3 Releases published by 1 person

• v4.49.0-Gemma-3 Gemma 3 (Based on v4.49.0)

  published Mar 18, 2025

• v4.49.0-Mistral-3 Mistral 3 (Based on v4.49.0)

  published Mar 18, 2025

• v4.50.0 Release v4.50.0

  published Mar 21, 2025

89 Pull requests merged by 54 people

• Update installation.md

  #36826 merged Mar 21, 2025

• [docs] Model docs

  #36469 merged Mar 21, 2025

• Fix Pan and Scan on batched images Gemma3

  #36864 merged Mar 21, 2025

• Simplify keep_in_fp32_modules logic

  #36722 merged Mar 21, 2025

• fix: loss computation after embeddings resize - mllama

  #36840 merged Mar 21, 2025

• Fix: dtype cannot be str

  #36262 merged Mar 21, 2025

• Minor Gemma 3 fixes

  #36884 merged Mar 21, 2025

• Use deformable_detr kernel from the Hub

  #36853 merged Mar 21, 2025

• Gemma 3 tests expect greedy decoding

  #36882 merged Mar 21, 2025

• 🔴 🔴 🔴 supersede paligemma forward to shift pos id indexing

  #36859 merged Mar 21, 2025

• [generate] model defaults being inherited only happens for newer models

  #36881 merged Mar 21, 2025

• Revert "Update deprecated Jax calls (#35919)"

  #36880 merged Mar 21, 2025

• Make ViTPooler configurable

  #36517 merged Mar 21, 2025

• chore: fix typos in the tests directory

  #36813 merged Mar 21, 2025

• Remove call to .item in get_batch_samples

  #36861 merged Mar 21, 2025

• FIX FSDP plugin update for QLoRA

  #36720 merged Mar 21, 2025

• [CI] doc builder without custom image

  #36862 merged Mar 21, 2025

• Mllama: raise better error

  #35934 merged Mar 21, 2025

• Refactor Aya Vision with modular

  #36688 merged Mar 20, 2025

• Add support for seed in DataCollatorForLanguageModeling

  #36497 merged Mar 20, 2025

• [CI] fix update metadata job

  #36850 merged Mar 20, 2025

• Gemma3: fix test

  #36820 merged Mar 20, 2025

• [torchao] revert to get_apply_tensor_subclass

  #36849 merged Mar 20, 2025

• Add model visual debugger

  #36798 merged Mar 20, 2025

• Add Prompt Depth Anything Model

  #35401 merged Mar 20, 2025

• Refactor Attention implementation for ViT-based models

  #36545 merged Mar 20, 2025

• DeepSpeed tensor parallel+ZeRO

  #36825 merged Mar 20, 2025

• Support loading Quark quantized models in Transformers

  #36372 merged Mar 20, 2025

• Use pyupgrade --py39-plus to improve code

  #36843 merged Mar 20, 2025

• Fix hqq skipped modules and dynamic quant

  #36821 merged Mar 20, 2025

• Fix ONNX export for sequence classification head

  #36332 merged Mar 20, 2025

• Shieldgemma2

  #36678 merged Mar 20, 2025

• Fix: remove the redundant snippet of _whole_word_mask

  #36759 merged Mar 20, 2025

• Gemma 3: Adding explicit GenerationConfig and refactoring conversion …

  #36833 merged Mar 20, 2025

• Fix import for torch 2.0, 2.1 - guard typehint for "device_mesh"

