LVkt@<cC+MLTX@GL33tQn}isFZ(A*bd2|6>WyTi
z^+nAAs{O)2P+{()mlfv%{NuQ7a7&=F>ld(7`lJj`Lvm5lubp71?Su}J<Ui*pEG_c7
z66zb<(^neIVZp`|E<}~zTV>3FAIs|JkA<l8Eqt^fqvNT?ZM9i{2U@=!b>zC@T74nO
z63FG2gQ5qYULCjX{yOxqk}loukhwfah|=eOsY>Q@%jB_qGM;HD9K=wZN^5EIiG$73
z<i-4CPMd8{^70Fz@e-frBgx*7=9+Agl(F9B`D0$4&4xT+GF_*YT#XoZ&|$@!1Q}z5
zWd{P4R&rCkId2|V5kh~!RSK0%@fwgMoE7CXqUqN$#cPC{Go<cL`1;K);I(#MzB_{+
zjtB^Bp$DSYq(#E>i@piM3_2|1_hf58c9Cv<>A6n8Gl<2c5RBDj!K1}=J=^iwd%>(=
zs(j$hbFAoa<&=fvK;haYojV-_xi^ck)Ro*z$B4G!239(lkxXeo-Zht_PQJ<oP=b~f
zN`Pdr?8v>2i|v_DsMw3p`ZV+1_|5F{x>4hl9VfT@&7{*%>Ud^O_{qwGmI87EyA`>S
z^3cnb)J+gTY+Lm<wValzK7|t6!7vCXb1MR-Lti2*+_rm%1I0jYqOxb$Ii94N6ny>0
zBCi^^1$&?!TyBRGK6u3L;O@pWY9YY<8&(E^IDBin-?>ZX-x5ba4UCA@H&-GAIr}8j
zz63PPt+~`O%R)5@t0DuatrMqkg>_gzDC9v1ZOmsa!6<VZganrTEo7_O+E05az%%nS
z!>*T25Ae?)JV5sm72uUe5lstpJ+ByrJ>mEyUCTp`9+ZCP<VJt&p0|>;Ft$yMNqxO6
zp|ppi4{3&e_;Vrms_tYnZcIYM#MxX;KuE?eGAK+%s>m`lXcen(dd~nRq;~t<YG}yP
zdN18B>yyS0?9s2v8WcdfoJWWw7gbMMRCXUdpeAYji(F4`Ey~3Ih-g><wFSvTMZTaZ
ztclg-Jbb2IKY_5Vy@JhP3Z#?5mO`JD2ttlNS8jPpZk@k)NO_d;n40AKZ=Xe#4oV#|
zaHn3{mh66@dlipGkNVOEjxv(IY@a=1I0&OY-;W+YDX#1*VFX~004x!O{f5AVN=W}Y
zJ?k1pzM()!BfU9FjG^&Ad=LmJ%;C@&eGdu<wiRN?V~R^(US`rJzOYK9Pf6P79eh^x
z8Nw*f9ToTp688Qteuk`OCTWJoWOJ9CNf6ph_>hu!<dorAt2P&owTbW>BSURFJ*{?A
zGTR7_V`I2?V(O}|zf|c1njt6ppWPL5ixI{`-~7qO*xzwUx(CD5^lwU+k~ZSoQd^k8
z^#vPRU87Oz&Q>dO41b0wx~4|&T|UG_QM-4Ih>#2fp+nNx#E?zlz>q!(?Q3HGtgiXH
zC3L`|W{DKfimusIDA76DVr|x5;ythyA&)P*3JSV%UpVh}64G^t!70T=y~Jyjv@Crj
z_d%S(N43tT-74uWb49*?_vWxat^had!|5muwk&5g&g|Xv?SvzNq@*<UoKL<)AHDiw
zAh6=zPk}#A9z0<92OigfrhYXQ$8tkcGo;UUxUu?prw?#DEUwi#tnMJ(Ii}klQ5cTm
zZ@I;}V+nSOKmIkkoaFvxbO9!IQ#j>Nw<>2X4r0Z?D(0#NB?qFy636{U!y5v|ME6zY
zEGA|koe4$WreL1;V}?>L`NIsAGzpR2ssCWdOELAs+VF3j5{gB?8w)hlv(L16tG+~;
zQ#O^2DV<KvrEv7uS$JNthnsz7y)8QEL?$RUJ8s=)$xmU~KD-4g2mkJ4K2Y>p&V1`5
z?NE>CHq6uX7;s%qe@;`*UXHUN>&+di>u`m3DkWNb2R{g%Z>9%F9Wn(RJL{QU+gKaV
z?#&RlR<Y~mh;;JK`PqRHCH=AX{<M<i-bvKXAsCR?6)25-5&hz`07KVM8HPJBPOAbC
zuke{OMtz}y&yGVZGrPBSh&huk=yQfftT(PhLb0eVN0&>F07gh)3e*y=vi2Q%I)ds%
zD~kGue+n$?74HKA0I<Z!%()o3f7qRT5sEq&vwQUH;PxIu-;9($;0yX?g?P4i%F2@~
ze^t<3p;|`%V=HV*?5uWC`e`wB9A#_MLja1M!M$ZZO$jdmU(K=BlMm<sg;n*9zi)(I
zeL*VAo>=lZ>rZHDmiwHh6v5T<Amk^wOr@0Ft1k{ac`Xt13-fVY&APMp&Q^_iZZDP>
zSYF#E2K!o^F*??=ax?<uczw?A8v#6A{F;IIhU(sYbHnNYvkY$ib~(@IF84prhN%9a
zG)&LF;sI;s%buR^tT?sy1$PBXo12qj0IdLNwnz9)pw?AS#Xzkc!4tFNKjB!XoQ&&7
z3@T1U7pj9*N90vyUv&W_tc+PtL4(qG=@tlUs$~v1nEOA2mFP%6VI}a(w`Ds3QLgl=
zsQCb>L?)QmFX=1;%1ie-Yyk9)(T0?k8b(w?p2=NFthA{vyIUVIX{V4WX4PXP<Xy1s
zHs^odEI`SOvVK-`*-@ekJRC_xWe9LA6n%?)1uRI3CO|<IU(S@E^+CK0$#f)qM}bhV
zJ_N;G4)6?1Ce`1Zjz&@NxSh8;M@rm-)BA#|dvFR^j_|T3fm-vhO$Qm^)anlQB4!&)
z2B7JT@Je-!se|*_Hn-g1@}0)g%LJ6H#d;1d)4VKCkK-Ej(*%^Z$3M|HaBvgVtM@4J
zwFVmS2#DA$#4%wle0vQy!3*c>MT(bIn>6Y-+BUD*bX2zAPPG9yrAB~u`+rV{|LdF%
z@D<|A5U20)xkD?;*AY~{{Jmwd)I;mxBmua$W61X=6+MtC|DO#T%}nlRz0C=pK{p0B
z<0Tmto)^c3TS@-Oo|3rS?FcY8=Se49zISe%_@W*ofk(R4JHC`3W_RG=9R-7quaeY`
zok9k<!V2MCO+;$GFK;6XF5<TNdS{v2F<X2?-VgY03R8bm*g<hPZOXD7K?0n7HZJ&0
zS=iaL567gN(UHdwMa8rBEk-RnM#HC2&zAJu)kJS@U$hB+5bm+#7G05h2XH$`;MkNB
z5mL<v>(2G~Iy%;a(XArPWSnf5&0Yr5G@3g?RZULS!{!}eM{nIYsd3gsRtapP7nhlh
zIS8pqEKUy8`W;n85*<#(!O=T~VR5BtJgGmiWwj4Z!oI=3cRO6ZxNFo;sEcJrgPWQ-
zROM5lu{Epq>1M9_sMYXaw>ywopPkHGz~@@^Z1N6%wA9p=5En2qd|U2NL6_h_Hc6T{
z7hz~+$NGjrfOY7E_D_obvnAcWct2Ug*riMCr*>&=ip{LBY~VSa_LZxin=+CjpF228
zPNH2gN^)9VM^zq-rHIe=Ha7I<f_YVpWJay-m;R;+o;{k?WJS+Y5Lq`QYHO7eNpMq`
z7E}zh5|1p2_p$qGVy?CYV^irq4+rh~SyC9Yw$J4yoy51pD;a$U<OO8aOHxX#%#BZr
zGGd4je4oL*U?`<#^KNEmA|0iSS%<ijG1<(aa4QD>Q77e)6SL-aJg*6s%kuem=3kFY
zm?f`ZQPfLXfL1^3nqKy8e<7jL<%96SuZ7k#Uw4AuNzdmINEsO}t30-*R+iTJ;1Q@P
zVJGT837Wy3($6@|f5^Nz`vzN$7}v=6qf%9J`k$;Ivg$g9(8dQ<lFkMMT}4@v^7(BV
z#BLz#LkHhn6igVl=H60Iz(h~EH+*86S4*vWx-8Ox$`Th*ltj~-BQBgmQGeq$(zpcD
zQG%ATIPmu^WD7xUWiYwmMFJfB2H3btq#_sp&qV)!kcmF=Q{awi8TON)Nvdul!NVw+
z)+4J7-nf{z0)KMM<4OM$$2>XpXgz<yoOXrc{JmXKHGm=mj->Z_*n-QS(BLx8pT&RL
zHB$2sU~E}|`QtqQ6AV_r=VC(Xna3;`Vc%gw)DQ@VCL7bQ*F)gPlzUe<FcGO)dn~IS
z@ArO%5^A3BYwrCckN&5bFbM_dQ2Kt@T}SA_1vVrOS&5y#p0&#J#1S<+$$-szM=rN0
zM)63*2U`7)-a!gd>PiE%wHU^i(=caS2TykemWb?PTPb{uGeuOu26}j$@4wriD6a$T
zRnapL`k<T#T+X~w8nx_wwN!~|?A7U(A&c=7rU9o6@v3gIv{-|HD3r18z<n*L2UZ>a
zuaYSMWBOZ?ic(C#yw_92JeP!W|Ge5LU~l<a!;m0hy1?UCf3nBWyG9{{P3)wAsoUo@
z4=LCuh!D?;umyN762LY`7`4Jr58!5eDuYs*HYB6H4No1?-@iM6_n1?>b3nPwL|J&z
z5QYD|__J*1KR{KaTO;S1AZqL9Zy1^Z%&9mJwvmblp57Fe13bM6dhFI=Rq_sD&rT)8
zt?J^<6@$Hb>3L;AlJiyL9p1+Db!S{z8@rYPfl&D2;QJp`I;GytYb&i*b)|fu2UFhu
z-XRINfmI)k5~o<qj`jMDd<6$v08ZB@+uJ}!SLF`&sbLD+sBy3)i6aJYr?mnO#eD{l
zel+jVnSQtZ2vE@^qL9T-T|TErmElK;+ea50|7cLu=(#7A&jOh=?gq9}SP|mA(fud`
z^DL1f3Qp*zGHdpB*zIK1kP>I7dZ6r19mrzmVf^kN?}UuM&wtN|zcMykqNI_j8!2ul
zyd_dCm?9iLIT$o{dpqrQYJW8~=Kn6O`5-h{$>TgsKaIt=-uR|QC*}+j&;@}q;vc%;
ziga0gOgYLAAOu+0J^^tQcWTkqh<Dt2Cx;1Ngr%?#baVBp!>O6pk6sy6NY`joP|}po
zNT5w#$ublBx~)E%TqwgSW*q4?`lk@g2mU7nivTtqQnqn#h|h&m`sCAIa^B_y`^%GV
z3qTb>GT-5QDK~E3N!GJN38;Iu(u*$}SYaV`-Y&Y{GnK)kYvK1Z1BS)3<|_nOlUG&Q
zp*xO75)7wh3`OJKMeAH+88Lul1}IAmTp?F!3q|X-y}2*7ZXf_w0NmmPk@ww3-=c{w
z<a1XQO71P=BVty7HB7}tyriCnvqFZr091L6m*WGwHTz;>M<!ZHo5Np!pj@hs9Hr7M
zt3*Wn<=*RQldrE6Q(4=Lr5H{VDe!QsI*SLy=aMFEca4F~TOfJjT=lgpjonx#&V617
zDwV=o;My8A*N?)##oTXrO))na*l^EZz3N5zLN)!fEK9=`&@am`411UF?#XFgha>dS
z5}9YQycI>?*Rh`kF=c`5&Fu{5s*&3!)XpwLHKtEbf;~e@2KslBSHYNu!oWze?DYsJ
z_3#TIFP1I~NM_oA2PBiNKl;^~fV2CDT^;gv4vptrhLM7{#$bKR8;Hfd_(&H4;P8eb
z#GrPj@v1Ot+YGF(dnFstQH5!4ll9X8|9V1#mo>A?mgb#S!SF~~)O&+=D}U4HmzjUL
zEM<ZyhIPMIR9)bcgzO1?Eip3WW1#J{QN+5yx1xhG-iY0*j96n@rdtWi#PFDvPoI)P
z{ahw)lGozU>J>`iiF<juaDLuo?eNZQeLKby*oQuNm>|P;Zvj@=pBmjkmMUIo{3(E;
za+E$$pjeG!eM@YvA)h32J3;s(vYVyRup#WDvY9Y`O>{t+*e>Z3rfej`fbFKgb|uyP
zl1#TK{7wBDz#*UmB_O^e4-9}ddjk>5CeMy67tpn7Hz@${`hNht<ygGFH#T4{Ag(1a
zCl3s?a%-0Lwe~9j(@|$VOLG74#?LwregmEwXfa<vCH!b2bxcP8(98btf7Q!gIV!WQ
zX?l<F7QOyU2+t=E)a5*6v;XwIO6kEPiVxrUu)Q-u2+m%1KkJo5U>|(}ZM^MzD|G+=
z@7bY%&i{ocnm_(}<IMv`VmuR2=ac()zy0*U$d?_r0Ri}&H!tv>yf;Yar{V{B-GA=p
bvvW^ou+}3aa|@>jz<;8GQUYK;?GOJ0%ne7+

literal 0
HcmV?d00001

diff --git a/docs/zh_cn/reward_model/overview.md b/docs/zh_cn/reward_model/overview.md
index 84b5ab14b..6c7c976ac 100644
--- a/docs/zh_cn/reward_model/overview.md
+++ b/docs/zh_cn/reward_model/overview.md
@@ -24,6 +24,20 @@ XTuner 中的 Reward Model 训练具备以下显著优势：
 
