3 Releases published by 1 person

• v4.43.0 v4.43.0: Llama 3.1, Chameleon, ZoeDepth, Hiera

  published Jul 23, 2024

• v4.43.1 v4.43.1: Patch release

  published Jul 23, 2024

• v4.43.2 v4.43.2: Patch release

  published Jul 24, 2024

75 Pull requests merged by 44 people

• Flash-Attn: fix generation when no attention mask or no pading

  #32241 merged Jul 26, 2024

• [tests] fix static cache implementation is not compatible with attn_implementation==flash_attention_2

  #32039 merged Jul 26, 2024

• Add check for target_sizes is None in post_process_image_guided_detection for owlv2

  #31934 merged Jul 26, 2024

• Adds: extra_repr for RMSNorm layers in most models

  #32204 merged Jul 26, 2024

• Refactor: Removed un-necessary object base class

  #32230 merged Jul 26, 2024

• don't log base model architecture in wandb if log model is false

  #32143 merged Jul 26, 2024

• Resize embeds with DeepSpeed

  #32214 merged Jul 26, 2024

• Llava: generate without images

  #32183 merged Jul 26, 2024

• Generation: stop at eos for assisted decoding

  #31301 merged Jul 26, 2024

• Fix code snippet for Grounding DINO

  #32229 merged Jul 25, 2024

• Allow a specific microphone to be used by the ffmpeg audio pipeline utility functions. Default to using the currently active microphone on Mac

  #31846 merged Jul 25, 2024

• translate philosophy.md to chinese

  #32177 merged Jul 25, 2024

• Follow up for #31973

  #32025 merged Jul 25, 2024

• [warnings] fix E721 warnings

  #32223 merged Jul 25, 2024

• [BigBird Pegasus] set _supports_param_buffer_assignment to False

  #32222 merged Jul 25, 2024

• Update question_answering.py

  #32208 merged Jul 25, 2024

• remove unnecessary guard code related with pytorch versions 1.4.2 ~ 1.7.0

  #32210 merged Jul 25, 2024

• [whisper] fix short-form output type

  #32178 merged Jul 25, 2024

• fix: Replaced deprecated unittest method with the correct one

  #32198 merged Jul 24, 2024

• 🚨 No more default chat templates

  #31733 merged Jul 24, 2024

• Support dequantizing GGUF FP16 format

  #31783 merged Jul 24, 2024

• Fix float8_e4m3fn in modeling_utils

  #32193 merged Jul 24, 2024

• Fix resize embedding with Deepspeed

  #32192 merged Jul 24, 2024

• let's not warn when someone is running a forward

  #32176 merged Jul 24, 2024

• RoPE: relaxed rope validation

  #32182 merged Jul 24, 2024

• Remove conversational pipeline tests

  #32099 merged Jul 24, 2024

• Update qwen2.md

  #32108 merged Jul 24, 2024

• fix: default value reflects the runtime environment variables rather than the ones present at import time.

  #32153 merged Jul 24, 2024

• adds: extra_repr() to MambaRMSNorm to include hidden size / size of weights in the layer

  #32171 merged Jul 24, 2024

• [docs] change temperature to a positive value

  #32077 merged Jul 23, 2024

• fix: Fixed an if condition that is always evaluating to true

  #32160 merged Jul 23, 2024

• fix

  #32162 merged Jul 23, 2024

• Updated ruff to the latest version

  #31926 merged Jul 23, 2024

• Enhancing SFT Training Efficiency Using Packing and FlashAttention2 with Position IDs

  #31629 merged Jul 23, 2024

• Added additional kwarg for successful running of optuna hyperparameter search

  #31924 merged Jul 23, 2024

• feat(cache): StaticCache uses index_copy_ to avoid useless copy

  #31857 merged Jul 23, 2024

• Fix typing to be compatible with later py versions

  #32155 merged Jul 23, 2024

• Revert "Incorrect Whisper long-form decoding timestamps "