  #36768 merged Mar 20, 2025

• Update min safetensors bis

  #36823 merged Mar 20, 2025

• [generate] clarify docstrings: when to inherit GenerationMixin

  #36605 merged Mar 20, 2025

• [modular] Sort modular skips

  #36304 merged Mar 20, 2025

• Pass state dict

  #35234 merged Mar 20, 2025

• [qwen2 audio] remove redundant code and update docs

  #36282 merged Mar 20, 2025

• Update deprecated Jax calls

  #35919 merged Mar 20, 2025

• Fix fp16 ONNX export for RT-DETR and RT-DETRv2

  #36460 merged Mar 20, 2025

• Pass num_items_in_batch directly to loss computation

  #36753 merged Mar 20, 2025

• Saving Trainer.collator.tokenizer in when Trainer.processing_class is None

  #36552 merged Mar 20, 2025

• fix tiktoken convert to pass AddedToken to Tokenizer

  #36566 merged Mar 20, 2025

• [ForCausalLMLoss] allow users to pass shifted labels

  #36607 merged Mar 20, 2025

• Disable inductor config setter by default

  #36608 merged Mar 20, 2025

• Fix swanlab global step

  #36728 merged Mar 20, 2025

• rewrite main method in Qwen2, making it more clear

  #36772 merged Mar 20, 2025

• Move the warning to the documentation for DataCollatorWithFlattening

  #36707 merged Mar 20, 2025

• Remove our AdamW implementation

  #36177 merged Mar 19, 2025

• Update configuration_qwen2.py

  #36735 merged Mar 19, 2025

• quick fix fast_image_processor register error

  #36716 merged Mar 19, 2025

• Add Space to Bitsandbytes doc

  #36834 merged Mar 19, 2025

• Support tracable dynamicKVcache

  #36311 merged Mar 19, 2025

• One more fix for reviewer assignment

  #36829 merged Mar 19, 2025

• [gemma 3] multimodal checkpoints + AutoModelForCausalLM

  #36741 merged Mar 19, 2025

• enable OffloadedCache on XPU from PyTorch 2.7

  #36654 merged Mar 19, 2025

• Add option for ao base configs

  #36526 merged Mar 19, 2025

• Add attention visualization tool

  #36630 merged Mar 19, 2025

• [Generation] remove leftover code from end-to-end compilation

  #36685 merged Mar 19, 2025

• Fix Device map for bitsandbytes tests

  #36800 merged Mar 19, 2025

• Remove dist": "loadfile" for pytest for CircleCI jobs

  #36811 merged Mar 19, 2025

• fix "Cannot copy out of meta tensor; no data!" issue for BartForConditionalGeneration model