 3. **高效训练**：借助 XTuner 的 QLoRA 训练功能，我们能够仅对 Reward Model 的 Value Head 进行全参数训练，而对语言模型本身使用 QLoRA 微调，大幅降低了模型训练的显存开销。
 
+4. **长文本训练**: 借助 XTuner 的序列并行功能，能够对长文本数据进行训练。
+
+![img](./images/sequence_parallel.png)
+
 ### 开始训练
 
 请参[阅快速上手](./quick_start.md)来了解最基本的概念，若希望了解更多训练参数配置相关的内容，请参考[修改Reward Model配置](./modify_settings.md)章节。
+
+### 开源模型
+
+我们使用 XTuner 训练了 InternLM2 技术报告中的 Reward Model，欢迎下载使用：
+
+| Model                     | Transformers(HF)                                                                 | ModelScope(HF)                                                                                             | OpenXLab(HF)                                                                                                                                                | RewardBench Score |
+| ------------------------- | -------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------- |
+| **InternLM2-1.8B-Reward** | [🤗internlm2-1_8b-reward](https://huggingface.co/internlm/internlm2-1_8b-reward) | [internlm2-1_8b-reward](https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-1_8b-reward/summary) | [![Open in OpenXLab](https://cdn-static.openxlab.org.cn/header/openxlab_models.svg)](https://openxlab.org.cn/models/detail/OpenLMLab/internlm2-1_8b-reward) | 80.6              |
+| **InternLM2-7B-Reward**   | [🤗internlm2-7b-reward](https://huggingface.co/internlm/internlm2-7b-reward)     | [internlm2-7b-reward](https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-7b-reward/summary)     | [![Open in OpenXLab](https://cdn-static.openxlab.org.cn/header/openxlab_models.svg)](https://openxlab.org.cn/models/detail/OpenLMLab/internlm2-7b-reward)   | 86.6              |
+| **InternLM2-20B-Reward**  | [🤗internlm2-20b-reward](https://huggingface.co/internlm/internlm2-20b-reward)   | [internlm2-20b-reward](https://modelscope.cn/models/Shanghai_AI_Laboratory/internlm2-20b-reward/summary)   | [![Open in OpenXLab](https://cdn-static.openxlab.org.cn/header/openxlab_models.svg)](https://openxlab.org.cn/models/detail/OpenLMLab/internlm2-20b-reward)  | 89.5              |

From 27cf856135ef9db63ec96257c1cb8caf4b87d329 Mon Sep 17 00:00:00 2001
From: LDLINGLINGLING <47373076+LDLINGLINGLING@users.noreply.github.com>
Date: Fri, 19 Jul 2024 17:14:55 +0800
Subject: [PATCH 19/29] =?UTF-8?q?Added=20minicpm=20config=20file=20to=20su?=
 =?UTF-8?q?pport=20sft=E3=80=81qlora=E3=80=81lora=E3=80=81dpo=20(#847)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* 增加了minicpm的template

* 增加了minicpm自定义预训练数据级的脚本demo

* 增加了自定义sft数据集训练的minicpmdemo

* 增加了minicpm_1b的dpo训练config

* 增加了minicpm_1b中文的lora训练cofig

* 增加了minicpm1b全量训练中文的config

* 增加了中英双语的minicpm1b的qlora训练config

* 增加了minicpm2b使用qlora训练dpo的config

* 增加了minicpm2b在中文上全量训练的config

* 增加了minicpm2b在中文下训练的loraconfig'

* 增加了中英双语下使用qlora训练minicpm2b的config
---
 .../minicpm_1b_full_custom_pretrain_e1.py     | 200 +++++++++++++++
 .../minicpm_2b_full_custom_pretrain_e1.py     | 200 +++++++++++++++
 .../minicpm_1b_full_custom_pretrain_e1.py     | 200 +++++++++++++++
 .../minicpm_2b_full_custom_pretrain_e1.py     | 200 +++++++++++++++
 .../minicpm/1_2b/minicpm_1b_dpo_qlora.py      | 219 ++++++++++++++++
 .../1_2b/minicpm_1b_full_alpaca_zh_e3.py      | 201 +++++++++++++++
 .../1_2b/minicpm_1b_lora_alpaca_zh_e3.py      | 213 ++++++++++++++++
 .../1_2b/minicpm_1b_qlora_alpaca_enzh_e3.py   | 238 ++++++++++++++++++
 .../1_2b/minicpm_1b_qlora_alpaca_zh_e3.py     | 221 ++++++++++++++++
 .../minicpm/2b/minicpm_2b_dpo_qlora.py        | 219 ++++++++++++++++
 .../2b/minicpm_2b_full_alpaca_zh_e3.py        | 201 +++++++++++++++
 .../2b/minicpm_2b_lora_alpaca_zh_e3.py        | 213 ++++++++++++++++
 .../2b/minicpm_2b_qlora_alpaca_enzh_e3.py     | 238 ++++++++++++++++++
 .../2b/minicpm_2b_qlora_alpaca_zh_e3.py       | 221 ++++++++++++++++
 xtuner/utils/templates.py                     |   3 +
 15 files changed, 2987 insertions(+)
 create mode 100644 xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py
 create mode 100644 xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_2b_full_custom_pretrain_e1.py
 create mode 100644 xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py
 create mode 100644 xtuner/configs/custom_dataset/sft/minicpm/minicpm_2b_full_custom_pretrain_e1.py
 create mode 100644 xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py
 create mode 100644 xtuner/configs/minicpm/1_2b/minicpm_1b_full_alpaca_zh_e3.py
 create mode 100644 xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py
 create mode 100644 xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_enzh_e3.py
 create mode 100644 xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_zh_e3.py
 create mode 100644 xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py
 create mode 100644 xtuner/configs/minicpm/2b/minicpm_2b_full_alpaca_zh_e3.py
 create mode 100644 xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py
 create mode 100644 xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_enzh_e3.py
 create mode 100644 xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_zh_e3.py

diff --git a/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py b/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py
new file mode 100644
index 000000000..bafe7f42c
--- /dev/null
+++ b/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py
@@ -0,0 +1,200 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""Data format:
+
+[
+  {
+      "text": "xxx"
+  },
+  {
+      "text": "xxx"
+  },
+  ...
+]
+"""  # noqa: E501
+
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import pretrain_map_fn
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+data_files = ['/root/ld/pull_request/xtuner/xtuner/configs/custom_dataset/pretrain/minicpm/pretrain.json']
+max_length = 2048
+pack_to_max_length = True
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1  # bs = 1 GPU * 1 batch_size_per_device * 16 acc
+dataloader_num_workers = 0
+max_epochs = 1
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = ['上海是', 'Shanghai is']
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path='json', data_files=data_files),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=pretrain_map_fn,
+    template_map_fn=None,
+    remove_unused_columns=True,
+    shuffle_before_pack=False,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_2b_full_custom_pretrain_e1.py b/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_2b_full_custom_pretrain_e1.py
new file mode 100644
index 000000000..160495a86
--- /dev/null
+++ b/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_2b_full_custom_pretrain_e1.py
@@ -0,0 +1,200 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""Data format:
+
+[
+  {
+      "text": "xxx"
+  },
+  {
+      "text": "xxx"
+  },
+  ...
+]
+"""  # noqa: E501
+
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import pretrain_map_fn
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+data_files = ['/path/to/json/file.json']
+max_length = 2048
+pack_to_max_length = True
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16  # bs = 1 GPU * 1 batch_size_per_device * 16 acc
+dataloader_num_workers = 0
+max_epochs = 1
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = ['上海是', 'Shanghai is']
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path='json', data_files=data_files),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=pretrain_map_fn,
+    template_map_fn=None,
+    remove_unused_columns=True,
+    shuffle_before_pack=False,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py b/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py
new file mode 100644
index 000000000..bafe7f42c
--- /dev/null
+++ b/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py
@@ -0,0 +1,200 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""Data format:
+
+[
+  {
+      "text": "xxx"
+  },
+  {
+      "text": "xxx"
+  },
+  ...
+]
+"""  # noqa: E501
+
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import pretrain_map_fn
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+data_files = ['/root/ld/pull_request/xtuner/xtuner/configs/custom_dataset/pretrain/minicpm/pretrain.json']
+max_length = 2048
+pack_to_max_length = True
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1  # bs = 1 GPU * 1 batch_size_per_device * 16 acc
+dataloader_num_workers = 0
+max_epochs = 1
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = ['上海是', 'Shanghai is']
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path='json', data_files=data_files),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=pretrain_map_fn,
+    template_map_fn=None,
+    remove_unused_columns=True,
+    shuffle_before_pack=False,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/custom_dataset/sft/minicpm/minicpm_2b_full_custom_pretrain_e1.py b/xtuner/configs/custom_dataset/sft/minicpm/minicpm_2b_full_custom_pretrain_e1.py
new file mode 100644
index 000000000..160495a86
--- /dev/null
+++ b/xtuner/configs/custom_dataset/sft/minicpm/minicpm_2b_full_custom_pretrain_e1.py
@@ -0,0 +1,200 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""Data format:
+
+[
+  {
+      "text": "xxx"
+  },
+  {
+      "text": "xxx"
+  },
+  ...
+]
+"""  # noqa: E501
+
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import pretrain_map_fn
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+data_files = ['/path/to/json/file.json']
+max_length = 2048
+pack_to_max_length = True
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16  # bs = 1 GPU * 1 batch_size_per_device * 16 acc
+dataloader_num_workers = 0
+max_epochs = 1
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = ''
+evaluation_inputs = ['上海是', 'Shanghai is']
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path='json', data_files=data_files),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=pretrain_map_fn,
+    template_map_fn=None,
+    remove_unused_columns=True,
+    shuffle_before_pack=False,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py
new file mode 100644
index 000000000..ed48f29d0
--- /dev/null
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py
@@ -0,0 +1,219 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from datasets import load_dataset
+import torch
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig
+from peft import LoraConfig
+from xtuner.dataset.collate_fns.preference_collate_fn import \
+    preference_collate_fn
+from xtuner.dataset.preference_dataset import (build_preference_dataset,
+                                               orpo_dpo_mix_40k_map_fn)
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model.dpo import DPO
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+dpo_loss_type = 'sigmoid'  # One of ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'sppo_hard', 'nca_pair', 'robust']  # noqa: E501
+loss_beta = 0.1
+label_smoothing = 0.0
+
+# Data
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 5e-7  # refer to alignment handbook
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    'What famous British author, known for his tales of mystery and the macabre, shares his initials with a common abbreviation for "rest in peace"?',  # noqa: E501
+    'Please tell me five scenic spots in Shanghai',
+    '890729 - 425663? Only respond with math and no words.'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=DPO,
+    use_varlen_attn=use_varlen_attn,
+    loss_type=dpo_loss_type,
+    beta=loss_beta,
+    label_smoothing=label_smoothing,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(type=load_dataset, path='mlabonne/orpo-dpo-mix-40k'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=True,
+    is_reward=False,
+    reward_token_id=-1,
+    num_proc=32,
+    use_varlen_attn=use_varlen_attn,
+    shuffle_before_pack=True,
+)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(
+        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_full_alpaca_zh_e3.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_full_alpaca_zh_e3.py
new file mode 100644
index 000000000..2c1e37ff3
--- /dev/null
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_full_alpaca_zh_e3.py
@@ -0,0 +1,201 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_en_path = 'silk-road/alpaca-data-gpt4-chinese'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_en,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py
new file mode 100644
index 000000000..428bdcd68
--- /dev/null
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py
@@ -0,0 +1,213 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+gradient_checkpointing = True
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        ),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_zh = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_zh_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_zh,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_enzh_e3.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_enzh_e3.py
new file mode 100644
index 000000000..0adc91aec
--- /dev/null
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_enzh_e3.py
@@ -0,0 +1,238 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import ConcatDataset, process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import (alpaca_map_fn, alpaca_zh_map_fn,
+                                    template_map_fn_factory)
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese'
+alpaca_en_path = 'tatsu-lab/alpaca'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+alpaca_zh = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_zh_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+train_dataset = dict(type=ConcatDataset, datasets=[alpaca_en, alpaca_zh])
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_zh_e3.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_zh_e3.py
new file mode 100644
index 000000000..ca7816c0a
--- /dev/null
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_qlora_alpaca_zh_e3.py
@@ -0,0 +1,221 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-1B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+gradient_checkpointing = True
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_zh = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_zh_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_zh,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py b/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py
new file mode 100644
index 000000000..6b395ae3e
--- /dev/null
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py
@@ -0,0 +1,219 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from datasets import load_dataset
+import torch
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig
+from peft import LoraConfig
+from xtuner.dataset.collate_fns.preference_collate_fn import \
+    preference_collate_fn
+from xtuner.dataset.preference_dataset import (build_preference_dataset,
+                                               orpo_dpo_mix_40k_map_fn)
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model.dpo import DPO
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+dpo_loss_type = 'sigmoid'  # One of ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'sppo_hard', 'nca_pair', 'robust']  # noqa: E501
+loss_beta = 0.1
+label_smoothing = 0.0
+
+# Data
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 5e-7  # refer to alignment handbook
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    'What famous British author, known for his tales of mystery and the macabre, shares his initials with a common abbreviation for "rest in peace"?',  # noqa: E501
+    'Please tell me five scenic spots in Shanghai',
+    '890729 - 425663? Only respond with math and no words.'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=DPO,
+    use_varlen_attn=use_varlen_attn,
+    loss_type=dpo_loss_type,
+    beta=loss_beta,
+    label_smoothing=label_smoothing,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+train_dataset = dict(
+    type=build_preference_dataset,
+    dataset=dict(type=load_dataset, path='mlabonne/orpo-dpo-mix-40k'),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=orpo_dpo_mix_40k_map_fn,
+    is_dpo=True,
+    is_reward=False,
+    reward_token_id=-1,
+    num_proc=32,
+    use_varlen_attn=use_varlen_attn,
+    shuffle_before_pack=True,
+)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(
+        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_full_alpaca_zh_e3.py b/xtuner/configs/minicpm/2b/minicpm_2b_full_alpaca_zh_e3.py
new file mode 100644
index 000000000..c699ff876
--- /dev/null
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_full_alpaca_zh_e3.py
@@ -0,0 +1,201 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_en_path = 'silk-road/alpaca-data-gpt4-chinese'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_en,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py b/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py
new file mode 100644
index 000000000..ff1988a68
--- /dev/null
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py
@@ -0,0 +1,213 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+gradient_checkpointing = True
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        ),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_zh = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_zh_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_zh,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_enzh_e3.py b/xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_enzh_e3.py
new file mode 100644
index 000000000..2082e4c24
--- /dev/null
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_enzh_e3.py
@@ -0,0 +1,238 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import ConcatDataset, process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import (alpaca_map_fn, alpaca_zh_map_fn,
+                                    template_map_fn_factory)
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese'
+alpaca_en_path = 'tatsu-lab/alpaca'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+alpaca_zh = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_zh_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+train_dataset = dict(type=ConcatDataset, datasets=[alpaca_en, alpaca_zh])
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_zh_e3.py b/xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_zh_e3.py
new file mode 100644
index 000000000..86d3564da
--- /dev/null
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_qlora_alpaca_zh_e3.py
@@ -0,0 +1,221 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_zh_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'openbmb/MiniCPM-2B-sft-bf16'
+use_varlen_attn = False
+
+# Data
+alpaca_zh_path = 'silk-road/alpaca-data-gpt4-chinese'
+prompt_template = PROMPT_TEMPLATE.minicpm
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 16
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+gradient_checkpointing = True
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right',
+    eos_token='</s>')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_zh = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_zh_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_zh_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_zh,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/utils/templates.py b/xtuner/utils/templates.py
index da1d96d2c..ae8ba1523 100644
--- a/xtuner/utils/templates.py
+++ b/xtuner/utils/templates.py
@@ -130,6 +130,9 @@
         SYSTEM=('[INST] {system} [/INST]\n'),
         INSTRUCTION=('[INST] {input} [/INST]'),
         SEP='\n'),
+    minicpm=dict(
+        INSTRUCTION=('<用户> {input} <AI>'),
+        SEP='\n'),
     gemma=dict(
         # `system` field is extended by xtuner
         SYSTEM=('<start_of_turn>system\n{system}<end_of_turn>\n'),