  #32148 merged Jul 23, 2024

• Rename Phi-3 rope scaling type

  #31436 merged Jul 23, 2024

• Added mamba.py backend

  #30139 merged Jul 23, 2024

• Fix video batching to videollava

  #32139 merged Jul 23, 2024

• Fix flash attention speed issue

  #32028 merged Jul 23, 2024

• gguf conversion add_prefix_space=None for llama3

  #31937 merged Jul 23, 2024

• Llama: RoPE refactor

  #32135 merged Jul 23, 2024

• Modify resize_token_embeddings to ensure output type is same as input

  #31979 merged Jul 23, 2024

• Disable quick init for TapasPreTrainedModel

  #32149 merged Jul 23, 2024

• Add YaRN and Dynamic-YaRN RoPE Scaling Methods

  #30910 merged Jul 23, 2024

• Add method to retrieve used chat template

  #32032 merged Jul 23, 2024

• Fix mask creations of GPTNeoX and GPT2

  #31944 merged Jul 23, 2024

• [modelling] remove un-necessary transpose for fa2 attention

  #31749 merged Jul 23, 2024

• Remove trust_remote_code when loading Libri Dummy

  #31748 merged Jul 23, 2024

• LLaVaNeXT: pad on right if training

  #32134 merged Jul 23, 2024

• Add llama3-llava-next-8b to llava_next conversion script

  #31395 merged Jul 23, 2024

• Add new quant method

  #32047 merged Jul 22, 2024

• set warning level to info for special tokens have been added

  #32138 merged Jul 22, 2024

• Don't default to other weights file when use_safetensors=True

  #31874 merged Jul 22, 2024

• Return assistant generated tokens mask in apply_chat_template

  #30650 merged Jul 22, 2024

• [RoBERTa] Minor clarifications to model doc

  #31949 merged Jul 22, 2024

• fix: Fixed raising TypeError instead of ValueError for invalid type

  #32111 merged Jul 22, 2024

• Update ko/_toctree.yml and remove custom_tools.md to reflect latest changes

  #31969 merged Jul 22, 2024

• Fix failing test with race condition

  #32140 merged Jul 22, 2024

• [generate] fix eos/pad id check on mps devices

  #31695 merged Jul 22, 2024

• Mention model_info.id instead of model_info.modelId

  #32106 merged Jul 22, 2024

• fix: Replaced deprecated mktemp() function

  #32123 merged Jul 22, 2024

• Generate: store special token tensors under a unique variable name

  #31980 merged Jul 22, 2024

• Fix shard order

  #32023 merged Jul 22, 2024

• Agents planning

  #31702 merged Jul 22, 2024

• Fix tests after huggingface_hub 0.24

  #32054 merged Jul 19, 2024

• Chameleon: not supported with fast load

  #32091 merged Jul 19, 2024

• Disable quick init for deepspeed

  #32066 merged Jul 19, 2024

• Support generating with fallback for short form audio in Whisper

  #30984 merged Jul 19, 2024

• Add image-text-to-text task guide

  #31777 merged Jul 19, 2024

• Fixes to chameleon docs

  #32078 merged Jul 19, 2024

• Fix progress callback deepcopy

  #32070 merged Jul 19, 2024

• VideoLLaVa: fix chat format in docs

  #32083 merged Jul 19, 2024

41 Pull requests opened by 35 people

• added warning to Trainer when label_names is not specified for PeftModel

  #32085 opened Jul 19, 2024

• fix: SeamlessM4TFeatureExtractor stride remainder

  #32088 opened Jul 19, 2024

• Add sdpa support for Albert

  #32092 opened Jul 19, 2024

• Fix `.push_to_hub(..., create_pr=True, revision="my-branch")` when creating PR on not-owned repo

  #32094 opened Jul 19, 2024

• Custom beam search scorer argument in generate function

  #32097 opened Jul 19, 2024

• Create django.yml

  #32100 opened Jul 19, 2024

• [wip][meta-llama][torch.compile] Fix issues with torch.compile

  #32102 opened Jul 19, 2024

• support 3D attention mask in bert

  #32105 opened Jul 20, 2024

• fix: resolve bug with `use_mps_device` setting not taking effect

  #32114 opened Jul 20, 2024

• Update stateful_callbacks state before saving checkpoint

  #32115 opened Jul 20, 2024

• Update benchmark.py-- Enhance Benchmarking with Multi-Commit Support …

  #32116 opened Jul 21, 2024

• [WIP] Add Depth Anything V2 Metric models

  #32126 opened Jul 21, 2024

• DINOv2 register support

  #32127 opened Jul 22, 2024

• [whisper] alternative fix for long-form timestamps

  #32131 opened Jul 22, 2024

• Add Qwen2-Audio

  #32137 opened Jul 22, 2024

• add scaling_factor to GemmaRotaryEmbedding for fix error in GemmaLine…

  #32141 opened Jul 22, 2024

• Add stream messages from agent run for gradio chatbot

  #32142 opened Jul 22, 2024

• Cache: create docs

  #32150 opened Jul 23, 2024

• [build-ci-image] add tiktoken

  #32152 opened Jul 23, 2024

• Added error when sequence length is bigger than max_position_embeddings

  #32156 opened Jul 23, 2024

• fix: (issue #32124) Exception raised when running `transformers/examples/flax/language-modeling/t5_tokenizer_model.py`.

  #32157 opened Jul 23, 2024

• Test

  #32158 opened Jul 23, 2024

• support copies

  #32159 opened Jul 23, 2024

• Add a static cache that offloads to the CPU or other device

  #32161 opened Jul 23, 2024

• Fixed Hybrid Cache Shape Initialization.

  #32163 opened Jul 23, 2024

• Make static cache compatible with torch.export

  #32168 opened Jul 23, 2024

• Uniformize kwargs for Layoutlm (2, 3, X) processors

  #32180 opened Jul 24, 2024

• Uniformize kwargs for chameleon processor

  #32181 opened Jul 24, 2024

• Gemma2 and flash-attention

  #32188 opened Jul 24, 2024

• [WIP] - Enable speculative decoding with batch size >1

  #32189 opened Jul 24, 2024

• Fix slow GemmaTokenizer and improve SPM slow -> fast conversion process

  #32191 opened Jul 24, 2024

• warning about weight_g/weight_v missing on WeightNorm on PyTorch

  #32194 opened Jul 24, 2024

• Whisper tokenizer word level timestamps

  #32197 opened Jul 24, 2024

• Bump torch from 1.13.1 to 2.2.0 in /examples/research_projects/visual_bert

  #32220 opened Jul 25, 2024

• Fix attention propagation for vision towers of llava-like models

  #32221 opened Jul 25, 2024

• >3-5x faster torch.compile forward compilation for autoregressive decoder models

  #32227 opened Jul 25, 2024

• [WIP] Add support for XTR

  #32231 opened Jul 25, 2024

• fix: warmup_steps check for training_args

  #32236 opened Jul 26, 2024

• VLMs: dispatch sdpa to each sub model

  #32238 opened Jul 26, 2024

• 🌐 [i18n-KO] Translated `image_feature_extraction.md` to Korean

  #32239 opened Jul 26, 2024

• #32184 save total_vocab_size

  #32240 opened Jul 26, 2024

61 Issues closed by 23 people

• Idefics2 generation erroring with flash_attention_2

  #32237 closed Jul 26, 2024

• `target_sizes` in Owlv2 `post_process_image_guided_detection`

  #31915 closed Jul 26, 2024

• Transformers 4.36 use_cache issue

  #28056 closed Jul 26, 2024

• Model saving when using Trainer with Accelerate

  #29792 closed Jul 26, 2024

• Running out of memory while finetuning and inferencing VideoMAE due to which script is being killed.

