Shape error when using torchtune.modules.RotaryPositionalEmbeddings

[Leo-Lifeblood]

🐛 Describe the bug

When using the position encoding layer strange shape errors occur I dont have the time or insight to resolve

import torch
import torchtune

#max_value = max(tokenizer_causal.vocab.values()) + 1

max_value = 50

class causallm(torch.nn.Module):
def init(self, d_model, num_heads, d_ff, num_layers):
super().init()

    self.embeddings = torch.nn.Embedding(max_value, d_model)
    self.pos_embeddings = torchtune.modules.RotaryPositionalEmbeddings(d_model)

    self.encoder_layer = torch.nn.TransformerEncoderLayer(d_model, num_heads, d_ff, activation=torch.nn.GELU(), batch_first=True)

    self.transformer_encoder = torch.nn.TransformerEncoder(self.encoder_layer, num_layers=num_layers)

    self.output = torch.nn.Linear(d_model, max_value)

    self.num_heads = num_heads
    self.head_dim = d_model // num_heads

def forward(self, x, attention_mask=None):
    x = self.embeddings(x)
    
    #seq_len = x.shape[1]
    #batch_size = x.shape[0]
    #x = x.view(batch_size, seq_len, self.num_heads, self.head_dim)
    x = self.pos_embeddings(x)


    x = self.transformer_encoder(x, src_mask=attention_mask, is_causal=True)

    x = self.output(x)

    return x

model = causallm(d_model=512, num_heads=8, d_ff=2048, num_layers=2)

input_ids = torch.randint(0, max_value, (32, 128)) # input tensor with batch size 32 and sequence length 128
attention_mask = torch.ones((32, 128), dtype=torch.bool) # attention mask

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)
input_ids = input_ids.to(device)
attention_mask = attention_mask.to(device)

outputs = model(input_ids, attention_mask=attention_mask)

---

RuntimeError Traceback (most recent call last)
in <cell line: 11>()
9
10 # Forward pass
---> 11 outputs = model(input_ids, attention_mask=attention_mask)

5 frames
/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py in _wrapped_call_impl(self, *args, **kwargs)
1530 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1531 else:
-> 1532 return self._call_impl(*args, **kwargs)
1533
1534 def _call_impl(self, *args, **kwargs):

/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py in _call_impl(self, *args, **kwargs)
1539 or _global_backward_pre_hooks or _global_backward_hooks
1540 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1541 return forward_call(*args, **kwargs)
1542
1543 try:

in forward(self, x, attention_mask)
23 #batch_size = x.shape[0]
24 #x = x.view(batch_size, seq_len, self.num_heads, self.head_dim)
---> 25 x = self.pos_embeddings(x)
26
27

/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py in _wrapped_call_impl(self, *args, **kwargs)
1530 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1531 else:
-> 1532 return self._call_impl(*args, **kwargs)
1533
1534 def _call_impl(self, *args, **kwargs):

/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py in _call_impl(self, *args, **kwargs)
1539 or _global_backward_pre_hooks or _global_backward_hooks
1540 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1541 return forward_call(*args, **kwargs)
1542
1543 try:

/usr/local/lib/python3.10/dist-packages/torchtune/modules/position_embeddings.py in forward(self, x, input_pos)
109 # reshape the cache for broadcasting
110 # tensor has shape [1, s, 1, n_d // 2, 2]
--> 111 rope_cache = rope_cache.view(1, xshaped.size(1), 1, xshaped.size(3), 2)
112
113 # tensor has shape [b, s, n_h, n_d // 2, 2]

RuntimeError: shape '[1, 128, 1, 2, 2]' is invalid for input of size 65536

Versions

PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.27.9
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.1.85+-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.6
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 2
On-line CPU(s) list: 0,1
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) CPU @ 2.20GHz
CPU family: 6
Model: 79
Thread(s) per core: 2
Core(s) per socket: 1
Socket(s): 1
Stepping: 0
BogoMIPS: 4399.99
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm rdseed adx smap xsaveopt arat md_clear arch_capabilities
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 32 KiB (1 instance)
L1i cache: 32 KiB (1 instance)
L2 cache: 256 KiB (1 instance)
L3 cache: 55 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0,1
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Mitigation; PTE Inversion
Vulnerability Mds: Vulnerable; SMT Host state unknown
Vulnerability Meltdown: Vulnerable
Vulnerability Mmio stale data: Vulnerable
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Vulnerable: __user pointer sanitization and usercopy barriers only; no swapgs barriers
Vulnerability Spectre v2: Vulnerable; IBPB: disabled; STIBP: disabled; PBRSB-eIBRS: Not affected; BHI: Vulnerable (Syscall hardening enabled)
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Vulnerable

Versions of relevant libraries:
[pip3] numpy==1.25.2
[pip3] torch==2.3.0+cu121
[pip3] torchao==0.1
[pip3] torchaudio==2.3.0+cu121
[pip3] torchsummary==1.5.1
[pip3] torchtext==0.18.0
[pip3] torchtune==0.1.1
[pip3] torchvision==0.18.0+cu121
[pip3] triton==2.3.0
[conda] Could not collect

[awgu]

This might be better filed in the torchtune repo?

[ebsmothers]

Hi @Leo-Lifeblood thanks for creating the issue. I believe you are seeing this error because you're using the RotaryPositionalEmbeddings class with an input tensor shape that doesn't line up with what's expected. Typically these are used inside of self-attention, where the tensor shape is (batch_size, seq_len, num_heads, head_dim) (ref). But here it seems like you are using it on the outputs of self.embeddings, which should be a 2D tensor based on your inputs. Depending on what you're trying to do, you can consider directly importing one of our model builder functions, e.g. llama3 will allow you to pass a basic set of params and will give you back an nn.Module that's somewhat similar to what you have here. If you're looking to do something a bit more custom, happy to provide some pointers on how you can achieve that as well.