From f49ac9895336ff8f6936b91ab04725020b5c53b3 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 19 Jul 2024 17:50:12 +0800
Subject: [PATCH 20/29] fix lint (#856)

* fix lint

* fix lint
---
 .../minicpm/minicpm_1b_full_custom_pretrain_e1.py         | 2 +-
 .../sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py     | 2 +-
 xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py       | 8 +++++---
 .../configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py  | 5 ++---
 xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py         | 8 +++++---
 xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py | 5 ++---
 xtuner/utils/templates.py                                 | 4 +---
 7 files changed, 17 insertions(+), 17 deletions(-)

diff --git a/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py b/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py
index bafe7f42c..fc0da5ed3 100644
--- a/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py
+++ b/xtuner/configs/custom_dataset/pretrain/minicpm/minicpm_1b_full_custom_pretrain_e1.py
@@ -36,7 +36,7 @@
 use_varlen_attn = False
 
 # Data
-data_files = ['/root/ld/pull_request/xtuner/xtuner/configs/custom_dataset/pretrain/minicpm/pretrain.json']
+data_files = ['/path/to/json/file.json']
 max_length = 2048
 pack_to_max_length = True
 
diff --git a/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py b/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py
index bafe7f42c..fc0da5ed3 100644
--- a/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py
+++ b/xtuner/configs/custom_dataset/sft/minicpm/minicpm_1b_full_custom_pretrain_e1.py
@@ -36,7 +36,7 @@
 use_varlen_attn = False
 
 # Data
-data_files = ['/root/ld/pull_request/xtuner/xtuner/configs/custom_dataset/pretrain/minicpm/pretrain.json']
+data_files = ['/path/to/json/file.json']
 max_length = 2048
 pack_to_max_length = True
 
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py
index ed48f29d0..b0fc4556a 100644
--- a/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_dpo_qlora.py
@@ -1,13 +1,15 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from datasets import load_dataset
 import torch
+from datasets import load_dataset
 from mmengine.dataset import DefaultSampler
 from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
                             LoggerHook, ParamSchedulerHook)
 from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
-from torch.optim import AdamW
-from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig
 from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
 from xtuner.dataset.collate_fns.preference_collate_fn import \
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
diff --git a/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py b/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py
index 428bdcd68..e0ed46147 100644
--- a/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py
+++ b/xtuner/configs/minicpm/1_2b/minicpm_1b_lora_alpaca_zh_e3.py
@@ -7,8 +7,7 @@
 from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
 from peft import LoraConfig
 from torch.optim import AdamW
-from transformers import (AutoModelForCausalLM, AutoTokenizer,
-                          BitsAndBytesConfig)
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
@@ -78,7 +77,7 @@
         pretrained_model_name_or_path=pretrained_model_name_or_path,
         trust_remote_code=True,
         torch_dtype=torch.float16,
-        ),
+    ),
     lora=dict(
         type=LoraConfig,
         r=64,
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py b/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py
index 6b395ae3e..abf1e7ef9 100644
--- a/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_dpo_qlora.py
@@ -1,13 +1,15 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from datasets import load_dataset
 import torch
+from datasets import load_dataset
 from mmengine.dataset import DefaultSampler
 from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
                             LoggerHook, ParamSchedulerHook)
 from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
-from torch.optim import AdamW
-from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig
 from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
 from xtuner.dataset.collate_fns.preference_collate_fn import \
     preference_collate_fn
 from xtuner.dataset.preference_dataset import (build_preference_dataset,
diff --git a/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py b/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py
index ff1988a68..a50fe91ab 100644
--- a/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py
+++ b/xtuner/configs/minicpm/2b/minicpm_2b_lora_alpaca_zh_e3.py
@@ -7,8 +7,7 @@
 from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
 from peft import LoraConfig
 from torch.optim import AdamW
-from transformers import (AutoModelForCausalLM, AutoTokenizer,
-                          BitsAndBytesConfig)
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from xtuner.dataset import process_hf_dataset
 from xtuner.dataset.collate_fns import default_collate_fn
@@ -78,7 +77,7 @@
         pretrained_model_name_or_path=pretrained_model_name_or_path,
         trust_remote_code=True,
         torch_dtype=torch.float16,
-        ),
+    ),
     lora=dict(
         type=LoraConfig,
         r=64,
diff --git a/xtuner/utils/templates.py b/xtuner/utils/templates.py
index ae8ba1523..59b472731 100644
--- a/xtuner/utils/templates.py
+++ b/xtuner/utils/templates.py
@@ -130,9 +130,7 @@
         SYSTEM=('[INST] {system} [/INST]\n'),
         INSTRUCTION=('[INST] {input} [/INST]'),
         SEP='\n'),
-    minicpm=dict(
-        INSTRUCTION=('<用户> {input} <AI>'),
-        SEP='\n'),
+    minicpm=dict(INSTRUCTION=('<用户> {input} <AI>'), SEP='\n'),
     gemma=dict(
         # `system` field is extended by xtuner
         SYSTEM=('<start_of_turn>system\n{system}<end_of_turn>\n'),

From 5a93e7d7518d347a9c381ce261fa4390a4385b7b Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 19 Jul 2024 17:50:31 +0800
Subject: [PATCH 21/29] bump version to 0.1.22 (#855)

---
 xtuner/version.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/xtuner/version.py b/xtuner/version.py
index 85cbb5561..c77f4bf8f 100644
--- a/xtuner/version.py
+++ b/xtuner/version.py
@@ -1,5 +1,5 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-__version__ = '0.1.21'
+__version__ = '0.1.22'
 short_version = __version__
 
 

From f30ad4c4302573f64e703bcf2aa12de3fce98f75 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Haian=20Huang=28=E6=B7=B1=E5=BA=A6=E7=9C=B8=29?=
 <huanghaian@sensetime.com>
Date: Mon, 22 Jul 2024 14:45:09 +0800
Subject: [PATCH 22/29] Support InternVL 1.5/2.0 finetune (#737)

* support internvl finetune

* fix

* fix

* fix

* update

* update

* update

* update cfg

* fix

* support phi3

* support full+lora+qlora

* support internvl 26b

* fix lora cfg

* update all

* update

* update

* update config

* update config

* update config

* fix type and add readme

* update readme

* RENAME

* fix

* update

* support internvl2

* update
---
 xtuner/configs/internvl/README.md             | 152 +++++++
 xtuner/configs/internvl/README_zh-CN.md       | 152 +++++++
 .../internvl/v1_5/convert_to_official.py      |  56 +++
 .../internvl_v1_5_internlm2_26b_finetune.py   | 170 ++++++++
 ...ternvl_v1_5_internlm2_26b_lora_finetune.py | 183 ++++++++
 ...ernvl_v1_5_internlm2_26b_qlora_finetune.py | 185 ++++++++
 .../internvl_v1_5_internlm2_2b_finetune.py    | 170 ++++++++
 ...nternvl_v1_5_internlm2_2b_lora_finetune.py | 183 ++++++++
 ...ternvl_v1_5_internlm2_2b_qlora_finetune.py | 185 ++++++++
 .../v1_5/internvl_v1_5_phi3_4b_finetune.py    | 170 ++++++++
 .../internvl_v1_5_phi3_4b_lora_finetune.py    | 183 ++++++++
 .../internvl_v1_5_phi3_4b_qlora_finetune.py   | 185 ++++++++
 .../v2/internvl_v2_internlm2_26b_finetune.py  | 170 ++++++++
 ...internvl_v2_internlm2_26b_lora_finetune.py | 183 ++++++++
 ...nternvl_v2_internlm2_26b_qlora_finetune.py | 185 ++++++++
 .../v2/internvl_v2_internlm2_2b_finetune.py   | 170 ++++++++
 .../internvl_v2_internlm2_2b_lora_finetune.py | 183 ++++++++
 ...internvl_v2_internlm2_2b_qlora_finetune.py | 185 ++++++++
 .../v2/internvl_v2_internlm2_5_8b_finetune.py | 170 ++++++++
 ...nternvl_v2_internlm2_5_8b_lora_finetune.py | 183 ++++++++
 ...ternvl_v2_internlm2_5_8b_qlora_finetune.py | 185 ++++++++
 .../v2/internvl_v2_phi3_4b_finetune.py        | 170 ++++++++
 .../v2/internvl_v2_phi3_4b_lora_finetune.py   | 183 ++++++++
 .../v2/internvl_v2_phi3_4b_qlora_finetune.py  | 185 ++++++++
 xtuner/dataset/__init__.py                    |   3 +-
 .../dataset/collate_fns/default_collate_fn.py |   3 +-
 xtuner/dataset/internvl_dataset.py            | 409 ++++++++++++++++++
 xtuner/dataset/samplers/length_grouped.py     |   6 +
 xtuner/model/__init__.py                      |   3 +-
 xtuner/model/internvl.py                      | 320 ++++++++++++++
 30 files changed, 4867 insertions(+), 3 deletions(-)
 create mode 100644 xtuner/configs/internvl/README.md
 create mode 100644 xtuner/configs/internvl/README_zh-CN.md
 create mode 100644 xtuner/configs/internvl/v1_5/convert_to_official.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_qlora_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_qlora_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_qlora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_qlora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_qlora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_qlora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_phi3_4b_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_phi3_4b_lora_finetune.py
 create mode 100644 xtuner/configs/internvl/v2/internvl_v2_phi3_4b_qlora_finetune.py
 create mode 100644 xtuner/dataset/internvl_dataset.py
 create mode 100644 xtuner/model/internvl.py