  #30939 closed Jul 26, 2024

• Can't create transformer pipeline because pytorch failed to be detected

  #31454 closed Jul 26, 2024

• cannot import name 'AutoModelForImageToImage' from 'transformers.models.auto.modeling_auto' (/opt/conda/lib/python3.10/site-packages/transformers/models/auto/modeling_auto.py)

  #31460 closed Jul 26, 2024

• ValueError: too many values to unpack (expected 2) when use glm4v or cogvlm2

  #32226 closed Jul 25, 2024

• [Whisper] Inconsistent return types for Whisper generation

  #32202 closed Jul 25, 2024

• 🐛 `attn_implementation="sdpa"` slower than `BetterTransformer.transform`?

  #31245 closed Jul 25, 2024

• Low retrieval and generation performance if evaluate rag model using consolidate_rag_checkpoint initialized with BART-LARGE as generator

  #31349 closed Jul 25, 2024

• pipeline gives a different result than the other approach in predicting word probability

  #31995 closed Jul 25, 2024

• NaNs when running bitsandbytes quantized Chameleon

  #32174 closed Jul 25, 2024

• Metrics - Pipeline

  #32190 closed Jul 24, 2024

• The features calculated by transformers dinov2 are different from the official ones

  #32175 closed Jul 24, 2024

• AttributeError: module 'torch' has no attribute 'float8_e4m3fn'

  #32185 closed Jul 24, 2024

• Llama 3 - RuntimeError: shape '[-1, 0]' is invalid for input of size 41041920

  #32170 closed Jul 24, 2024

• Backwards compatibility broken for RoPE: "rope_type"

  #32166 closed Jul 24, 2024

• KeyError: 'rope_type'

  #32167 closed Jul 24, 2024

• cannot import name 'Conversation' from 'transformers'

  #32096 closed Jul 24, 2024

• callback to implement how the predictions should be stored

  #32186 closed Jul 24, 2024

• Using `AutoTokenizer.from_pretrained`'s `.encode()` function fails to add BOS token for new Llama-3.1 model

  #32172 closed Jul 24, 2024

• phi-3's LlamaTokenizer ignores newline character.

  #32136 closed Jul 24, 2024

• Why MPS can never be used successfully?

  #32035 closed Jul 24, 2024

• Unable to load wavlm-large from pretrained in offline mode

  #32147 closed Jul 23, 2024

• Extra dataset features not passing to the custom collator

  #32093 closed Jul 23, 2024

• Allow additional keyword args to be passed to optuna hyperparameter search

  #31923 closed Jul 23, 2024

• The behavior of the tokenizer loaded from GGUF file is incorrect.

  #31630 closed Jul 23, 2024

• Table question answering pipeline failing to save

  #32128 closed Jul 23, 2024

• Very different output depending on whether an attention mask is passed when using caching

  #31943 closed Jul 23, 2024

• the script convert_wav2vec2_original_pytorch_checkpoint_to_pytorch.py does not work on fairseq wav2vec2-xls-r fine-tuned model

  #29182 closed Jul 23, 2024

• Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained

  #30762 closed Jul 23, 2024

• AttributeError: 'BertModel' object has no attribute 'attn_implementation'

  #30965 closed Jul 23, 2024

• MultiScaleDeformableAttentionFunction different results on different devices

  #31399 closed Jul 23, 2024

• LlavaNextVideo always assumes left padding when batch size is 1

  #32112 closed Jul 23, 2024

• Add llama3-llava-next-8b to convert_llava_next_weights_to_hf.py

  #31394 closed Jul 23, 2024

• "use_safetensors" not enforced with "local_files_only", loads bin file

  #31649 closed Jul 22, 2024

• Source link to `LlamaForSequenceClassification` seems broken, if so, update it.

  #31531 closed Jul 22, 2024

• Licence

  #32104 closed Jul 22, 2024

• Gemma-7b model set my own lm_head but cannot saved and changed pretrained embedding_layer's weights too.

  #31467 closed Jul 22, 2024

• Metadata

  #32121 closed Jul 22, 2024

• When tranining the RWKV, it report "backward error"

  #31413 closed Jul 22, 2024

• Transformer compatibility Python 3.9, ComfyUI

  #31806 closed Jul 22, 2024

• Model load when dtypes match is broken

  #32089 closed Jul 22, 2024

• unexpected keyword argument 'torch_empty_cache_steps' in TrainingArguments

  #32071 closed Jul 22, 2024

• HuggingFace `from_pretrained()` called multiple times when launching with `torchrun`

  #29843 closed Jul 21, 2024

• recent version of Transformers seems to mess with forward/__call__. Breaks patching loss function

  #30753 closed Jul 21, 2024

• Trained tokenizer has broken encoding for cyrillic

  #30937 closed Jul 21, 2024

• sdpa for bert causes nan output when mix-precision enabled.

  #31038 closed Jul 21, 2024

• Using a single 'RecurrentGemmaRglru' layer - "Trying to backward through the graph a second time" Error

  #31324 closed Jul 21, 2024

• https://github.com/VikParuchuri/surya can not convert model to onnx

  #31384 closed Jul 21, 2024

• ValueError: The checkpoint you are trying to load has model type `chameleon` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date.