  #36572 merged Mar 19, 2025

• Expectations test utils

  #36569 merged Mar 18, 2025

• [generate] ✨ vectorized beam search ✨

  #35802 merged Mar 18, 2025

• Support custom dosctrings in modular

  #36726 merged Mar 18, 2025

• Fix chameleon's TypeError because inputs_embeds may None

  #36673 merged Mar 18, 2025

• Fix casting dtype for qunatization

  #36799 merged Mar 18, 2025

• Fix Mistral3 tests

  #36797 merged Mar 18, 2025

• Loading optimizations

  #36742 merged Mar 18, 2025

• Update SHA for tj-actions/changed-files

  #36795 merged Mar 18, 2025

• fix hqq due to recent modeling changes

  #36771 merged Mar 18, 2025

• Add Mistral3

  #36790 merged Mar 18, 2025

• Fix gemma3_text tokenizer in mapping

  #36793 merged Mar 18, 2025

• Fixing typo in gemma3 image_processor_fast and adding a small test

  #36776 merged Mar 18, 2025

• chore: fix typos in tests directory

  #36785 merged Mar 18, 2025

• fix typos in the tests directory

  #36717 merged Mar 17, 2025

• doc: Clarify is_decoder usage in PretrainedConfig documentation

  #36724 merged Mar 17, 2025

• [docs] Update README

  #36265 merged Mar 17, 2025

• [CI] remove redundant checks in test_eager_matches_sdpa_inference

  #36740 merged Mar 17, 2025

• [MINOR:TYPO] Update hubert.md

  #36733 merged Mar 17, 2025

• Fix TrainingArguments.torch_empty_cache_steps post_init check

  #36734 merged Mar 17, 2025

• Fix test isolation for clear_import_cache utility

  #36345 merged Mar 17, 2025

• fix xpu tests

  #36656 merged Mar 17, 2025

• Allow ray datasets to be used with trainer

  #36699 merged Mar 17, 2025

• fix can_generate

  #36570 merged Mar 17, 2025

• enable/disable compile for quants methods

  #36519 merged Mar 17, 2025

• 🚨🚨🚨 Fix sdpa in sam and refactor relative position embeddings

  #36422 merged Mar 17, 2025

60 Pull requests opened by 49 people

• Remove extra tensor clone in PyTorch code

  #36748 opened Mar 16, 2025

• 🌐 [i18n-KO] Translated `qwen2_vl.md` to Korean

  #36750 opened Mar 16, 2025

• Add Qwen2.5-Omni

  #36752 opened Mar 16, 2025

• 🌐 [i18n-KO] Translated 'serving.md' to Korean

  #36756 opened Mar 17, 2025

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

  #36757 opened Mar 17, 2025

• feat: expose the strict flag to allow catching missing model layers while loading a checkpoint

  #36760 opened Mar 17, 2025

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

  #36763 opened Mar 17, 2025

• Add support for audios in apply_chat_template

  #36770 opened Mar 17, 2025

• Export for Phi4-mini

  #36780 opened Mar 18, 2025

• Use public export API on torch 2.5 and future

  #36781 opened Mar 18, 2025

• Fix attention_mask dimension issue in GPT2Model

  #36782 opened Mar 18, 2025

• Create modeling_ngen3.py for NGen3

  #36787 opened Mar 18, 2025

• Update configuration_auto.py for NGen3

  #36791 opened Mar 18, 2025

• Refactor `return_dict` logic to remove complicated if/else paths

  #36794 opened Mar 18, 2025

• [don't merge] check tokenizer ci job

  #36796 opened Mar 18, 2025

• Add Granite Speech Support

  #36801 opened Mar 18, 2025

• Add long vita

  #36807 opened Mar 19, 2025

• Support loading custom models (`trust_remote_code=True`) in offline mode from local

  #36808 opened Mar 19, 2025

• fix unexpected kws of input_ids when setup no speech detection of whisper

  #36809 opened Mar 19, 2025

• check tok

  #36818 opened Mar 19, 2025

• Dummies

  #36827 opened Mar 19, 2025

• [Modeling] Load FP8 safetensors such as DeepSeek

  #36828 opened Mar 19, 2025

• gemma3 fp16 fix

  #36832 opened Mar 19, 2025

• Enhance Model Loading By Providing Parallelism, Uses Optional Env Flag

  #36835 opened Mar 19, 2025

• Remove unnecessary attr assignment

  #36837 opened Mar 19, 2025

• Move `return_dict` logic into `can_return_tuple` decorator

  #36838 opened Mar 19, 2025

• Haocheng lu

  #36839 opened Mar 19, 2025

• Fix Optional type annotation

  #36841 opened Mar 20, 2025

• fix pegasus init weights and other copied models

  #36844 opened Mar 20, 2025

• [Utils] torch version checks optionally accept dev versions

  #36847 opened Mar 20, 2025

• Add support for specifying revisions when pushing to Hub via internal Trainer call

  #36852 opened Mar 20, 2025

• [2/N] Use pyupgrade --py39-plus to improve code

  #36857 opened Mar 20, 2025

• [DON'T MERGE] test doc builder

  #36860 opened Mar 20, 2025

• fix: prevent input side-effects in processor text args

  #36866 opened Mar 20, 2025

• Only count num items in batch when needed

  #36867 opened Mar 20, 2025

• Fix warning message for PEFT models in text-generation pipeline #36783

  #36868 opened Mar 20, 2025

• [docs] Fix image link

  #36869 opened Mar 20, 2025

• Improve Model Download Speeds By ~3x For Large Models

  #36870 opened Mar 21, 2025

• Bump ruff to 0.11.1

  #36871 opened Mar 21, 2025

• Correct Condition for Pixel Values in Chameleon PR to Address Embedding and Token Mismatch

  #36873 opened Mar 21, 2025

• [Fix] Add `original_max_position_embeddings` to YARN rope_scaling optional keys