diff --git a/xtuner/configs/internvl/README.md b/xtuner/configs/internvl/README.md
new file mode 100644
index 000000000..1f1acf191
--- /dev/null
+++ b/xtuner/configs/internvl/README.md
@@ -0,0 +1,152 @@
+# InterVL Full Pipeline
+
+English | [简体中文](./README_zh-CN.md)
+
+## InterVL 2
+
+> [InternVL-2: Better than the Best—Expanding Performance Boundaries of Open-Source Multimodal Models with the Progressive Scaling Strategy](https://internvl.github.io/blog/2024-07-02-InternVL-2.0/)
+
+We introduce InternVL-2, currently the most powerful open-source Multimodal Large Language Model (MLLM). The InternVL-2 family includes models ranging from a 2B model, suitable for edge devices, to a 108B model, which is significantly more powerful. With larger-scale language models, InternVL-2-Pro demonstrates outstanding multimodal understanding capabilities, matching the performance of commercial closed-source models across various benchmarks.
+
+InternVL-2 family is built upon the following designs:
+
+- Progressive with larger language models: We introduce a progressive alignment training strategy, resulting in the first vision foundation model aligned with large language models. By employing the progressive training strategy where the model scales from small to large while the data refines from coarse to fine, we have completed the training of large models at a relatively low cost. This approach has demonstrated excellent performance with limited resources.
+- Multimodal input: With one set of parameters, our model supports multiple modalities of input, including text, images, video, audio, and 3D point clouds.
+- Multitask output: Our model supports various output formats, such as images, bounding boxes, and masks, demonstrating extensive versatility. By connecting the MLLM with multiple downstream task decoders, InternVL-2 can be generalized to hundreds of vision-language tasks while achieving performance comparable to expert models.
+
+<div align="center">
+<img src="https://github.com/OpenGVLab/InternVL/assets/17425982/07845268-8b2c-4dc7-88dd-d10a173bdafe" alt="Image" />
+</div>
+
+### Basic Introduction
+
+- `./v2/` contains the configuration files for training InterVL 2
+- Supported fine-tuning of the InternVL 2B/4B/8B/26B model in full/LoRA/QLoRA single-image mode for now. We will support fine-tuning on multiple images and videos as soon as possible.
+- After training, you can use the `./v1_5/convert_to_official.py` script to convert the model trained by XTuner to the official format, so as to reuse all the official supported toolchains
+- All configurations are based on 8xA100 80G graphics cards, 2B/4B can use ZERO1 training, 8B models can use ZERO2, 26B models must run ZERO3, and there is no excessive adjustment of parameters, you can modify them according to your own needs
+- It is verified with LLaVA SFT data, which cannot fully reflect the fine-tuning performance. You can customize the data according to your own needs. We will provide a relatively fair fine-tuning dataset later
+
+### Data preparation
+
+If you also want to use the LLaVA SFT dataset for training, please refer to the [document](../../../docs/en/user_guides/dataset_prepare.md#llava-dataset) to prepare the data.
+
+For custom data, support multiple json and jsonl formats, the data organization can refer to the LLaVA SFT format, and support data sampling operations.
+
+**(1) Support multiple json or jsonl data**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+**(2) Support custom sampling**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    repeat_times=[2,0.5,3.5],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+### Training
+
+The provided configuration is mainly used for fine-tuning based on the official weights. After preparing the data, you can use the following command to train:
+
+```bash
+NPROC_PER_NODE=8 xtuner train internvl_v2_internlm2_5_8b_lora_finetune --deepspeed deepspeed_zero2
+```
+
+Default saved in `./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/`.
+
+### Model Conversion
+
+After training, we will get a set of weights, that is `./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/iter_xxx.pth`, in order to facilitate evaluation and dialogue, we can convert it to official weights.
+
+```bash
+python xtuner/configs/internvl/v1_5/convert_to_official.py xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_lora_finetune.py ./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/iter_xxx.pth ./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/convert_model/
+```
+
+Here, a complete set of official weights including configuration will be generated under `./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/convert_model`, you can use the [official toolchain](https://huggingface.co/OpenGVLab/InternVL2-8B) for evaluation and dialogue.
+
+If you encounter any problems during use, please feel free to contact us!!!
+
+## InterVL 1.5
+
+> [How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites](https://arxiv.org/abs/2404.16821)
+
+In this report, we introduce InternVL 1.5, an open-source multimodal large language model (MLLM) to bridge the capability gap between open-source and proprietary commercial models in multimodal understanding. We introduce three simple improvements: (1) Strong Vision Encoder: we explored a continuous learning strategy for the large-scale vision foundation model -- InternViT-6B, boosting its visual understanding capabilities, and making it can be transferred and reused in different LLMs. (2) Dynamic High-Resolution: we divide images into tiles ranging from 1 to 40 of 448×448 pixels according to the aspect ratio and resolution of the input images, which supports up to 4K resolution input. (3) High-Quality Bilingual Dataset: we carefully collected a high-quality bilingual dataset that covers common scenes, document images, and annotated them with English and Chinese question-answer pairs, significantly enhancing performance in OCR- and Chinese-related tasks. We evaluate InternVL 1.5 through a series of benchmarks and comparative studies. Compared to both open-source and proprietary models, InternVL 1.5 shows competitive performance, achieving state-of-the-art results in 8 of 18 benchmarks.
+
+<div align="center">
+<img src="https://github.com/InternLM/xtuner/assets/17425982/6dbe6a46-f01a-4c9d-ba44-0d857e5c0373" alt="Image" width="700" />
+</div>
+
+### Basic Introduction
+
+- `./v1_5/` contains the configuration files for training InterVL 1.5
+- Support InternVL 2B/4B/26B model full/LoRA/Qing efficiency and performance, it is recommended to choose the 4B model first
+- After training, you can use the `./v1_5/convert_to_official.py` script to convert the model trained by XTuner to the official format, so as to reuse all the official supported toolchains
+- All configurations are based on 8xA100 80G graphics cards, 2B/4B can use ZERO1 training, 8B models must run ZERO2, and there is no excessive adjustment of parameters, you can modify them according to your own needs
+- It is verified with LLaVA SFT data, which cannot fully reflect the fine-tuning performance. You can customize the data according to your own needs. We will provide a relatively fair fine-tuning dataset later
+
+### Data preparation
+
+If you also want to use the LLaVA SFT dataset for training, please refer to the [document](../../../docs/en/user_guides/dataset_prepare.md#llava-dataset) to prepare the data.
+
+For custom data, support multiple json and jsonl formats, the data organization can refer to the LLaVA SFT format, and support data sampling operations.
+
+**(1) Support multiple json or jsonl data**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+**(2) Support custom sampling**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    repeat_times=[2,0.5,3.5],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+### Training
+
+The provided configuration is mainly used for fine-tuning based on the official weights. After preparing the data, you can use the following command to train:
+
+```bash
+NPROC_PER_NODE=8 xtuner train internvl_v1_5_phi3_4b_lora_finetune --deepspeed deepspeed_zero1
+# NPROC_PER_NODE=8 xtuner train internvl_v1_5_internlm2_26b_lora_finetune.py --deepspeed deepspeed_zero3
+```
+
+Default saved in `./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/`.
+
+### Model Conversion
+
+After training, we will get a set of weights, that is `./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/iter_xxx.pth`, in order to facilitate evaluation and dialogue, we can convert it to official weights.
+
+```bash
+python xtuner/configs/internvl/v1_5/convert_to_official.py xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_lora_finetune.py ./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/iter_xxx.pth ./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/internvl_v1_5_phi3_4b/
+```
+
+Here, a complete set of official weights including configuration will be generated under `./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/internvl_v1_5_phi3_4b/`, you can use the [official toolchain](https://github.com/OpenGVLab/InternVL) for evaluation and dialogue.
+
+If you encounter any problems during use, please feel free to contact us!!!
diff --git a/xtuner/configs/internvl/README_zh-CN.md b/xtuner/configs/internvl/README_zh-CN.md
new file mode 100644
index 000000000..cdaa59348
--- /dev/null
+++ b/xtuner/configs/internvl/README_zh-CN.md
@@ -0,0 +1,152 @@
+# InterVL 全流程
+
+[English](./README.md) | 简体中文
+
+## InterVL 2
+
+> [InternVL-2: Better than the Best—Expanding Performance Boundaries of Open-Source Multimodal Models with the Progressive Scaling Strategy](https://internvl.github.io/blog/2024-07-02-InternVL-2.0/)
+
+我们引入了 InternVL-2,目前最强大的开源多模态大语言模型(MLLM)。InternVL-2 系列包括从适合于边缘设备的 2B 模型到强大的 108B 模型等多种规模的模型。借助更大规模的语言模型,InternVL-2-Pro 展现出了出色的多模态理解能力,在各种基准测试中的性能与商业闭源模型相匹配。
+
+InternVL-2 系列基于以下设计:
+
+- 渐进式的大型语言模型:我们引入了一种渐进式对齐训练策略,实现了首个与大型语言模型对齐的视觉基础模型。通过采用从小到大模型扩展、从粗到细数据优化的渐进式训练策略,我们以较低的成本完成了大模型的训练。这种方法已经展示了出色的性能,资源有限的情况下也能取得良好的结果。
+- 多模态输入:使用一套参数,我们的模型支持文本、图像、视频、音频和 3D 点云等多种输入模态。
+- 多任务输出:我们的模型支持图像、边界框和掩码等各种输出格式,展现出广泛的多功能性。通过将 MLLM 与多个下游任务解码器相连接,InternVL-2 可以泛化到数百个视觉语言任务,并取得与专家模型相当的性能。
+
+<div align="center">
+<img src="https://github.com/OpenGVLab/InternVL/assets/17425982/07845268-8b2c-4dc7-88dd-d10a173bdafe" alt="Image" />
+</div>
+
+### 基本说明
+
+- `./v2/` 包含着 InterVL 2 训练配置的配置文件
+- 支持了 InternVL 2B/4B/8B/26B 模型全量/LoRA/QLoRA 单图模式的微调，会尽快支持多图和视频的微调。
+- 在训练完成后，可以使用 `./v1_5/convert_to_official.py` 脚本将 XTuner 训练的模型转换为官方格式，从而复用官方所支持的所有工具链
+- 目前所有配置都是以 8xA100 80G 显卡为基准，2B/4B 可以使用 ZERO1 训练，8B 模型要 ZERO2 运行，26B 模型必须要 ZERO3，并且没有对参数进行过多的调整，你可以按照你自己的需求进行修改
+- 目前是以 LLaVA SFT 数据进行验证，无法充分反应微调性能，你可以根据自己的需求进行数据自定义，后续我们会提供一个相对公平的微调数据集
+
+### 数据准备
+
+如果你也想使用 LLaVA SFT 数据集进行训练，请参考[文档](../../../docs/zh_cn/user_guides/dataset_prepare.md#llava-dataset) 准备数据。
+
+对于自定义数据，支持多种 json 和 jsonl 格式，内部数据组织可以参考 LLaVA SFT 格式，且支持数据采样操作。
+
+**(1) 支持多个 json 或者 jsonl 数据**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+**(2) 支持自定义采样**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    repeat_times=[2,0.5,3.5],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+### 训练流程
+
+所提供的配置主要用于基于官方权重继续微调。在准备好数据后，你可以使用以下命令进行训练：
+
+```bash
+NPROC_PER_NODE=8 xtuner train internvl_v2_internlm2_5_8b_lora_finetune --deepspeed deepspeed_zero2
+```
+
+默认保存在 `./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/`。
+
+### 模型转换
+
+训练后，我们将获得一组权重即 `./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/iter_xxx.pth`，为了方便评测和对话，可以将其转换为官方权重。
+
+```bash
+python xtuner/configs/internvl/v1_5/convert_to_official.py xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_lora_finetune.py ./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/iter_xxx.pth ./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/convert_model/
+```
+
+此时，会在 `./work_dirs/internvl_v2_internlm2_5_8b_lora_finetune/convert_model` 下生成一组包括配置的完整官方权重，你可以使用[官方工具链](https://huggingface.co/OpenGVLab/InternVL2-8B)进行评测和对话。
+
+如果你在使用中碰到任何问题，欢迎联系我们！！！
+
+## InterVL 1.5
+
+> [How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites](https://arxiv.org/abs/2404.16821)
+
+在本报告中,我们介绍了开源多模态大语言模型 InternVL 1.5,以弥补开源模型与商业专有模型在多模态理解能力上的差距。我们引入了三项简单的改进:(1) 强大的视觉编码器:我们探索了大规模视觉基础模型 InternViT-6B 的连续学习策略,提升了其视觉理解能力,并使其可以在不同的大语言模型中进行迁移和重复利用。(2) 动态高分辨率:我们根据输入图像的长宽比和分辨率,将图像划分为从1到40个448×448像素的瓦片,支持高达4K分辨率的输入。(3) 高质量双语数据集:我们精心收集了一个高质量的双语数据集,涵盖了常见场景、文档图像,并用英语和中文问答对进行了注释,显著提升了在OCR和中文相关任务中的性能。我们通过一系列基准测试和对比研究评估了 InternVL 1.5。与开源和专有模型相比,InternVL 1.5 表现出了竞争力,在18个基准中的8个中取得了最先进的结果。
+
+<div align="center">
+<img src="https://github.com/InternLM/xtuner/assets/17425982/6dbe6a46-f01a-4c9d-ba44-0d857e5c0373" alt="Image" width="700" />
+</div>
+
+### 基本说明
+
+- `./v1_5/` 包含着 InterVL 1.5 训练配置的配置文件
+- 支持 InternVL 2B/4B/26B 模型全量/LoRA/QLoRA 微调，综合考虑效率性能，建议你优先选择 4B 模型
+- 在训练完成后，可以使用 `./v1_5/convert_to_official.py` 脚本将 XTuner 训练的模型转换为官方格式，从而复用官方所支持的所有工具链
+- 目前所有配置都是以 8xA100 80G 显卡为基准，2B/4B 可以使用 ZERO1 训练，26B 模型必须要 ZERO3 运行，并且没有对参数进行过多的调整，你可以按照你自己的需求进行修改
+- 目前是以 LLaVA SFT 数据进行验证，无法充分反应微调性能，你可以根据自己的需求进行数据自定义，后续我们会提供一个相对公平的微调数据集
+
+### 数据准备
+
+如果你也想使用 LLaVA SFT 数据集进行训练，请参考[文档](../../../docs/zh_cn/user_guides/dataset_prepare.md#llava-dataset) 准备数据。
+
+对于自定义数据，支持多种 json 和 jsonl 格式，内部数据组织可以参考 LLaVA SFT 格式，且支持数据采样操作。
+
+**(1) 支持多个 json 或者 jsonl 数据**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+**(2) 支持自定义采样**
+
+```text
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=['a.json','b.jsonl','c.json'],
+    image_folders=['a',None,'c'],
+    repeat_times=[2,0.5,3.5],
+    template=prompt_template,
+    max_length=max_length)
+```
+
+### 训练流程
+
+所提供的配置主要用于基于官方权重继续微调。在准备好数据后，你可以使用以下命令进行训练：
+
+```bash
+NPROC_PER_NODE=8 xtuner train internvl_v1_5_phi3_4b_lora_finetune --deepspeed deepspeed_zero1
+# NPROC_PER_NODE=8 xtuner train internvl_v1_5_internlm2_26b_lora_finetune.py --deepspeed deepspeed_zero3
+```
+
+默认保存在 `./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/`。
+
+### 模型转换
+
+训练后，我们将获得一组权重即 `./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/iter_xxx.pth`，为了方便评测和对话，可以将其转换为官方权重。
+
+```bash
+python xtuner/configs/internvl/v1_5/convert_to_official.py xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_lora_finetune.py ./work_dirs/iter_xxx.pth ./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/internvl_v1_5_phi3_4b/
+```
+
+此时，会在 `./work_dirs/internvl_v1_5_phi3_4b_lora_finetune/internvl_v1_5_phi3_4b/` 下生成一组包括配置的完整官方权重，你可以使用[官方工具链](https://github.com/OpenGVLab/InternVL)进行评测和对话。
+
+如果你在使用中碰到任何问题，欢迎联系我们！！！
diff --git a/xtuner/configs/internvl/v1_5/convert_to_official.py b/xtuner/configs/internvl/v1_5/convert_to_official.py
new file mode 100644
index 000000000..765855daa
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/convert_to_official.py
@@ -0,0 +1,56 @@
+import argparse
+import os.path as osp
+
+import torch
+from mmengine.config import Config
+from transformers import AutoTokenizer
+
+from xtuner.model.utils import LoadWoInit
+from xtuner.registry import BUILDER
+
+
+def convert_to_official(config, trained_path, save_path):
+    cfg = Config.fromfile(config)
+    cfg.model.pretrained_pth = trained_path
+    cfg.model.quantization_vit = False
+    cfg.model.quantization_llm = False
+
+    with LoadWoInit():
+        model = BUILDER.build(cfg.model)
+    model.to(torch.bfloat16)
+
+    if model.use_visual_encoder_lora:
+        vision_model = model.model.vision_model.merge_and_unload()
+        model.model.vision_model = vision_model
+
+    if model.use_llm_lora:
+        language_model = model.model.language_model.merge_and_unload()
+        model.model.language_model = language_model
+
+    model.model.save_pretrained(save_path)
+
+    tokenizer = AutoTokenizer.from_pretrained(
+        cfg.model.model_path, trust_remote_code=True)
+    tokenizer.save_pretrained(save_path)
+
+    print(model)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description='Convert the pth model to HuggingFace model')
+    parser.add_argument('config', help='config file name or path.')
+    parser.add_argument('trained_model_pth', help='The trained model path.')
+    parser.add_argument(
+        'save_path', help='The path to save the converted model.')
+    args = parser.parse_args()
+
+    if osp.realpath(args.trained_model_pth) == osp.realpath(args.save_path):
+        raise ValueError(
+            'The trained path and save path should not be the same.')
+
+    convert_to_official(args.config, args.trained_model_pth, args.save_path)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_finetune.py
new file mode 100644
index 000000000..d5eec7829
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL-Chat-V1-5'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+max_length = 4096
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 8
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 2e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.01
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_lora_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_lora_finetune.py
new file mode 100644
index 000000000..0fb511d42
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL-Chat-V1-5'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+max_length = 4096
+
+# Scheduler & Optimizer
+batch_size = 2  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 2e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.01
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_qlora_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_qlora_finetune.py
new file mode 100644
index 000000000..8d994c81d
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_26b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL-Chat-V1-5'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+max_length = 4096
+
+# Scheduler & Optimizer
+batch_size = 2  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 2e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.01
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_finetune.py
new file mode 100644
index 000000000..09fb01e3f
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/Mini-InternVL-Chat-2B-V1-5'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_lora_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_lora_finetune.py
new file mode 100644
index 000000000..193e2f269
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/Mini-InternVL-Chat-2B-V1-5'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_qlora_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_qlora_finetune.py
new file mode 100644
index 000000000..6bb28e490
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_internlm2_2b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/Mini-InternVL-Chat-2B-V1-5'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_finetune.py
new file mode 100644
index 000000000..5d34a928b
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.phi3_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_lora_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_lora_finetune.py
new file mode 100644
index 000000000..19588cb95
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.phi3_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_qlora_finetune.py b/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_qlora_finetune.py
new file mode 100644
index 000000000..cb150f0c4
--- /dev/null
+++ b/xtuner/configs/internvl/v1_5/internvl_v1_5_phi3_4b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/Mini-InternVL-Chat-4B-V1-5'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.phi3_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_finetune.py
new file mode 100644
index 000000000..0916df44a
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-26B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 8
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_lora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_lora_finetune.py
new file mode 100644
index 000000000..045fd7055
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-26B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 2  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_qlora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_qlora_finetune.py
new file mode 100644
index 000000000..60717b312
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_26b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-26B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 2  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_finetune.py
new file mode 100644
index 000000000..a921cf0c0
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-2B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_lora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_lora_finetune.py
new file mode 100644
index 000000000..44b3c3944
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-2B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_qlora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_qlora_finetune.py
new file mode 100644
index 000000000..5840a593f
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_2b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-2B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_finetune.py
new file mode 100644
index 000000000..2a92c017f
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-8B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_lora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_lora_finetune.py
new file mode 100644
index 000000000..d9fa7ab3a
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-8B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_qlora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_qlora_finetune.py
new file mode 100644
index 000000000..b3d04bb43
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_internlm2_5_8b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-8B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_finetune.py
new file mode 100644
index 000000000..41a712569
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_finetune.py
@@ -0,0 +1,170 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-4B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.phi3_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 4
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=False,
+    freeze_visual_encoder=True  # or False
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_lora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_lora_finetune.py
new file mode 100644
index 000000000..64a20450f
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_lora_finetune.py
@@ -0,0 +1,183 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-4B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.phi3_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_qlora_finetune.py b/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_qlora_finetune.py
new file mode 100644
index 000000000..8302fa5cc
--- /dev/null
+++ b/xtuner/configs/internvl/v2/internvl_v2_phi3_4b_qlora_finetune.py
@@ -0,0 +1,185 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset import InternVL_V1_5_Dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import InternVL_V1_5
+from xtuner.utils import PROMPT_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+path = 'OpenGVLab/InternVL2-4B'
+
+# Data
+data_root = './data/llava_data/'
+data_path = data_root + 'LLaVA-Instruct-150K/llava_v1_5_mix665k.json'
+image_folder = data_root + 'llava_images'
+prompt_template = PROMPT_TEMPLATE.phi3_chat
+max_length = 8192
+
+# Scheduler & Optimizer
+batch_size = 8  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+# official 1024 -> 4e-5
+lr = 1e-6
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 1000
+save_total_limit = 1  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+model = dict(
+    type=InternVL_V1_5,
+    model_path=path,
+    freeze_llm=True,
+    freeze_visual_encoder=True,
+    quantization_llm=True,  # or False
+    quantization_vit=False,  # or True and uncomment visual_encoder_lora
+    # comment the following lines if you don't want to use Lora in llm
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        target_modules=None,
+        task_type='CAUSAL_LM'),
+    # uncomment the following lines if you don't want to use Lora in visual encoder # noqa
+    # visual_encoder_lora=dict(
+    #     type=LoraConfig, r=64, lora_alpha=16, lora_dropout=0.05,
+    #     target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'])
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=InternVL_V1_5_Dataset,
+    model_path=path,
+    data_paths=data_path,
+    image_folders=image_folder,
+    template=prompt_template,
+    max_length=max_length)
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=llava_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=default_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=path,
+    trust_remote_code=True)
+
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        save_optimizer=False,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/dataset/__init__.py b/xtuner/dataset/__init__.py
index bcfe0dcc3..2ad3d7bd9 100644
--- a/xtuner/dataset/__init__.py
+++ b/xtuner/dataset/__init__.py
@@ -6,6 +6,7 @@
 from .intern_repo import (build_packed_dataset,
                           load_intern_repo_tokenized_dataset,
                           load_intern_repo_untokenized_dataset)
+from .internvl_dataset import InternVL_V1_5_Dataset
 from .json_dataset import load_json_file
 from .llava import LLaVADataset
 from .modelscope import process_ms_dataset
@@ -24,5 +25,5 @@
     'load_intern_repo_tokenized_dataset',
     'load_intern_repo_untokenized_dataset', 'build_packed_dataset',
     'RefCOCOJsonDataset', 'RefCOCOJsonEvalDataset', 'InvRefCOCOJsonDataset',
-    'load_json_file'
+    'load_json_file', 'InternVL_V1_5_Dataset'
 ]
diff --git a/xtuner/dataset/collate_fns/default_collate_fn.py b/xtuner/dataset/collate_fns/default_collate_fn.py
index 0ca9264f0..3d9fe18fb 100644
--- a/xtuner/dataset/collate_fns/default_collate_fn.py
+++ b/xtuner/dataset/collate_fns/default_collate_fn.py
@@ -89,7 +89,8 @@ def default_collate_fn(instances: Sequence[Dict],
         }
 