  #32107 closed Jul 20, 2024

• past_key_values for SeamlessM4Tv2ForSpeechToText is not working as expected

  #29139 closed Jul 20, 2024

• `num_input_tokens_seen` included the `pad` tokens if sample padding strategy used

  #29889 closed Jul 20, 2024

• stop_strings not work at all

  #31363 closed Jul 20, 2024

• loading AutoTokenizer.from_pretrained with gated model

  #31367 closed Jul 20, 2024

• TypeError: Block.forward() got an unexpected keyword argument 'past_key_value'

  #31371 closed Jul 20, 2024

• linear_sum_assignment error in the object_detection.py guide

  #31461 closed Jul 20, 2024

• Support generating with fallback for short form audio in Whisper

  #29508 closed Jul 19, 2024

• Can you please provide:

  #32087 closed Jul 19, 2024

• The ProgressCallback triggers a `cannot pickle '_thread.lock' object` failure

  #32064 closed Jul 19, 2024

48 Issues opened by 45 people

• A question about code on Mistral-7B attention

  #32235 opened Jul 26, 2024

• LLaVA cannot use beam search after 4.43.0

  #32234 opened Jul 26, 2024

• SinkCache with Qwen1.5 broken in 4.43.0+

  #32233 opened Jul 25, 2024

• can't load the llama-3.1-8b-instruct model

  #32232 opened Jul 25, 2024

• [Whisper] Attention mask not detected in `Whisper.generate()`

  #32228 opened Jul 25, 2024

• Add New Optimizer

  #32225 opened Jul 25, 2024

• `BarkModel` can't be saved anymore

  #32224 opened Jul 25, 2024

• flashattention3

  #32219 opened Jul 25, 2024

• Parallel inference on generative models throws an exception

  #32217 opened Jul 25, 2024

• auto_find_batch_size for OOM during evaluation

  #32215 opened Jul 25, 2024

• Chat Assistant Prefill

  #32213 opened Jul 25, 2024

• RuntimeError: Failed to import transformers.pipelines because of the following error (look up to see its traceback):

  #32211 opened Jul 25, 2024

• get error when running the chatglm3: 'GenerationConfig' object has no attribute '_eos_token_tensor'

  #32207 opened Jul 25, 2024

• Does GroundingDINO support batched inference?

  #32206 opened Jul 25, 2024

• Broken accuracy on LLaMa 3.1 70B -- worse than even 8B

  #32205 opened Jul 24, 2024

• Cannot build documentation on Mac OS

  #32203 opened Jul 24, 2024

• Load Phi 3 small on Nvidia Tesla V100 - Flash Attention

  #32201 opened Jul 24, 2024

• Support `from_pretrained` of `FlaxPretrainedModel` from sharded `.safetensors` weights

  #32200 opened Jul 24, 2024

• Model loading is uneven on GPUs with AutomodelforCasualLM

  #32199 opened Jul 24, 2024

• error occur in the resize_embedding

  #32196 opened Jul 24, 2024

• "inverted" form required for 4D masking not defined / 4D attention masks breaks with transformers >=4.40

  #32195 opened Jul 24, 2024

• error occured while running _compute_llama3_parameters in modeling_repe_utils.py with torch.device('meta').

  #32187 opened Jul 24, 2024

• DataCollatorForLanguageModeling is (unnecessary) slow

  #32184 opened Jul 24, 2024

• Static KV cache with CPU offloading

  #32179 opened Jul 24, 2024

• LlavaNextForConditionalGeneration - incorrect way of creating `final_labels` in `_merge_input_ids_with_image_features`

  #32173 opened Jul 24, 2024

• `dataloader_prefetch_factor` is left unused for datasets of type `IterableDataset`

  #32169 opened Jul 23, 2024

• Enable speculative decoding with batch size >1

  #32165 opened Jul 23, 2024

• Add Matching Anything by Segmenting Anything (MASA) MOT tracking model

  #32164 opened Jul 23, 2024

• Adding warnings or errors when provided sequence length is bigger than config.max_position_embeddings

  #32154 opened Jul 23, 2024

• [i18n-<languageCode>] Translating docs to <languageName>

  #32146 opened Jul 22, 2024

• callback to implement how the predictions should be stored.

  #32145 opened Jul 22, 2024

• TimmBackbone.from_pretrained out_indices not working correctly (out_indices is sorted using strings blocks.11 before blocks.3)

  #32133 opened Jul 22, 2024

• Wav2Vec2ProcessorWithLM doesn't handle unknown token well for BPE

  #32132 opened Jul 22, 2024

• Is apply_chat_template support function call usage?

  #32130 opened Jul 22, 2024

• No module named 'transformers.modeling_flash_attention_utils'

  #32129 opened Jul 22, 2024

• TF Lite model created from TFWhisperForConditionalGeneration.from_pretrained craches

  #32125 opened Jul 21, 2024

• Exception raised when running `T5-like span-masked language modeling` example in `examples/flax/language-modeling/`

  #32124 opened Jul 21, 2024

• RuntimeError: Failed to import transformers.pipelines because of the following error (look up to see its traceback): module 'tensorflow' has no attribute 'data'

  #32122 opened Jul 21, 2024

• Output from model.Generate & model.forward not same when output attention/hidden_state is True

  #32117 opened Jul 21, 2024

• _prepare_4d_causal_attention_mask mask inversion should work boolean masks

  #32113 opened Jul 20, 2024

• Gemma template won't end with eos_token

  #32110 opened Jul 20, 2024

• gemma2 + flash atten Error: RuntimeError: linalg.vector_norm: Expected a floating point or complex tensor as input. Got Long

  #32103 opened Jul 20, 2024

• Using Trainer + a pretrained tokenizer + 4D attention mask is extremely slow

  #32101 opened Jul 19, 2024

• Unrecognized configuration class ChameleonConfig

  #32098 opened Jul 19, 2024

• max_length calculation for padding the generation outputs in the Seq2SeqTrainer prediction_step function

  #32095 opened Jul 19, 2024

• [Error] with Trainer: TypeError: Unsupported types (<class 'NoneType'>) passed to `_gpu_broadcast_one`.

  #32090 opened Jul 19, 2024

• The implementations of `LlamaAttention` and `LlamaSdpaAttention` are not equivalent.