  #36877 opened Mar 21, 2025

• Adding Qwen3 and Qwen3MoE

  #36878 opened Mar 21, 2025

• Fix `resume_from_checkpoint` not recognising `"last-checkpoint"`

  #36883 opened Mar 21, 2025

• Optimize `to_py_obj` for python-native numeric lists and scalars

  #36885 opened Mar 21, 2025

• Fix warning message for PEFT models in text-generation pipeline #36783

  #36887 opened Mar 21, 2025

• Allow easy registration of custom attention functions

  #36889 opened Mar 21, 2025

• Fix processor kwargs qwen2 vl

  #36890 opened Mar 21, 2025

• Fix SDPA implementation in Qwen2-VL (issues with torch==2.6.0)

  #36891 opened Mar 21, 2025

• [WIP] Computer vision util: vision visualizer

  #36892 opened Mar 21, 2025

• fix Gemma3 Config

  #36893 opened Mar 21, 2025

• Enable tracing for Moshi

  #36894 opened Mar 21, 2025

• Add RF-DETR

  #36895 opened Mar 21, 2025

• tests: fix asyncio.wait() usage for python>=3.11

  #36898 opened Mar 22, 2025

• Adding ArlowGPT

  #36899 opened Mar 22, 2025

• Add NGen3

  #36901 opened Mar 22, 2025

• Added support for seed in `DataCollatorForWholeWordMask`

  #36903 opened Mar 22, 2025

• LogfireCallback: Integrating Logfire with Hugging Face’s Trainer

  #36905 opened Mar 22, 2025

• Fix torch version guard at import

  #36907 opened Mar 22, 2025

• fix cached file error when repo type is dataset

  #36909 opened Mar 23, 2025

• Fix typos

  #36910 opened Mar 23, 2025

43 Issues closed by 16 people

• Unable to export GLM models to ONNX

  #35021 closed Mar 23, 2025

• ModernBERT inference fails on CPU: ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)

  #35388 closed Mar 23, 2025

• `modular_model_converter` can not handle objects import via try - except

  #35414 closed Mar 23, 2025

• `TFViTModel` and `interpolate_pos_encoding=True`

  #36155 closed Mar 23, 2025

• Inference with FSDP during training affects checkpoints

  #34530 closed Mar 22, 2025

• [BART] Cannot copy out of meta tensor; no data!

  #36247 closed Mar 21, 2025

• Bug introduced in `from_pretrained` `v4.48.3`..`v4.49.0`

  #36258 closed Mar 21, 2025

• <spam>

  #36876 closed Mar 21, 2025

• Gemma3 (and Paligemma) position_ids 1-indexed?

  #36856 closed Mar 21, 2025

• torch._subclasses.fake_tensor.DataDependentOutputException: aten._local_scalar_dense.default with `_prepare_4d_attention_mask_for_sdpa(

  #36123 closed Mar 21, 2025

• Speaker Verification: All Speakers Getting Perfect 1.000 Similarity Scores

  #36124 closed Mar 21, 2025

• Allow setting a seed for DataCollatorForLanguageModeling

  #36357 closed Mar 20, 2025

• LlamaAttention has no attribute `rotary_emb` (4.50.0.dev0)

  #36758 closed Mar 20, 2025

• GPT2 repetition of words in output

  #36848 closed Mar 20, 2025

• num_items_in_batch unexpected in vision encoder decoder

  #36744 closed Mar 20, 2025

• Convert RT-DETR model to coreml

  #35905 closed Mar 20, 2025

• [bug] fast_image_processor register error

  #36715 closed Mar 19, 2025

• When what needs to be loaded is in the cache directory, there is no need to make a request to the remote

  #36762 closed Mar 19, 2025

• In the _speculative_sampling function, it seems that the "squeeze" method is being used incorrectly.

  #36810 closed Mar 19, 2025

• AttributeError: 'Gemma3Config' object has no attribute 'vocab_size'

  #36683 closed Mar 19, 2025

• text-to-video_app

  #36747 closed Mar 19, 2025

• model from_pretrained bug in 4.50.dev0 in these days

  #36506 closed Mar 19, 2025

• Subtle difference with Pytorch AdamW?