     if has_image:
-        pixel_values = torch.stack(pixel_values)
+        if all(x.shape == pixel_values[0].shape for x in pixel_values):
+            pixel_values = torch.stack(pixel_values, dim=0)
         data_dict['pixel_values'] = pixel_values
 
     if return_hf_format:
diff --git a/xtuner/dataset/internvl_dataset.py b/xtuner/dataset/internvl_dataset.py
new file mode 100644
index 000000000..82904ae87
--- /dev/null
+++ b/xtuner/dataset/internvl_dataset.py
@@ -0,0 +1,409 @@
+import copy
+import io
+import json
+import os
+import random
+import warnings
+
+import numpy as np
+import torch
+import torchvision.transforms as T
+from mmengine import print_log
+from mmengine.fileio import get
+from PIL import Image
+from torch.utils.data import Dataset
+from torchvision.transforms.functional import InterpolationMode
+from transformers import AutoConfig, AutoTokenizer
+
+from xtuner.utils import IGNORE_INDEX
+
+
+# Referenced from InternVL
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height,
+                              image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    return best_ratio
+
+
+def dynamic_preprocess(image,
+                       min_num=1,
+                       max_num=6,
+                       image_size=448,
+                       use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = {(i, j)
+                     for n in range(min_num, max_num + 1)
+                     for i in range(1, n + 1) for j in range(1, n + 1)
+                     if i * j <= max_num and i * j >= min_num}
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(aspect_ratio,
+                                                    target_ratios, orig_width,
+                                                    orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = ((i % (target_width // image_size)) * image_size,
+               (i // (target_width // image_size)) * image_size,
+               ((i % (target_width // image_size)) + 1) * image_size,
+               ((i // (target_width // image_size)) + 1) * image_size)
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images
+
+
+def total_image_token(orig_size,
+                      min_num=1,
+                      max_num=12,
+                      image_size=448,
+                      use_thumbnail=True):
+    orig_width, orig_height = orig_size
+
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = {(i, j)
+                     for n in range(min_num, max_num + 1)
+                     for i in range(1, n + 1) for j in range(1, n + 1)
+                     if max_num >= i * j >= min_num}
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(aspect_ratio,
+                                                    target_ratios, orig_width,
+                                                    orig_height, image_size)
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    if use_thumbnail:
+        blocks += 1
+
+    return blocks
+
+
+def load_json_or_jsonl(json_path):
+    if json_path.endswith('.json'):
+        with open(json_path) as f:
+            data = json.load(f)
+    elif json_path.endswith('.jsonl'):
+        with open(json_path) as f:
+            data = [json.loads(line) for line in f]
+    else:
+        raise ValueError(f'Unsupported file format: {json_path}, '
+                         f'only support .json and .jsonl.')
+    return data
+
+
+class InternVL_V1_5_Dataset(Dataset):
+    os.environ['TOKENIZERS_PARALLELISM'] = 'true'
+    IMG_CONTEXT_TOKEN = '<IMG_CONTEXT>'
+    IMG_START_TOKEN = '<img>'
+    IMG_END_TOKEN = '</img>'
+
+    IMAGENET_MEAN = (0.485, 0.456, 0.406)
+    IMAGENET_STD = (0.229, 0.224, 0.225)
+
+    def __init__(self,
+                 model_path,
+                 template,
+                 data_paths,
+                 image_folders=None,
+                 repeat_times=1,
+                 max_length=8192):
+        self.template = template
+        self.max_length = max_length
+
+        self.cfg = AutoConfig.from_pretrained(
+            model_path, trust_remote_code=True)
+
+        # The following modifications are only to ensure full
+        # consistency with the official template,
+        # without investigating the impact on performance.
+        if self.cfg.llm_config.architectures[0] == 'Phi3ForCausalLM':
+            self._system = 'You are an AI assistant whose name is Phi-3.'
+            self.template[
+                'INSTRUCTION'] = '<|user|>\n{input}<|end|><|assistant|>\n'
+        elif self.cfg.llm_config.architectures[0] == 'InternLM2ForCausalLM':
+            self._system = 'You are an AI assistant whose name ' \
+                           'is InternLM (书生·浦语).'
+            self.template['SYSTEM'] = '<|im_start|>system\n{system}<|im_end|>'
+            self.template[
+                'INSTRUCTION'] = '<|im_start|>user\n{input}' \
+                                 '<|im_end|><|im_start|>assistant\n'
+        else:
+            raise NotImplementedError
+
+        self.min_dynamic_patch = self.cfg.min_dynamic_patch
+        self.max_dynamic_patch = self.cfg.max_dynamic_patch
+        self.downsample_ratio = self.cfg.downsample_ratio
+        self.image_size = self.cfg.force_image_size
+        self.use_thumbnail = self.cfg.use_thumbnail
+        patch_size = self.cfg.vision_config.patch_size
+        self.patch_token = int(
+            (self.image_size // patch_size)**2 * (self.downsample_ratio**2))
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_path, trust_remote_code=True)
+        self.transformer = T.Compose([
+            T.Lambda(lambda img: img.convert('RGB')
+                     if img.mode != 'RGB' else img),
+            T.Resize((self.image_size, self.image_size),
+                     interpolation=InterpolationMode.BICUBIC),
+            T.ToTensor(),
+            T.Normalize(mean=self.IMAGENET_MEAN, std=self.IMAGENET_STD)
+        ])
+
+        if not isinstance(data_paths, (list, tuple)):
+            data_paths = [data_paths]
+        if not isinstance(image_folders, (list, tuple)):
+            image_folders = [image_folders]
+        if not isinstance(repeat_times, (list, tuple)):
+            repeat_times = [repeat_times]
+        assert len(data_paths) == len(image_folders) == len(repeat_times)
+
+        print_log('Starting to loading data and calc length', logger='current')
+        self.data = []
+        self.image_folder = []
+        self.group_length = []
+        self.conv2length_text = {
+        }  # using dict to speedup the calculation of token length
+
+        for data_file, image_folder, repeat_time in zip(
+                data_paths, image_folders, repeat_times):
+            print_log(
+                f'=======Starting to process {data_file} =======',
+                logger='current')
+            assert repeat_time > 0
+            json_data = load_json_or_jsonl(data_file)
+            if repeat_time < 1:
+                json_data = random.sample(json_data,
+                                          int(len(json_data) * repeat_time))
+            elif repeat_time > 1:
+                int_repeat_time = int(repeat_time)
+                remaining_repeat_time = repeat_time - repeat_time
+                if remaining_repeat_time > 0:
+                    remaining_json_data = random.sample(
+                        json_data, int(len(json_data) * remaining_repeat_time))
+                    json_data = json_data * int_repeat_time
+                    json_data.extend(remaining_json_data)
+                else:
+                    json_data = json_data * int_repeat_time
+
+            self.data.extend(json_data)
+            self.image_folder.extend([image_folder] * len(json_data))
+
+            # TODO: multi process
+            for data_item in json_data:
+                if 'length' in data_item:
+                    token_length = data_item['length']  # include image token
+                else:
+                    conversations = '\n'.join(
+                        [temp['value'] for temp in data_item['conversations']])
+                    str_length = len(conversations)
+
+                    if str_length not in self.conv2length_text:
+                        token_length = self.tokenizer(
+                            conversations,
+                            return_tensors='pt',
+                            padding=False,
+                            truncation=False,
+                        ).input_ids.size(1)
+                        self.conv2length_text[str_length] = token_length
+                    else:
+                        token_length = self.conv2length_text[str_length]
+
+                    if 'image' in data_item and data_item['image'] is not None:
+                        if 'image_wh' in data_item and data_item[
+                                'image_wh'] is not None:
+                            # more accurate calculation of image token
+                            image_wh = data_item['image_wh']
+                            if isinstance(image_wh[0], list):
+                                image_wh = image_wh[0]
+                            image_token = total_image_token(
+                                image_wh, self.min_dynamic_patch,
+                                self.max_dynamic_patch, self.image_size,
+                                self.use_thumbnail)
+                            image_token = self.patch_token * image_token
+                        else:
+                            # max_dynamic_patch + use_thumbnail
+                            image_token = self.patch_token * (
+                                self.max_dynamic_patch + self.use_thumbnail)
+
+                        token_length = token_length + image_token
+                    else:
+                        token_length = -token_length
+
+                self.group_length.append(token_length)
+            print_log(
+                f'=======total {len(json_data)} samples of {data_file}=======',
+                logger='current')
+
+        assert len(self.group_length) == len(self.data)
+        print_log('end loading data and calc length', logger='current')
+        print_log(
+            f'=======total {len(self.data)} samples=======', logger='current')
+        self._max_refetch = 1000
+
+    def __getitem__(self, index):
+        for _ in range(self._max_refetch + 1):
+            data = self.prepare_data(index)
+            # Broken images may cause the returned data to be None
+            if data is None:
+                index = self._rand_another()
+                continue
+            return data
+
+    def __len__(self):
+        return len(self.data)
+
+    @property
+    def modality_length(self):
+        return self.group_length
+
+    @property
+    def length(self):
+        group_length = np.array(self.group_length)
+        group_length = np.abs(group_length).tolist()
+        return group_length
+
+    def prepare_data(self, index):
+        data_dict: dict = self.data[index]
+        image_folder = self.image_folder[index]
+
+        out_data_dict = {}
+        if data_dict.get('image', None) is not None:
+            image_file = data_dict['image']
+            if isinstance(image_file, (list, tuple)):
+                assert len(image_file) == 1
+                image_file = image_file[0]
+
+            try:
+                image = self.get_image(os.path.join(image_folder, image_file))
+            except Exception as e:
+                print(f'Error: {e}', flush=True)
+                print_log(f'Error: {e}', logger='current')
+                return None
+
+            images = dynamic_preprocess(image, self.min_dynamic_patch,
+                                        self.max_dynamic_patch,
+                                        self.image_size, self.use_thumbnail)
+            pixel_values = [self.transformer(image) for image in images]
+            pixel_values = torch.stack(pixel_values)
+            out_data_dict['pixel_values'] = pixel_values
+
+            num_image_tokens = pixel_values.shape[0] * self.patch_token
+            image_token_str = f'{self.IMG_START_TOKEN}' \
+                              f'{self.IMG_CONTEXT_TOKEN * num_image_tokens}' \
+                              f'{self.IMG_END_TOKEN}'
+            token_dict = self.get_inputid_labels(data_dict['conversations'],
+                                                 image_token_str)
+            out_data_dict.update(token_dict)
+        else:
+            token_dict = self.get_inputid_labels(data_dict['conversations'],
+                                                 None)
+            out_data_dict.update(token_dict)
+            out_data_dict['pixel_values'] = torch.zeros(
+                1, 3, self.image_size, self.image_size)
+        return out_data_dict
+
+    def _rand_another(self) -> int:
+        return np.random.randint(0, len(self.data))
+
+    def get_image(self, path):
+        if 's3://' in path:
+            img_bytes = get(path)
+            with io.BytesIO(img_bytes) as buff:
+                img = Image.open(buff).convert('RGB')
+            return img
+        else:
+            return Image.open(path).convert('RGB')
+
+    def get_inputid_labels(self, conversations, image_token_str) -> dict:
+        input = ''
+        out_conversation = []
+        while conversations and conversations[0]['from'] == 'gpt':
+            # Skip the first one if it is from gpt
+            conversations = conversations[1:]
+        for msg in conversations:
+            if msg['from'] == 'human':
+                if image_token_str is None and '<image>' in msg['value']:
+                    warnings.warn(
+                        f'The current data << {msg["value"]} >> is '
+                        f'in plain text mode, but '
+                        'there are <image> tags present in the data. '
+                        'We need to remove the <image> tags.')
+                    msg['value'] = msg['value'].replace('<image>', '')
+                if '<image>' in msg['value']:
+                    msg['value'] = msg['value'].replace('<image>', '').strip()
+                    msg['value'] = image_token_str + '\n' + msg['value']
+                    msg['value'] = msg['value'].strip()
+                input += msg['value'].strip()
+            elif msg['from'] == 'gpt':
+                out_conversation.append({
+                    'input': input,
+                    'output': msg['value'].strip()
+                })
+                input = ''
+            else:
+                raise NotImplementedError
+
+        input_ids, labels = [], []
+        for i, single_turn_conversation in enumerate(out_conversation):
+            input = single_turn_conversation.get('input', '')