  #32086 opened Jul 19, 2024

• Training multiple adapters

  #32084 opened Jul 19, 2024

189 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [WIP] - Add Microsoft CLAP model

  #31929 commented on Jul 25, 2024 • 52 new comments

• Add DAB-DETR Object detection/segmentation model

  #30803 commented on Jul 24, 2024 • 48 new comments

• Improve support for image generation with Chameleon & Anole

  #32013 commented on Jul 24, 2024 • 43 new comments

• Add GLM-4 and Later GLM Model (Draft)

  #31977 commented on Jul 26, 2024 • 41 new comments

• Support Kosmos-2.5

  #31711 commented on Jul 25, 2024 • 36 new comments

• [WIP] Add OmDet-Turbo

  #31843 commented on Jul 23, 2024 • 35 new comments

• Uniform kwargs for processors + Docs update - GroundingDINO

  #31964 commented on Jul 25, 2024 • 31 new comments

• Import structure & first three model refactors

  #31329 commented on Jul 25, 2024 • 25 new comments

• [GroundingDino] Fix grounding dino loss 🚨

  #31828 commented on Jul 23, 2024 • 24 new comments

• 🚨 Bloom support for cache class

  #31445 commented on Jul 26, 2024 • 22 new comments

• Support reading tiktoken tokenizer.model file

  #31656 commented on Jul 23, 2024 • 18 new comments

• RoPE: model-agnostic RoPE refactor

  #31999 commented on Jul 22, 2024 • 13 new comments

• Add codestral mamba2

  #32080 commented on Jul 26, 2024 • 12 new comments

• Add Nemotron HF Support

  #31699 commented on Jul 26, 2024 • 11 new comments

• Offloaded KV Cache

  #31325 commented on Jul 25, 2024 • 8 new comments

• Granite language models

  #31502 commented on Jul 24, 2024 • 8 new comments

• Implement SuperGlue model

  #29886 commented on Jul 20, 2024 • 7 new comments

• Add Descript-Audio-Codec model

  #31494 commented on Jul 26, 2024 • 7 new comments

• Add DINOv2 with registers

  #31832 commented on Jul 23, 2024 • 5 new comments

• _is_peft_model update to recognise peft submodules, allowing training quantised models with peft submodules

  #30884 commented on Jul 22, 2024 • 3 new comments

• Interpolate clip

  #31900 commented on Jul 22, 2024 • 3 new comments

• [pipeline] fix padding for 1-d tensors

  #31776 commented on Jul 25, 2024 • 3 new comments

• [whisper] compile compatibility with long-form decoding

  #31772 commented on Jul 25, 2024 • 3 new comments

• clean_up_tokenization_spaces=False if unset

  #31938 commented on Jul 25, 2024 • 2 new comments

• Add ViTPose

  #30530 commented on Jul 25, 2024 • 2 new comments

• Cache: new Cache format in decoder-only models

  #31421 commented on Jul 26, 2024 • 2 new comments

• Fix conflicting key in init kwargs in PreTrainedTokenizerBase

  #31233 commented on Jul 23, 2024 • 2 new comments

• Implement MambaForSequenceClassification

  #31155 commented on Jul 23, 2024 • 2 new comments

• Add Flax Dinov2

  #31960 commented on Jul 26, 2024 • 1 new comment

• SPLIT PR: eos bos tokens

  #31316 commented on Jul 22, 2024 • 1 new comment

• [RoBERTa-based] Add support for sdpa

  #30510 commented on Jul 22, 2024 • 1 new comment

• Add Zamba

  #30950 commented on Jul 24, 2024 • 0 new comments

• Add IRIS

  #30883 commented on Jul 25, 2024 • 0 new comments

• fix: multilingual midel convert to tflite get wrong token

  #32079 commented on Jul 19, 2024 • 0 new comments

• fixes clip interpolate

  #30783 commented on Jul 20, 2024 • 0 new comments

• pipeline 'text-classification' in >=4.40.0 throwing TypeError: Got unsupported ScalarType BFloat16

  #30542 commented on Jul 21, 2024 • 0 new comments

• filter flash_attn optional imports loading remote code

  #30954 commented on Jul 22, 2024 • 0 new comments

• [WIP] enable cpu bnb path

  #31098 commented on Jul 22, 2024 • 0 new comments

• Fix perceiver latent initialization modeling_idefics2.py

  #31151 commented on Jul 26, 2024 • 0 new comments

• [WIP] Add Tokenizer for MyT5 Model

  #31286 commented on Jul 22, 2024 • 0 new comments

• Reducing memory usage: removing useless logits computation in generate()

  #31292 commented on Jul 24, 2024 • 0 new comments

• fix: [whisper] don't overwrite GenerationConfig's `return_timestamps` when `return_timestamps` is not passed to `generate` function

  #31296 commented on Jul 26, 2024 • 0 new comments

• Added HHCache class implementing H2O Cache

  #31623 commented on Jul 26, 2024 • 0 new comments

• The last ut test of the QDQBert model ”test_inference_no_head_absolute_embedding” did not pass when using official safetensors

  #31486 commented on Jul 26, 2024 • 0 new comments

• "from_pretrained" read wrong config file. not "tokenizer_config.json", but "config.json"

  #31282 commented on Jul 26, 2024 • 0 new comments

• Error During Training with PatchTSMixerForTimeSeriesClassification for Time Series Classification

  #30614 commented on Jul 26, 2024 • 0 new comments

• Inconsitent module names (state_dict keys).