[iankur]

@ebsmothers why do we apply rope on the expanded key tensor? it seems wasteful as rope is applied on head_dim anyway.

[Leo-Lifeblood]

Hi @Leo-Lifeblood thanks for creating the issue. I believe you are seeing this error because you're using the RotaryPositionalEmbeddings class with an input tensor shape that doesn't line up with what's expected. Typically these are used inside of self-attention, where the tensor shape is (batch_size, seq_len, num_heads, head_dim) (ref). But here it seems like you are using it on the outputs of self.embeddings, which should be a 2D tensor based on your inputs. Depending on what you're trying to do, you can consider directly importing one of our model builder functions, e.g. llama3 will allow you to pass a basic set of params and will give you back an nn.Module that's somewhat similar to what you have here. If you're looking to do something a bit more custom, happy to provide some pointers on how you can achieve that as well.

rope = torchtune.modules.RotaryPositionalEmbeddings(32)
rope(torch.rand(32, 10, 4 ,8))

RuntimeError Traceback (most recent call last)
Cell In[106], line 1
----> 1 rope(torch.rand(32, 10, 4 ,8))

File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
1551 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1552 else:
-> 1553 return self._call_impl(*args, **kwargs)

File /opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1562, in Module._call_impl(self, *args, **kwargs)
1557 # If we don't have any hooks, we want to skip the rest of the logic in
1558 # this function, and just call forward.
1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1560 or _global_backward_pre_hooks or _global_backward_hooks
1561 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1562 return forward_call(*args, **kwargs)
1564 try:
1565 result = None

File /opt/conda/lib/python3.10/site-packages/torchtune/modules/position_embeddings.py:117, in RotaryPositionalEmbeddings.forward(self, x, input_pos)
112 rope_cache = rope_cache.view(-1, xshaped.size(1), 1, xshaped.size(3), 2)
114 # tensor has shape [b, s, n_h, h_d // 2, 2]
115 x_out = torch.stack(
116 [
--> 117 xshaped[..., 0] * rope_cache[..., 0]
118 - xshaped[..., 1] * rope_cache[..., 1],
119 xshaped[..., 1] * rope_cache[..., 0]
120 + xshaped[..., 0] * rope_cache[..., 1],
121 ],
122 -1,
123 )
125 # tensor has shape [b, s, n_h, h_d]
126 x_out = x_out.flatten(3)

RuntimeError: The size of tensor a (32) must match the size of tensor b (4) at non-singleton dimension 0

[Leo-Lifeblood]

the rope implementation somehow ends up with 1/8th the required batch dimension:
ope.cache.shape
torch.Size([4096, 16, 2])
add Codeadd Markdown
10
rope_cache = rope.cache[:10]
add Codeadd Markdown
torch.rand(32, 10, 4 ,8).reshape(*torch.rand(32, 10, 4 ,8).shape[:-1], -1, 2).shape
torch.Size([32, 10, 4, 4, 2])
add Codeadd Markdown
arrow_upwardarrow_downwarddelete
.shape
rope_cache.view(-1, 10, 1, 4, 2).shape
torch.Size([4, 10, 1, 4, 2])

[ebsmothers]

@iankur my sincere apologies, somehow I completely missed your previous comment. Probably too late now but this has actually recently been changed in #1558.

[ebsmothers]

@Leo-Lifeblood in your first comment:

rope = torchtune.modules.RotaryPositionalEmbeddings(32)
rope(torch.rand(32, 10, 4 ,8))

you are setting RoPE's dim to 32, while in your input tensor you have batch_size=32, seq_len=10, num_heads=4, head_dim=8. But as you can see here RoPE's dim should be the head_dim (i.e. 8, not 32 as you've set it).

I don't follow your second example. But I suspect that it's due to the same reason: if your RoPE dim is off by a factor of 8 from what's in your input data it makes sense that the inferred dimension from a view of the RoPE cache based on your input data would be off by a factor of 8 as well.

[Leo-Lifeblood]

Ok I have tried what you have suggested It has not worked though I have the code below and i'll try to explain whats wrong with it from my perspective:

import torch
import torch.functional as F
import torch.nn as nn
import numpy as np
import torchtune as tune

posenc = tune.modules.RotaryPositionalEmbeddings(64//4)

test = torch.randn(1, 1, 64)

num_heads = 4

batch_size, seq_len, hidden_dim = test.shape

test = test.view(batch_size, seq_len, -1, num_heads)
print(test.shape)

test = posenc(test)
print(test.shape)

test.view(batch_size, seq_len, hidden_dim)

From this code I get:

torch.Size([1, 1, 16, 4])
torch.Size([4, 1, 16, 4])
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
[<ipython-input-18-25a064cc2a57>](https://localhost:8080/#) in <cell line: 13>()
     11 print(test.shape)
     12 
---> 13 test.view(batch_size, seq_len, hidden_dim)

RuntimeError: shape '[1, 1, 64]' is invalid for input of size 256

In my understanding the batch size should not change here.

[ebsmothers]

@Leo-Lifeblood in your most recent example I think this line is not correct:

test = test.view(batch_size, seq_len, -1, num_heads)

For RoPE your input tensor should have shape (batch_size, seq_len, num_heads, head_dim), not (batch_size, seq_len, head_dim, num_heads) as you have here. See this comment in the code. So then when you apply RoPE with embed_dim=16 (i.e. head_dim=16) to this tensor it thinks that your head_dim is actually 4, since that's the last dimension of test. This explains why the total number of elements in the output is off by a factor of 4 (i.e. 16/4).