+            if input is None:
+                input = ''
+            input_text = self.template.INSTRUCTION.format(
+                input=input, round=i + 1)
+
+            if i == 0:
+                system = self.template.SYSTEM.format(system=self._system)
+                input_text = system + input_text
+                input_encode = self.tokenizer.encode(
+                    input_text, add_special_tokens=True)
+            else:
+                input_encode = self.tokenizer.encode(
+                    input_text, add_special_tokens=False)
+            input_ids += input_encode
+            labels += [IGNORE_INDEX] * len(input_encode)
+
+            output_text = single_turn_conversation.get('output', '')
+            if self.template.get('SUFFIX', None):
+                output_text += self.template.SUFFIX
+            output_encode = self.tokenizer.encode(
+                output_text, add_special_tokens=False)
+            input_ids += output_encode
+            labels += copy.deepcopy(output_encode)
+
+        if len(input_ids) > self.max_length:
+            input_ids = input_ids[:self.max_length]
+            labels = labels[:self.max_length]
+            print_log(
+                f'Warning: input_ids length({len(input_ids)}) '
+                f'is longer than max_length, cut to {self.max_length}',
+                logger='current')
+        return {'input_ids': input_ids, 'labels': labels}
diff --git a/xtuner/dataset/samplers/length_grouped.py b/xtuner/dataset/samplers/length_grouped.py
index ad37957f2..184827837 100644
--- a/xtuner/dataset/samplers/length_grouped.py
+++ b/xtuner/dataset/samplers/length_grouped.py
@@ -4,6 +4,7 @@
 
 import torch
 from mmengine.dist import get_dist_info, sync_random_seed
+from mmengine.logging import print_log
 from torch.utils.data import ConcatDataset as TorchConcatDataset
 from torch.utils.data import Sampler
 
@@ -78,6 +79,7 @@ def __init__(self,
                  mega_batch_mult: Optional[int] = None,
                  seed: Optional[int] = None,
                  round_up: bool = True) -> None:
+        print_log('LengthGroupedSampler is used.', logger='current')
         rank, world_size = get_dist_info()
         self.rank = rank
         self.world_size = world_size
@@ -120,6 +122,10 @@ def __init__(self,
         assert isinstance(self.length, (list, tuple))
 
         self.total_batch_size = total_batch_size
+        print_log(
+            f'LengthGroupedSampler construction is complete, '
+            f'and the selected attribute is {length_property}',
+            logger='current')
 
     def __iter__(self) -> Iterator[int]:
         """Iterate the indices."""
diff --git a/xtuner/model/__init__.py b/xtuner/model/__init__.py
index 39547b2d7..1b3a501d4 100644
--- a/xtuner/model/__init__.py
+++ b/xtuner/model/__init__.py
@@ -1,5 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+from .internvl import InternVL_V1_5
 from .llava import LLaVAModel
 from .sft import SupervisedFinetune
 
-__all__ = ['SupervisedFinetune', 'LLaVAModel']
+__all__ = ['SupervisedFinetune', 'LLaVAModel', 'InternVL_V1_5']
diff --git a/xtuner/model/internvl.py b/xtuner/model/internvl.py
new file mode 100644
index 000000000..0358266a9
--- /dev/null
+++ b/xtuner/model/internvl.py
@@ -0,0 +1,320 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from collections import OrderedDict
+from typing import List, Optional, Tuple, Union
+
+import torch
+from mmengine import print_log
+from mmengine.config import Config, ConfigDict
+from mmengine.model import BaseModel
+from peft import get_peft_model, prepare_model_for_kbit_training
+from torch.nn import CrossEntropyLoss
+from transformers import (AutoConfig, AutoModel, AutoTokenizer,
+                          BitsAndBytesConfig)
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from xtuner.registry import BUILDER
+from .utils import (find_all_linear_names, get_peft_model_state_dict,
+                    guess_load_checkpoint, make_inputs_require_grad)
+
+
+class InternVL_V1_5(BaseModel):
+
+    def __init__(self,
+                 model_path,
+                 freeze_llm=False,
+                 freeze_visual_encoder=False,
+                 llm_lora=None,
+                 visual_encoder_lora=None,
+                 quantization_vit=False,
+                 quantization_llm=False,
+                 pretrained_pth=None):
+        print_log('Start to load InternVL_V1_5 model.', logger='current')
+        super().__init__()
+        self.freeze_llm = freeze_llm
+        self.freeze_visual_encoder = freeze_visual_encoder
+        self.use_llm_lora = llm_lora is not None
+        self.use_visual_encoder_lora = visual_encoder_lora is not None
+        self.quantization_vit = quantization_vit
+        self.quantization_llm = quantization_llm
+        if quantization_vit:
+            assert visual_encoder_lora is not None
+        if quantization_llm:
+            assert quantization_llm and llm_lora is not None
+
+        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+        if config.llm_config.model_type == 'internlm2':
+            config.llm_config.attn_implementation = 'flash_attention_2'
+        else:
+            config.llm_config._attn_implementation = 'flash_attention_2'
+
+        if quantization_vit is False and quantization_llm is False:
+            quantization = None
+        else:
+            llm_int8_skip_modules = ['mlp1']
+            if quantization_llm and not quantization_vit:
+                llm_int8_skip_modules.append('vision_model')
+
+            if quantization_vit and not quantization_llm:
+                llm_int8_skip_modules.append('language_model')
+
+            quantization_config = dict(
+                type=BitsAndBytesConfig,
+                llm_int8_skip_modules=llm_int8_skip_modules,
+                load_in_4bit=True,
+                load_in_8bit=False,
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=False,
+                bnb_4bit_compute_dtype=torch.float16,
+                bnb_4bit_use_double_quant=True,
+                bnb_4bit_quant_type='nf4')
+            quantization_clazz = quantization_config.pop('type')
+            quantization = quantization_clazz(**quantization_config)
+
+        self.model = AutoModel.from_pretrained(
+            model_path,
+            torch_dtype=torch.bfloat16,
+            quantization_config=quantization,
+            config=config,
+            trust_remote_code=True)
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_path, trust_remote_code=True)
+        img_context_token_id = tokenizer.convert_tokens_to_ids('<IMG_CONTEXT>')
+        self.model.img_context_token_id = img_context_token_id
+
+        if self.freeze_llm:
+            self.model.language_model.requires_grad_(False)
+        if self.freeze_visual_encoder:
+            self.model.vision_model.requires_grad_(False)
+
+        if hasattr(self.model.language_model, 'enable_input_require_grads'):
+            self.model.language_model.enable_input_require_grads()
+        else:
+            self.model.language_model.get_input_embeddings(
+            ).register_forward_hook(make_inputs_require_grad)
+
+        self.gradient_checkpointing_enable()
+
+        if self.use_llm_lora:
+            self._prepare_llm_for_lora(llm_lora)
+
+        if self.use_visual_encoder_lora:
+            self._prepare_visual_encoder_for_lora(visual_encoder_lora)
+
+        if pretrained_pth is not None:
+            pretrained_state_dict = guess_load_checkpoint(pretrained_pth)
+
+            self.load_state_dict(pretrained_state_dict, strict=False)
+            print(f'Load pretrained weight from {pretrained_pth}')
+
+        self._count = 0
+        print_log(self, logger='current')
+        print_log('InternVL_V1_5 construction is complete', logger='current')
+
+    def _parse_lora_config(self, lora_config):
+        if isinstance(lora_config, dict) or isinstance(
+                lora_config, Config) or isinstance(lora_config, ConfigDict):
+            lora_config = BUILDER.build(lora_config)
+        return lora_config
+
+    def _prepare_llm_for_lora(self,
+                              lora_config,
+                              use_activation_checkpointing=True):
+        lora_config = self._parse_lora_config(lora_config)
+        self.model.language_model = prepare_model_for_kbit_training(
+            self.model.language_model, use_activation_checkpointing)
+        if lora_config.target_modules is None:
+            modules = find_all_linear_names(self.model.language_model)
+            lora_config.target_modules = modules
+        self.model.language_model = get_peft_model(self.model.language_model,
+                                                   lora_config)
+
+    def _prepare_visual_encoder_for_lora(self, lora_config):
+        lora_config = self._parse_lora_config(lora_config)
+        if lora_config.target_modules is None:
+            modules = find_all_linear_names(self.model.vision_model)
+            lora_config.target_modules = modules
+        self.model.vision_model = get_peft_model(self.model.vision_model,
+                                                 lora_config)
+
+    def gradient_checkpointing_enable(self):
+        self.activation_checkpointing_enable()
+
+    def activation_checkpointing_enable(self):
+        self.model.language_model.gradient_checkpointing_enable()
+
+    def gradient_checkpointing_disable(self):
+        self.activation_checkpointing_disable()
+
+    def activation_checkpointing_disable(self):
+        self.model.language_model.gradient_checkpointing_disable()
+
+    def state_dict(self, *args, **kwargs):
+        state_dict = super().state_dict(*args, **kwargs)
+        to_return = OrderedDict()
+        # Step 1. visual_encoder
+        if self.use_visual_encoder_lora:
+            to_return.update(
+                get_peft_model_state_dict(
+                    self.model.vision_model, state_dict=state_dict))
+        elif not self.freeze_visual_encoder:
+            to_return.update({
+                k: v
+                for k, v in state_dict.items() if 'model.vision_model.' in k
+            })
+        # Step 2. LLM
+        if self.use_llm_lora:
+            to_return.update(
+                get_peft_model_state_dict(
+                    self.model.language_model, state_dict=state_dict))
+        elif not self.freeze_llm:
+            to_return.update({
+                k: v
+                for k, v in state_dict.items() if 'model.language_model.' in k
+            })
+        # Step 3. Projector
+        to_return.update(
+            {k: v
+             for k, v in state_dict.items() if 'model.mlp1.' in k})
+        return to_return
+
+    def init_weights(self):
+        pass
+
+    def forward(self, data, data_samples=None, mode='loss'):
+        pixel_values = data['pixel_values']
+
+        if type(pixel_values) is list or pixel_values.ndim == 5:
+            if type(pixel_values) is list:
+                pixel_values = [
+                    x.unsqueeze(0) if x.ndim == 3 else x for x in pixel_values
+                ]
+            # b*n, c, h, w
+            concat_images = torch.cat([
+                image.to(self.model.vision_model.dtype)
+                for image in pixel_values
+            ],
+                                      dim=0)
+        else:
+            raise NotImplementedError()
+
+        input_ids = data['input_ids']
+        position_ids = data['position_ids']
+        attention_mask = data['attention_mask']
+        # sum is 0 are text
+        image_flags = torch.sum(concat_images, dim=(1, 2, 3)) != 0
+        image_flags = image_flags.long()
+
+        labels = data['labels']
+        use_cache = False
+
+        # Directly calling this code in LORA fine-tuning
+        # will result in an error,so we must rewrite it.
+        # TODO: Once the official is fixed, we can remove it.
+        # outputs = self.model(input_ids=input_ids,
+        #                      position_ids=position_ids,
+        #                      attention_mask=attention_mask,
+        #                      image_flags=image_flags,
+        #                      pixel_values=concat_images,
+        #                      labels=labels,
+        #                      use_cache=use_cache)
+        outputs = self._llm_forward(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            image_flags=image_flags,
+            pixel_values=concat_images,
+            labels=labels,
+            use_cache=use_cache)
+        loss_dict = {'loss': outputs.loss}
+        return loss_dict
+
+    def _llm_forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        image_flags: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        return_dict = return_dict if return_dict is not None \
+            else self.model.config.use_return_dict
+
+        image_flags = image_flags.squeeze(-1)
+        # We only added the clone code here to avoid the error.
+        input_embeds = self.model.language_model.get_input_embeddings()(
+            input_ids).clone()
+
+        vit_embeds = self.model.extract_feature(pixel_values)
+        vit_embeds = vit_embeds[image_flags == 1]
+        vit_batch_size = pixel_values.shape[0]
+
+        B, N, C = input_embeds.shape
+        input_embeds = input_embeds.reshape(B * N, C)
+
+        if torch.distributed.get_rank() == 0 and self._count % 100 == 0:
+            print(f'dynamic ViT batch size: {vit_batch_size}, '
+                  f'images per sample: {vit_batch_size / B}, '
+                  f'dynamic token length: {N}')
+        self._count += 1
+
+        input_ids = input_ids.reshape(B * N)
+        selected = (input_ids == self.model.img_context_token_id)
+        try:
+            input_embeds[
+                selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(
+                    -1, C)
+        except Exception as e:
+            vit_embeds = vit_embeds.reshape(-1, C)
+            print(f'warning: {e}, input_embeds[selected].shape='
+                  f'{input_embeds[selected].shape}, '
+                  f'vit_embeds.shape={vit_embeds.shape}')
+            n_token = selected.sum()
+            input_embeds[
+                selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]
+
+        input_embeds = input_embeds.reshape(B, N, C)
+
+        outputs = self.model.language_model(
+            inputs_embeds=input_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        logits = outputs.logits
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(
+                -1, self.model.language_model.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits, ) + outputs[1:]
+            return (loss, ) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