  #30124 commented on Jul 26, 2024 • 0 new comments

• Inconsistent special_token addition in EncoderDecoderModel forward pass

  #31729 commented on Jul 26, 2024 • 0 new comments

• Adding mixtral attention_bias in style of llama modeling

  #28440 commented on Jul 26, 2024 • 0 new comments

• Title: CUDA RuntimeError: Unspecified Launch Failure during Training

  #30913 commented on Jul 26, 2024 • 0 new comments

• Weights of LlamaForQuestionAnswering were not initialized from the model checkpoint

  #30381 commented on Jul 26, 2024 • 0 new comments

• [WIP] Improve multimodal processors - rely less on kwargs

  #28711 commented on Jul 24, 2024 • 0 new comments

• 🚨 Add Blip2ForImageTextRetrieval

  #29261 commented on Jul 24, 2024 • 0 new comments

• Fix from pretrained ignoring errors

  #29959 commented on Jul 26, 2024 • 0 new comments

• fix prompt tunning + deepspeed zero3 + checkpoint_saving hang issue

  #29980 commented on Jul 23, 2024 • 0 new comments

• schedulefree optimizers

  #30079 commented on Jul 23, 2024 • 0 new comments

• Add SDPA support for T5 Style Models

  #30375 commented on Jul 20, 2024 • 0 new comments

• Add trainer integration test for llava to ensure accelerate autocasting works correctly

  #30489 commented on Jul 19, 2024 • 0 new comments

• update based on tokenizers release

  #30574 commented on Jul 23, 2024 • 0 new comments

• Adding imagebind

  #30690 commented on Jul 23, 2024 • 0 new comments

• Remove device map for saving tokenizer config on TPU (fix for issue #31971)

  #32043 commented on Jul 19, 2024 • 0 new comments

• [`WIP`] Add Mamba2

  #32027 commented on Jul 20, 2024 • 0 new comments

• HFQuantizer implementation for compressed-tensors library

  #31704 commented on Jul 25, 2024 • 0 new comments

• [Demo][ExecuTorch] Lower and run native Gemma e2e in ExecuTorch

  #31706 commented on Jul 24, 2024 • 0 new comments

• Update kwargs validation for `preprocess` with decorator

  #32024 commented on Jul 24, 2024 • 0 new comments

• [WIP] Agents use grammar

  #31735 commented on Jul 25, 2024 • 0 new comments

• [docs] Redesign

  #31757 commented on Jul 25, 2024 • 0 new comments

• Adding mplugdocowl

  #31792 commented on Jul 25, 2024 • 0 new comments

• chore: move `conftest.py` to `tests/`

  #32011 commented on Jul 25, 2024 • 0 new comments

• Enable whisper encoder to accept any chunk length

  #31991 commented on Jul 25, 2024 • 0 new comments

• Add support for GGUF Phi-3

  #31844 commented on Jul 22, 2024 • 0 new comments

• Add cosine_with_min_lr_schedule_with_warmup_lr_rate scheduler in Trainer

  #31870 commented on Jul 24, 2024 • 0 new comments

• Add SOLO: A Single Transformer for Scalable Vision-Language Modeling

  #31918 commented on Jul 20, 2024 • 0 new comments

• Make special image tokens attribute of tokenizer

  #31967 commented on Jul 21, 2024 • 0 new comments

• Added optimizer adam mini

  #31933 commented on Jul 25, 2024 • 0 new comments

• # Fix an error of 'ValueError: mean must have 1 elements if it is an iterable, got 3' for method 'infer_channel_dimension_format' in 'image_utils.py'

  #31950 commented on Jul 20, 2024 • 0 new comments

• [feat] Apply rope_scaling for general use in phi3, llama

  #31966 commented on Jul 21, 2024 • 0 new comments

• make `p_mask` a numpy array before passing to `select_starts_ends`

  #32076 commented on Jul 26, 2024 • 0 new comments

• avoid padding for `num_frames` in `AutomaticSpeechRecognitionPipeline`

  #32074 commented on Jul 25, 2024 • 0 new comments

• Rest of model init refactors

  #31330 commented on Jul 25, 2024 • 0 new comments

• Uniformize model processors

  #31368 commented on Jul 24, 2024 • 0 new comments

• Add Cross-Attention to Bloom Model for VisionEncoderDecoder Compatibility

  #31432 commented on Jul 23, 2024 • 0 new comments

• [WIP] Standardize inputs and outputs for existing image-text-to-text models

  #32059 commented on Jul 25, 2024 • 0 new comments

• add changes in mistral model to avoid problems in pytorch hooks

  #31463 commented on Jul 19, 2024 • 0 new comments

• FIX / Hub: Also catch for `exceptions.ConnectionError`

  #31469 commented on Jul 22, 2024 • 0 new comments

• Update beam_constraints with KMP

  #31482 commented on Jul 19, 2024 • 0 new comments

• MixtralFlashAttention2: put "plus 1" inside parentheses when calculating rotary_seq_len, allowing None position_ids input.

  #31500 commented on Jul 25, 2024 • 0 new comments

• docs: ko: tasks/awq.md

  #32057 commented on Jul 25, 2024 • 0 new comments

• add bnb support for Ascend NPU

  #31512 commented on Jul 20, 2024 • 0 new comments

• Fix mininal version check for object_detection.md

  #31520 commented on Jul 21, 2024 • 0 new comments

• Sequence Length Invariant Text Models

  #31521 commented on Jul 21, 2024 • 0 new comments

• fix wav2vec2 with torch.compile

  #31538 commented on Jul 22, 2024 • 0 new comments

• handle when from_pretrained_id is a list

  #31541 commented on Jul 22, 2024 • 0 new comments

• Optimize 1st token for beam_search

  #31564 commented on Jul 24, 2024 • 0 new comments

• Allow infer_framework_load_model to use the originally specified config.

  #31580 commented on Jul 25, 2024 • 0 new comments

• activation_checkpointing error when using --fsdp

  #28499 commented on Jul 21, 2024 • 0 new comments

• Using batching with pipeline and transformers

  #31641 commented on Jul 22, 2024 • 0 new comments

• Uniform kwargs for processors

  #31911 commented on Jul 22, 2024 • 0 new comments

• Potential Bug in llava_next when calling pack_image_features function.