  #35504 closed Mar 19, 2025

• Qwen2VL exhibits significant performance differences under different attention implementations.

  #35749 closed Mar 19, 2025

• [Phi-3-mini-128k-instruct] Difference of encodings for Slow and Fast Tokenizer

  #35973 closed Mar 19, 2025

• Traning loss not showing with trainer

  #36102 closed Mar 19, 2025

• when model.generate with num_beams=2 and num_return_sequences=2,the output seqs are different from input_ids of stopping_criteria

  #34574 closed Mar 18, 2025

• `UnboundLocalError: cannot access local variable 'images_list'` when using Gemma 3 AutoProcessor with use_fast=True

  #36739 closed Mar 18, 2025

• Gemma3 minimal fine tuning example?

  #36714 closed Mar 18, 2025

• Shape mismatch in RoPE embeddings gpt_neox model when rotary_ndims is odd

  #35233 closed Mar 18, 2025

• incorrect special_tokens_mask

  #35897 closed Mar 18, 2025

• Llama tokenizer newline character inconsistency

  #35923 closed Mar 18, 2025

• flex_attention does not output the full attention_weights with output_attention option

  #36096 closed Mar 18, 2025

• bug in save checkpoint

  #36099 closed Mar 18, 2025

• qwen2_5_vl processor padding side is wrong.

  #36100 closed Mar 18, 2025

• ValueError: weight is on the meta device, we need a `value` to put in on 0. `Gemma3`

  #36766 closed Mar 17, 2025

• Misleading documentation for `is_decoder` configuration parameter

  #36482 closed Mar 17, 2025

• Gemma2 (quantized) inference is broken - torch._dynamo.exc.UserError: Dynamic control flow is not supported at the moment.

  #36485 closed Mar 17, 2025

• On MoE implementation in HuggingFace

  #36730 closed Mar 17, 2025

• bus error on version 4.43.0 with pretrained community CLIP model - MacOS

  #33357 closed Mar 17, 2025

• Cannot load siglip2 processor

  #36665 closed Mar 16, 2025

• SFTConfig.__init__() got an unexpected keyword argument 'optimizers'

  #36749 closed Mar 16, 2025

• Model.generate use_cache=True generates different results than use_cache=False

  #36536 closed Mar 16, 2025

41 Issues opened by 40 people

• `lm_head.weight` missing from `convert_mistral_weights_to_hf.STATE_DICT_MAPPING`

  #36908 opened Mar 23, 2025

• MacOs: register_pytree_node got an unexpected keyword argument 'flatten_with_keys_fn'

  #36906 opened Mar 22, 2025

• PixtralVisionModel does not support Flash Attention 2.0 yet

  #36904 opened Mar 22, 2025

• Warning: "No label_names provided for PeftModel" persists despite dataset containing "labels" column

  #36902 opened Mar 22, 2025

• groot n1

  #36900 opened Mar 22, 2025

• GPT2Model model output inconsistency between different transformers versions

  #36897 opened Mar 22, 2025

• Forced to hit `UserWarning` when generating with `temperature=0`

  #36896 opened Mar 21, 2025

• Issue with update

  #36888 opened Mar 21, 2025

• Florence2 stopped working after upgrade to 4.50.0 ("Unrecognized configuration class")

  #36886 opened Mar 21, 2025

• Add RF-DETR model

  #36879 opened Mar 21, 2025

• Qwen2-VL-7B-Instruct shape error when using TP=4

  #36875 opened Mar 21, 2025

• Support for SpatialLM series model

  #36874 opened Mar 21, 2025

• Optimize tokenizer.decode() Performance for `List[int]` Inputs

  #36872 opened Mar 21, 2025

• Multiple processor classes have input side-effects

  #36865 opened Mar 20, 2025

• Facing RunTime Attribute error while running different Flax models for RoFormer

  #36854 opened Mar 20, 2025

• Tansfomers_model

  #36846 opened Mar 20, 2025

• Unable to load google/siglip2-so400m-patch14-384/

  #36845 opened Mar 20, 2025

• GOT-OCR2 docs indicate model can produce markdown, but it only produces LaTeX.