From 9444b3423d3f9b0a409e879eb24d06a4a190a9b5 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Mon, 22 Jul 2024 14:46:06 +0800
Subject: [PATCH 23/29] [Bug] fix preference_collate_fn attn_mask (#859)

fix preference_collate_fn attn_mask
---
 .../dataset/collate_fns/preference_collate_fn.py  | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/xtuner/dataset/collate_fns/preference_collate_fn.py b/xtuner/dataset/collate_fns/preference_collate_fn.py
index ca21613bb..4b6a7f5c3 100644
--- a/xtuner/dataset/collate_fns/preference_collate_fn.py
+++ b/xtuner/dataset/collate_fns/preference_collate_fn.py
@@ -58,14 +58,14 @@ def preference_collate_fn(instances: Sequence[Dict],
         labels = torch.stack(labels)
 
     if use_varlen_attn:
-        attention_mask = torch.ones_like(input_ids).bool()
+        attention_mask = None
         position_ids = torch.stack(position_ids, dim=0)
     else:
         # Some tokenizers have the same eos token and pad token, so input_ids
         # cannot be masked directly based on the pad token id.
         attention_mask = torch.zeros_like(input_ids).bool()
-        for i in ori_length:
-            attention_mask[:i] = True
+        for i, length in enumerate(ori_length):
+            attention_mask[i, :length] = True
 
         bs, seq_len = input_ids.shape
         position_ids = torch.arange(seq_len).unsqueeze(0).long().repeat(bs, 1)
@@ -74,11 +74,12 @@ def preference_collate_fn(instances: Sequence[Dict],
         input_ids = pad_for_sequence_parallel(input_ids, pad_index)
         labels = pad_for_sequence_parallel(labels, IGNORE_INDEX)
         position_ids = pad_for_sequence_parallel(position_ids, 0)
-        # We use attention_mask to distinguish `input_ids` from
-        # (sequence parallel) pad tokens in `get_var_len_atten_logps` method of
-        # class `DPO` and `ORPO`
-        attention_mask = pad_for_sequence_parallel(attention_mask, 0)
+        if attention_mask is not None:
+            attention_mask = pad_for_sequence_parallel(attention_mask, 0)
         if use_varlen_attn:
+            # We use attention_mask to distinguish `input_ids` from
+            # (sequence parallel) pad tokens in `get_var_len_atten_logps`
+            # method of class `DPO` and `ORPO`
             (cumulative_len, attention_mask
              ) = pad_cumulative_len_for_sequence_parallel(cumulative_len)
 

From 94a4fcb0f53d63ceb85334bb456c073251669ccb Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Mon, 22 Jul 2024 20:17:44 +0800
Subject: [PATCH 24/29] bump version to 0.1.23 (#862)

---
 xtuner/version.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/xtuner/version.py b/xtuner/version.py
index c77f4bf8f..e4669c188 100644
--- a/xtuner/version.py
+++ b/xtuner/version.py
@@ -1,5 +1,5 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-__version__ = '0.1.22'
+__version__ = '0.1.23'
 short_version = __version__
 
 

From d2a173a284d5969e80669c6e59dffd055d2570f3 Mon Sep 17 00:00:00 2001
From: LDLINGLINGLING <47373076+LDLINGLINGLING@users.noreply.github.com>
Date: Mon, 29 Jul 2024 16:29:27 +0800
Subject: [PATCH 25/29] =?UTF-8?q?readme=E4=B8=AD=E5=A2=9E=E5=8A=A0?=
 =?UTF-8?q?=E4=BA=86MiniCPM=E7=9A=84=E6=94=AF=E6=8C=81=20(#869)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-authored-by: liudan <liudan@MacBook-Pro.local>
---
 README.md       | 3 ++-
 README_zh-CN.md | 3 ++-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 2a8eb2879..263d300c7 100644
--- a/README.md
+++ b/README.md
@@ -38,7 +38,7 @@ English | [简体中文](README_zh-CN.md)
 </div>
 
 ## 🎉 News
-
+- **\[2024/07\]** Support [MiniCPM](xtuner/configs/minicpm/) models!
 - **\[2024/07\]** Support [DPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/dpo), [ORPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/orpo) and [Reward Model](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/reward_model) training with packed data and sequence parallel! See [documents](https://xtuner.readthedocs.io/en/latest/dpo/overview.html) for more details.
 - **\[2024/07\]** Support [InternLM 2.5](xtuner/configs/internlm/internlm2_5_chat_7b/) models!
 - **\[2024/06\]** Support [DeepSeek V2](xtuner/configs/deepseek/deepseek_v2_chat/) models! **2x faster!**
@@ -113,6 +113,7 @@ XTuner is an efficient, flexible and full-featured toolkit for fine-tuning large
   <li><a href="https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1">Mixtral</a></li>
   <li><a href="https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat">DeepSeek V2</a></li>
   <li><a href="https://huggingface.co/google">Gemma</a></li>
+  <li><a href="https://huggingface.co/openbmb">MiniCPM</a></li>
   <li>...</li>
 </ul>
 </td>
diff --git a/README_zh-CN.md b/README_zh-CN.md
index 58076210f..f4f0b4b48 100644
--- a/README_zh-CN.md
+++ b/README_zh-CN.md
@@ -38,7 +38,7 @@
 </div>
 
 ## 🎉 更新
-
+- **\[2024/07\]** 支持 [MiniCPM](xtuner/configs/minicpm/) 模型!
 - **\[2024/07\]** 支持训练 [DPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/dpo)， [ORPO](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/orpo) 还有 [Reward Model](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/reward_model) ! 并且能够支持打包数据以及序列并行功能！ 请参考 [文档](https://xtuner.readthedocs.io/zh-cn/latest/dpo/overview.html) 了解更多信息。
 - **\[2024/07\]** 支持 [InternLM 2.5](xtuner/configs/internlm/internlm2_5_chat_7b/) 模型!
 - **\[2024/06\]** 支持 [DeepSeek V2](xtuner/configs/deepseek/deepseek_v2_chat/) models! **训练速度提升一倍！**
@@ -113,6 +113,7 @@ XTuner 是一个高效、灵活、全能的轻量化大模型微调工具库。
   <li><a href="https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1">Mixtral</a></li>
   <li><a href="https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat">DeepSeek V2</a></li>
   <li><a href="https://huggingface.co/google">Gemma</a></li>
+  <li><a href="https://huggingface.co/openbmb">MiniCPM</a></li>
   <li>...</li>
 </ul>
 </td>

From 01640b00790453de39a63b5019fdd14f7cc6e4d8 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Wed, 31 Jul 2024 18:54:56 +0800
Subject: [PATCH 26/29] [Bug] fix dsv2 attn dispatch (softmax_scale) (#873)

fix dsv2 attn dispatch (softmax_scale)
---
 xtuner/model/modules/dispatch/attention.py   | 4 ++++
 xtuner/model/modules/dispatch/deepseek_v2.py | 2 ++
 2 files changed, 6 insertions(+)

diff --git a/xtuner/model/modules/dispatch/attention.py b/xtuner/model/modules/dispatch/attention.py
index 2a225795d..e89bb511c 100644
--- a/xtuner/model/modules/dispatch/attention.py
+++ b/xtuner/model/modules/dispatch/attention.py
@@ -38,6 +38,7 @@ def flash_attn_w_mask(
         key_states,
         value_states,
         attention_mask,
+        softmax_scale=None,
         causal=True,
         dropout_p=0.0,
         window_size=(-1, -1),  # -1 means infinite context window
@@ -57,6 +58,7 @@ def flash_attn_w_mask(
         cu_seqlens_k=cu_seqlens_k,
         max_seqlen_q=max_seqlen_in_batch_q,
         max_seqlen_k=max_seqlen_in_batch_k,
+        softmax_scale=softmax_scale,
         dropout_p=dropout_p,
         causal=causal,
         window_size=window_size)
@@ -71,6 +73,7 @@ def varlen_flash_attn(
         value_states,
         cumulative_len,
         max_seqlen,
+        softmax_scale=None,
         dropout_p=0.,
         causal=True,
         window_size=(-1, -1),  # -1 means infinite context window
@@ -85,6 +88,7 @@ def varlen_flash_attn(
         cumulative_len,
         max_seqlen,
         max_seqlen,
+        softmax_scale=softmax_scale,
         dropout_p=dropout_p,
         return_attn_probs=False,
         causal=causal,
diff --git a/xtuner/model/modules/dispatch/deepseek_v2.py b/xtuner/model/modules/dispatch/deepseek_v2.py
index 667d2227c..bfa3ebb6d 100644
--- a/xtuner/model/modules/dispatch/deepseek_v2.py
+++ b/xtuner/model/modules/dispatch/deepseek_v2.py
@@ -279,6 +279,7 @@ def deepseek_varlen_attn_forward(
             value_states,
             cumulative_len,
             max_seqlen,
+            softmax_scale=self.softmax_scale,
             causal=causal,
             dropout_p=dropout_rate,
             training=True)
@@ -287,6 +288,7 @@ def deepseek_varlen_attn_forward(
             query_states,
             key_states,
             value_states,
+            softmax_scale=self.softmax_scale,
             causal=causal,
             dropout_p=dropout_rate,
             training=False)