  #31529 commented on Jul 22, 2024 • 0 new comments

• Incorrect docstring of `get_anyres_image_grid_shape`

  #31588 commented on Jul 22, 2024 • 0 new comments

• Problem with the masked language modeling tutorial

  #31545 commented on Jul 22, 2024 • 0 new comments

• Nested from_pretrained() gives warnings loading weights - "copying from a non-meta parameter"

  #31544 commented on Jul 22, 2024 • 0 new comments

• GenerationConfig throws Object is not JSON serializable when setting constraints

  #31070 commented on Jul 22, 2024 • 0 new comments

• Mismatched tensor size error when generating text with beam_search on mps

  #30662 commented on Jul 22, 2024 • 0 new comments

• transformers offline model loading is not working from version 4.40.0 for models without safetensors

  #30469 commented on Jul 22, 2024 • 0 new comments

• torchrun breaks with load_model_at_end and with metric_for_best_model=eval_f1 on question_answering example

  #30819 commented on Jul 22, 2024 • 0 new comments

• Add support for Apple's DCLM-Baseline-7B model

  #32000 commented on Jul 22, 2024 • 0 new comments

• `MixtralFlashAttention2` subscripts `position_ids` before checking if it is `None`

  #31326 commented on Jul 22, 2024 • 0 new comments

• Dropout sync across GPUs causes major performance drops

  #31412 commented on Jul 22, 2024 • 0 new comments

• [BUG] Offline loading of non-safe tensors fails

  #30920 commented on Jul 22, 2024 • 0 new comments

• ddp_time in TrainingArguments with deepspeed doesn't take effect

  #32036 commented on Jul 22, 2024 • 0 new comments

• FineWeb SLM Training doesn't start

  #31501 commented on Jul 22, 2024 • 0 new comments

• Have `_is_peft_model` check if there's any peft submodule/Allow quantised training

  #30878 commented on Jul 22, 2024 • 0 new comments

• RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1! (HF/Accelerate)

  #31504 commented on Jul 22, 2024 • 0 new comments

• Add Depth Anything v2 metric depth

  #31972 commented on Jul 22, 2024 • 0 new comments

• Add post_process_depth_estimation to image processors

  #30917 commented on Jul 22, 2024 • 0 new comments

• flash_attn ImportError breaking model loading (Florence-2-base-ft)

  #31793 commented on Jul 22, 2024 • 0 new comments

• Fixing Tensor Shape/Dimension Mismatch Errors in TimeSeries Transformer for Stock Price Prediction

  #31556 commented on Jul 24, 2024 • 0 new comments

• Batch is empty when fine-tuning flan-t5 using LoRA

  #31357 commented on Jul 19, 2024 • 0 new comments

• [Severe Bug] Performance Degradation Starting from v4.42.*

  #31890 commented on Jul 19, 2024 • 0 new comments

• Saved weights differ from the original model

  #30543 commented on Jul 19, 2024 • 0 new comments

• cannot use activation_checkpoint in torch native fsdp

  #32073 commented on Jul 19, 2024 • 0 new comments

• Open to contribution: adding `torch.nn.functional.scaled_dot_product_attention` support for more architectures

  #28005 commented on Jul 19, 2024 • 0 new comments

• ViTLayer.forward() needs to be in "eager" mode when `output_attentions=True`

  #30978 commented on Jul 19, 2024 • 0 new comments

• Add support for MiniCPM-V-2 and MiniCPM-Llama3-V-2_5

  #31836 commented on Jul 19, 2024 • 0 new comments

• More robust tests required for gradient checkpointing

  #32063 commented on Jul 19, 2024 • 0 new comments

• BertForSequenceClassification.from_pretrained broken when using FSDP

  #32068 commented on Jul 19, 2024 • 0 new comments

• Performance mismatch with best_epoch

  #32075 commented on Jul 19, 2024 • 0 new comments

• GPT-2 Model Logits and Loss are different on MPS

  #32005 commented on Jul 19, 2024 • 0 new comments

• 如果在单个GPU上out of memory 如何用两个GPU加载推理同一个模型？

  #31508 commented on Jul 20, 2024 • 0 new comments

• load qwen2-72b-instruct sft awq q4_0 gguf ValueError: Trying to set a tensor of shape torch.Size

  #31507 commented on Jul 20, 2024 • 0 new comments

• ImportError: cannot import name 'logging' from 'huggingface_hub'

  #31492 commented on Jul 20, 2024 • 0 new comments

• run_clm.py AttributeError: 'NoneType' object has no attribute 'get'

  #31487 commented on Jul 20, 2024 • 0 new comments

• RecurrentGemma Doesn't Support left padding?

  #31201 commented on Jul 20, 2024 • 0 new comments

• Unable to load starcoder2 finetuned version getting quantization errors

  #29990 commented on Jul 20, 2024 • 0 new comments

• Using accelerate launch FDSP cause weight saved after 2nd time onwards to be incomplete

  #31034 commented on Jul 20, 2024 • 0 new comments

• Plans to Integrate LongRoPE into LLaMA?

  #31992 commented on Jul 20, 2024 • 0 new comments

• OOM when loading 300B models with `AutoModelForCausalLM.from_pretrained` and `BitsAndBytesConfig` quantization.

  #31577 commented on Jul 21, 2024 • 0 new comments

• from_pretrained 加载checkpoint过慢的问题

  #31515 commented on Jul 21, 2024 • 0 new comments

• GenerationMixin sample() runs forever

  #31484 commented on Jul 21, 2024 • 0 new comments

• error when convert llama1 ckpts to hf formath

  #30723 commented on Jul 21, 2024 • 0 new comments

• Make fx traced model with the use of `past_key_values` pickable again?

  #30575 commented on Jul 21, 2024 • 0 new comments

• bart-large-xsum model: There were missing keys in the checkpoint model loaded: ['model.encoder.embed_tokens.weight', 'model.decoder.embed_tokens.weight', 'lm_head.weight'].

  #29128 commented on Jul 24, 2024 • 0 new comments

• (False?) warning about weight_g/weight_v missing on WeightNorm on PyTorch

  #26796 commented on Jul 24, 2024 • 0 new comments

• NotImplementedError: Cannot copy out of meta tensor; no data when embedding to meta

  #31560 commented on Jul 24, 2024 • 0 new comments

• Idefics2 fine-tuning: Error when unscale_gradients called on FP16 gradients during training with transformers and accelerate

  #30559 commented on Jul 24, 2024 • 0 new comments

• Optimised 4bit inference kernels

  #28568 commented on Jul 24, 2024 • 0 new comments

• Bug in whisper word-level timestamps (`tokenizer._decode_asr`)

  #31778 commented on Jul 24, 2024 • 0 new comments

• Converting gguf fp16 & bf16 to hf is not supported.

  #31762 commented on Jul 24, 2024 • 0 new comments

• Improving memory efficiency further 🚀

  #30860 commented on Jul 24, 2024 • 0 new comments

• `Gemma2Model` not returning cache

  #31981 commented on Jul 24, 2024 • 0 new comments

• KV cache with CPU offloading

  #30704 commented on Jul 24, 2024 • 0 new comments

• Implement Cross Attention in LLAMA Model

  #27285 commented on Jul 25, 2024 • 0 new comments

• RuntimeError: slow_conv2d_forward_mps: input(device='cpu') and weight(device=mps:0')

  #31571 commented on Jul 25, 2024 • 0 new comments

• Trainer: To keep unused columns for `compute_metrics`

  #31570 commented on Jul 25, 2024 • 0 new comments

• Tokenizers: Character encoding inconsistencies between __call__ and .convert_tokens_to_ids

  #31438 commented on Jul 25, 2024 • 0 new comments

• Whisper Translation on low resource languages

  #30592 commented on Jul 25, 2024 • 0 new comments

• `pip install accelerate` (and similar) error messages should specify min version

  #31583 commented on Jul 25, 2024 • 0 new comments

• Multi-GPU inference affects LLM's (Llama2-7b-chat-hf) generation.

  #31582 commented on Jul 25, 2024 • 0 new comments

• push_to_hub doesn't push checkpoint folder while training

  #30141 commented on Jul 25, 2024 • 0 new comments

• Bug version 4.42.4: KeyError: 'Cache only has 0 layers, attempted to access layer with index 0'

  #32060 commented on Jul 26, 2024 • 0 new comments

• Embedding class is replaced when calling `resize_token_embeddings`

  #31835 commented on Jul 26, 2024 • 0 new comments

• When max_steps < save_steps with deepspeed zero3 stage

  #31624 commented on Jul 26, 2024 • 0 new comments

• Unable to export Phi-3-vision model to PyTorch exported program

  #31622 commented on Jul 26, 2024 • 0 new comments

• HuggingFace GroundingDINO inference execution time is slower than the original groundingDINO (~100ms)

  #31533 commented on Jul 26, 2024 • 0 new comments

• Checkpoint validation as an option

  #32067 commented on Jul 22, 2024 • 0 new comments

• Whisper - get probability of detected language

  #29293 commented on Jul 22, 2024 • 0 new comments

• Support H100 training with FP8 in Trainer and Deepspeed

  #25333 commented on Jul 22, 2024 • 0 new comments

• Cannot export Deberta to TorchScript

  #20815 commented on Jul 22, 2024 • 0 new comments

• Index out of range when generate using optimum

  #31551 commented on Jul 23, 2024 • 0 new comments

• Error on fine tuning paligemma for object detection

  #31528 commented on Jul 23, 2024 • 0 new comments

• Mixtral's implementation of auxiliary loss seems incorrect

  #31464 commented on Jul 23, 2024 • 0 new comments

• rework `test_multi_gpu_data_parallel_forward`

  #31087 commented on Jul 23, 2024 • 0 new comments

• DPT implementation contains unused parameters

  #30633 commented on Jul 23, 2024 • 0 new comments

• `test_encode_decode_fast_slow_all_tokens` is failing

  #30045 commented on Jul 23, 2024 • 0 new comments

• SDPA gives nans/infs during sampling on ROCM w/ float16

  #30056 commented on Jul 23, 2024 • 0 new comments

• Fail to load model without .safetensors file

  #31552 commented on Jul 23, 2024 • 0 new comments

• Skipping cudagraphs for unknown reason

  #31645 commented on Jul 23, 2024 • 0 new comments

• Training Evaluation Display on VSCode

  #22694 commented on Jul 23, 2024 • 0 new comments

• kwargs pop "attn_implement" twice in modeling_utils.py and configuration_utils.py when use AutoConfig/AutoModel

  #32082 commented on Jul 23, 2024 • 0 new comments

• NonMatchingSplitsSizesError on Flax BART with wiki summary dataset

  #29596 commented on Jul 23, 2024 • 0 new comments

• [flax_llama] Why is the return value of the `create_sinusoidal_positions` truncated by `num_pos`?

  #29590 commented on Jul 23, 2024 • 0 new comments

• FP8 inference and FP8 KV cache

  #23660 commented on Jul 23, 2024 • 0 new comments

• SeamlessM4TFeatureExtractor fails with pad_to_multiple_of not being a multiple of stride

  #31916 commented on Jul 23, 2024 • 0 new comments

• Add MistralForQuestionAnswering

  #28908 commented on Jul 23, 2024 • 0 new comments

• Flash Attention with Gemma 2

  #31953 commented on Jul 23, 2024 • 0 new comments

• static cache implementation is not compatible with attn_implementation==flash_attention_2

  #32040 commented on Jul 23, 2024 • 0 new comments

• Quantization support for heads and embeddings

  #31474 commented on Jul 23, 2024 • 0 new comments

• Race condition when loading models from local folders with custom code

  #27421 commented on Jul 23, 2024 • 0 new comments