From 7dd779b3b4c5a7fb8e00f7511c1635b8ff33979b Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 9 Aug 2024 17:59:54 +0800
Subject: [PATCH 27/29] [Bug] fix openai_map_fn bugs (#885)

fix openai_map_fn bugs
---
 xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py b/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py
index c1798dc45..468e738f7 100644
--- a/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py
+++ b/xtuner/dataset/map_fns/dataset_map_fns/openai_map_fn.py
@@ -32,7 +32,8 @@ def openai_map_fn(example):
         elif msg['role'] == 'user':
             input += msg['content']
         elif msg['role'] == 'assistant':
-            output_with_loss = msg.get('loss', True)
+            output_with_loss = msg.get('loss', 'True')
+            output_with_loss = str(output_with_loss)
             output_with_loss = output_with_loss.lower() == 'true'
             conversation.append({
                 'system': system,

From d81b366d5bd5795d5892b0c9ae42ef229ac719f4 Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 9 Aug 2024 18:00:34 +0800
Subject: [PATCH 28/29] support transformers >= 4.43 (#878)

---
 xtuner/model/modules/dispatch/cohere.py  | 36 ++++++++++++++----
 xtuner/model/modules/dispatch/mistral.py | 40 +++++++++++++++-----
 xtuner/model/modules/dispatch/phi3.py    | 41 ++++++++++++++++-----
 xtuner/model/modules/dispatch/qwen2.py   | 47 +++++++++++++++++++-----
 4 files changed, 129 insertions(+), 35 deletions(-)

diff --git a/xtuner/model/modules/dispatch/cohere.py b/xtuner/model/modules/dispatch/cohere.py
index d3529f570..8acf06747 100644
--- a/xtuner/model/modules/dispatch/cohere.py
+++ b/xtuner/model/modules/dispatch/cohere.py
@@ -3,6 +3,8 @@
 
 import torch
 import torch.distributed as dist
+import transformers
+from mmengine.utils import digit_version
 from transformers.models.cohere.modeling_cohere import apply_rotary_pos_emb
 
 from xtuner.parallel.sequence import get_sequence_parallel_world_size
@@ -18,6 +20,14 @@ class Cache:
         pass
 
 
+TRANSFORMERS_VERSION = digit_version(transformers.__version__)
+IS_LOW_VERSION_TRANSFORMERS = TRANSFORMERS_VERSION < digit_version('4.43')
+
+if not IS_LOW_VERSION_TRANSFORMERS:
+    from transformers.modeling_flash_attention_utils import \
+        _flash_attention_forward
+
+
 def cohere_attn_forward(
     self,
     hidden_states: torch.Tensor,
@@ -110,13 +120,25 @@ def cohere_attn_forward(
         ori_num_head = self.num_heads
         self.num_heads = query_states.shape[-2]
 
-    attn_output = self._flash_attention_forward(
-        query_states,
-        key_states,
-        value_states,
-        attention_mask,
-        query_states.shape[1],
-        dropout=dropout_rate)
+    if IS_LOW_VERSION_TRANSFORMERS:
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.shape[1],
+            dropout=dropout_rate)
+    else:
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.shape[1],
+            dropout=dropout_rate,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=self.is_causal,
+        )
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
diff --git a/xtuner/model/modules/dispatch/mistral.py b/xtuner/model/modules/dispatch/mistral.py
index d08b0f00e..dc6c7fed8 100644
--- a/xtuner/model/modules/dispatch/mistral.py
+++ b/xtuner/model/modules/dispatch/mistral.py
@@ -6,7 +6,9 @@
 import torch
 import torch.distributed as dist
 import torch.nn as nn
+import transformers
 from mmengine import MessageHub
+from mmengine.utils import digit_version
 from transformers.cache_utils import Cache
 from transformers.models.mistral.modeling_mistral import (apply_rotary_pos_emb,
                                                           repeat_kv)
@@ -28,6 +30,13 @@
 except ImportError:
     pass
 
+TRANSFORMERS_VERSION = digit_version(transformers.__version__)
+IS_LOW_VERSION_TRANSFORMERS = TRANSFORMERS_VERSION < digit_version('4.43')
+
+if not IS_LOW_VERSION_TRANSFORMERS:
+    from transformers.modeling_flash_attention_utils import \
+        _flash_attention_forward
+
 
 class MistralRotaryEmbedding(nn.Module):
 
@@ -220,15 +229,28 @@ def mistral_attn_forward(
         ori_num_head = self.num_heads
         self.num_heads = query_states.shape[-2]
 
-    attn_output = self._flash_attention_forward(
-        query_states,
-        key_states,
-        value_states,
-        attention_mask,
-        query_length=query_states.shape[1],
-        dropout=dropout_rate,
-        use_sliding_windows=use_sliding_windows,
-    )
+    if IS_LOW_VERSION_TRANSFORMERS:
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_length=query_states.shape[1],
+            dropout=dropout_rate,
+            use_sliding_windows=use_sliding_windows,
+        )
+    else:
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.shape[1],
+            dropout=dropout_rate,
+            sliding_window=getattr(self.config, 'sliding_window', None),
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=self.is_causal,
+        )
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)
diff --git a/xtuner/model/modules/dispatch/phi3.py b/xtuner/model/modules/dispatch/phi3.py
index 97ebc8d33..10f60f939 100644
--- a/xtuner/model/modules/dispatch/phi3.py
+++ b/xtuner/model/modules/dispatch/phi3.py
@@ -1,10 +1,13 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+import inspect
 import warnings
 from typing import Optional, Tuple
 
 import torch
 import torch.distributed as dist
+import transformers
 from mmengine import MessageHub
+from mmengine.utils import digit_version
 
 from xtuner.parallel.sequence import (get_sequence_parallel_world_size,
                                       post_process_for_sequence_parallel_attn,
@@ -19,7 +22,12 @@ class Cache:
         pass
 
 
-import inspect
+TRANSFORMERS_VERSION = digit_version(transformers.__version__)
+IS_LOW_VERSION_TRANSFORMERS = TRANSFORMERS_VERSION < digit_version('4.43')
+
+if not IS_LOW_VERSION_TRANSFORMERS:
+    from transformers.modeling_flash_attention_utils import \
+        _flash_attention_forward
 
 _flash_supports_window_size = False
 try:
@@ -239,15 +247,28 @@ def phi3_attn_forward(
         ori_num_head = self.num_heads
         self.num_heads = query_states.shape[-2]
 
-    attn_output = self._flash_attention_forward(
-        query_states,
-        key_states,
-        value_states,
-        attention_mask,
-        query_states.shape[1],
-        dropout=attn_dropout,
-        use_sliding_windows=use_sliding_windows,
-    )
+    if IS_LOW_VERSION_TRANSFORMERS:
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.shape[1],
+            dropout=attn_dropout,
+            use_sliding_windows=use_sliding_windows,
+        )
+    else:
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.shape[1],
+            dropout=attn_dropout,
+            sliding_window=getattr(self.config, 'sliding_window', None),
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=self.is_causal,
+        )
 
     if enable_sequence_parallel:
         # (b, s, nd // sp_world_size, dim) -> (b, s // sp_world_size, nd, dim)
diff --git a/xtuner/model/modules/dispatch/qwen2.py b/xtuner/model/modules/dispatch/qwen2.py
index 1c8c5a8d0..20f2f40f3 100644
--- a/xtuner/model/modules/dispatch/qwen2.py
+++ b/xtuner/model/modules/dispatch/qwen2.py
@@ -5,7 +5,9 @@
 
 import torch
 import torch.distributed as dist
+import transformers
 from mmengine import MessageHub
+from mmengine.utils import digit_version
 from transformers.cache_utils import Cache
 from transformers.models.qwen2.modeling_qwen2 import (apply_rotary_pos_emb,
                                                       repeat_kv)
@@ -26,6 +28,13 @@
 except ImportError:
     pass
 
+TRANSFORMERS_VERSION = digit_version(transformers.__version__)
+IS_LOW_VERSION_TRANSFORMERS = TRANSFORMERS_VERSION < digit_version('4.43')
+
+if not IS_LOW_VERSION_TRANSFORMERS:
+    from transformers.modeling_flash_attention_utils import \
+        _flash_attention_forward
+
 
 def qwen2_attn_forward(
     self,
@@ -157,15 +166,35 @@ def qwen2_attn_forward(
         ori_num_head = self.num_heads
         self.num_heads = query_states.shape[-2]
 
-    attn_output = self._flash_attention_forward(
-        query_states,
-        key_states,
-        value_states,
-        attention_mask,
-        query_length=query_states.shape[1],
-        dropout=dropout_rate,
-        use_sliding_windows=use_sliding_windows,
-    )
+    if IS_LOW_VERSION_TRANSFORMERS:
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_length=query_states.shape[1],
+            dropout=dropout_rate,
+            use_sliding_windows=use_sliding_windows,
+        )
+    else:
+        if (self.config.use_sliding_window
+                and getattr(self.config, 'sliding_window', None) is not None
+                and self.layer_idx >= self.config.max_window_layers):
+            # There may be bugs here, but we are aligned with Transformers
+            sliding_window = self.config.sliding_window
+        else:
+            sliding_window = None
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.shape[1],
+            dropout=dropout_rate,
+            sliding_window=sliding_window,
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
 
     if enable_sequence_parallel:
         attn_output = post_process_for_sequence_parallel_attn(attn_output)

From 081c8ca874bdbf7a7f8cd0a9e4cba503eaaa7bba Mon Sep 17 00:00:00 2001
From: whcao <41630003+HIT-cwh@users.noreply.github.com>
Date: Fri, 9 Aug 2024 18:01:01 +0800
Subject: [PATCH 29/29] Add internlm2 5 cfgs (#872)

* add internlm2.5 configs

* limit transformers <= 4.42.4
---
 requirements/runtime.txt                      |   4 +-
 .../internlm2_5_chat_20b_alpaca_e3.py         | 202 ++++++++++++++++
 .../internlm2_5_chat_20b_qlora_alpaca_e3.py   | 219 ++++++++++++++++++
 3 files changed, 424 insertions(+), 1 deletion(-)
 create mode 100644 xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_alpaca_e3.py
 create mode 100644 xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_qlora_alpaca_e3.py

diff --git a/requirements/runtime.txt b/requirements/runtime.txt
index c93ea248b..3a4d2f84e 100644
--- a/requirements/runtime.txt
+++ b/requirements/runtime.txt
@@ -21,5 +21,7 @@ torchvision
 # Registering a causal mask in `LlamaModel` is not friendly for very large
 # `max_position_embeddings`. Refer to
 # https://github.com/huggingface/transformers/blob/v4.38.0/src/transformers/models/llama/modeling_llama.py#L921-L923
-transformers>=4.36.0,!=4.38.0,!=4.38.1,!=4.38.2
+# transformers >= 4.43.0 use _flash_attention_forward but not self._flash_attention_forward
+# to calculate attn output which lead to bc braeking
+transformers>=4.36.0,!=4.38.0,!=4.38.1,!=4.38.2,<=4.42.4
 transformers_stream_generator
diff --git a/xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_alpaca_e3.py b/xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_alpaca_e3.py
new file mode 100644
index 000000000..f67fc1a22
--- /dev/null
+++ b/xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_alpaca_e3.py
@@ -0,0 +1,202 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2_5-20b-chat'
+use_varlen_attn = False
+
+# Data
+alpaca_en_path = 'tatsu-lab/alpaca'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_en,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)
diff --git a/xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_qlora_alpaca_e3.py b/xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_qlora_alpaca_e3.py
new file mode 100644
index 000000000..f695e7922
--- /dev/null
+++ b/xtuner/configs/internlm/internlm2_5_chat_20b/internlm2_5_chat_20b_qlora_alpaca_e3.py
@@ -0,0 +1,219 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from datasets import load_dataset
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from peft import LoraConfig
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from xtuner.dataset import process_hf_dataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
+                                 VarlenAttnArgsToMessageHubHook)
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import SupervisedFinetune
+from xtuner.parallel.sequence import SequenceParallelSampler
+from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+pretrained_model_name_or_path = 'internlm/internlm2_5-20b-chat'
+use_varlen_attn = False
+
+# Data
+alpaca_en_path = 'tatsu-lab/alpaca'
+prompt_template = PROMPT_TEMPLATE.internlm2_chat
+max_length = 2048
+pack_to_max_length = True
+
+# parallel
+sequence_parallel_size = 1
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1
+accumulative_counts *= sequence_parallel_size
+dataloader_num_workers = 0
+max_epochs = 3
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 500
+SYSTEM = SYSTEM_TEMPLATE.alpaca
+evaluation_inputs = [
+    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
+]
+
+#######################################################################
+#                      PART 2  Model & Tokenizer                      #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=pretrained_model_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=SupervisedFinetune,
+    use_varlen_attn=use_varlen_attn,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=pretrained_model_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16,
+        quantization_config=dict(
+            type=BitsAndBytesConfig,
+            load_in_4bit=True,
+            load_in_8bit=False,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4')),
+    lora=dict(
+        type=LoraConfig,
+        r=64,
+        lora_alpha=16,
+        lora_dropout=0.1,
+        bias='none',
+        task_type='CAUSAL_LM'))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+alpaca_en = dict(
+    type=process_hf_dataset,
+    dataset=dict(type=load_dataset, path=alpaca_en_path),
+    tokenizer=tokenizer,
+    max_length=max_length,
+    dataset_map_fn=alpaca_map_fn,
+    template_map_fn=dict(
+        type=template_map_fn_factory, template=prompt_template),
+    remove_unused_columns=True,
+    shuffle_before_pack=True,
+    pack_to_max_length=pack_to_max_length,
+    use_varlen_attn=use_varlen_attn)
+
+sampler = SequenceParallelSampler \
+    if sequence_parallel_size > 1 else DefaultSampler
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=alpaca_en,
+    sampler=dict(type=sampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        system=SYSTEM,
+        prompt_template=prompt_template)
+]
+
+if use_varlen_attn:
+    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)