  #36836 opened Mar 19, 2025

• Build for Windows and VS 2022 does not compile CUDA sources

  #36830 opened Mar 19, 2025

• Support for Ovis2 models

  #36824 opened Mar 19, 2025

• Gemma 3 is broken with fp16

  #36822 opened Mar 19, 2025

• Need Option to Disable Flash Attention in VideoLLaMA2.1-7B-AV (SiglipVisionModel)

  #36819 opened Mar 19, 2025

• Add EuroBert Model To Config

  #36817 opened Mar 19, 2025

• Gemma3 can't be fine-tuned on multi-image examples

  #36816 opened Mar 19, 2025

• Gemma3

  #36815 opened Mar 19, 2025

• When I use BF16 or FP16 to perform Lora fine-tuning on GemMA-3-12B-it, there will be an error when saving the checkpoint, but FP32 is normal

  #36814 opened Mar 19, 2025

• Not able to trace GPT2DoubleHeadsModel

  #36812 opened Mar 19, 2025

• Logic Errors in Image_processing_gemma3_fast.py

  #36806 opened Mar 19, 2025

• Qwen2VLForConditionalGeneration.from_pretrained() hangs with v0.50.0-dev0

  #36803 opened Mar 18, 2025

• BERT is broken on `v4.49.0-Gemma-3`

  #36802 opened Mar 18, 2025

• Design question for integrating new model to Transformers?

  #36784 opened Mar 18, 2025

• Throw messages in text-generation task with deepseek r1 with PEFTModel

  #36783 opened Mar 18, 2025

• Please support GGUF format for UMT5EncoderModel

  #36774 opened Mar 17, 2025

• Inconsistent Documentation for `⁠dataset_index` Requirement Across ViTPose Models

  #36773 opened Mar 17, 2025

• Add Audio inputs available in apply_chat_template

  #36769 opened Mar 17, 2025

• Source link to Ray Tune API outdated

  #36765 opened Mar 17, 2025

• could not parse ModelProto from /home/imss/zxhhhh/llama-3-8b/tokenizer.model

  #36764 opened Mar 17, 2025

• tj-actions/changed-files action compromised

  #36761 opened Mar 17, 2025

• Add Gemma 3 For Sequence Classification

  #36755 opened Mar 16, 2025

• Unable to load google/siglip2-base-patch16-naflex

  #36754 opened Mar 16, 2025

• IdeficsProcessor cannot handle multiple images in one text

  #36751 opened Mar 16, 2025

130 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.

• Add Janus model

  #36053 commented on Mar 20, 2025 • 42 new comments

• Samhq model addition

  #35147 commented on Mar 22, 2025 • 27 new comments

• Add StyleTTS 2

  #35790 commented on Mar 19, 2025 • 27 new comments

• add FlashAttentionKwargs and seq_idx to flat collator

  #36456 commented on Mar 21, 2025 • 24 new comments

• Add evolla rebase main

  #36232 commented on Mar 21, 2025 • 18 new comments

• Add InternVL (2.5 MPO)

  #35968 commented on Mar 19, 2025 • 13 new comments

• Add DeepSeek V2 Model into Transformers

  #36400 commented on Mar 20, 2025 • 12 new comments

• Add Doge model

  #35891 commented on Mar 22, 2025 • 12 new comments

• Add TimesFM Time Series Forecasting Model

  #34082 commented on Mar 19, 2025 • 11 new comments

• Introduce modular files for speech models

  #35902 commented on Mar 21, 2025 • 10 new comments

• Add MLCD model

  #36182 commented on Mar 19, 2025 • 10 new comments

• Add support for MiniMax's MiniMax-Text-01

  #35831 commented on Mar 20, 2025 • 10 new comments

• [Feature] Support using FlashAttention2 on Ascend NPU

  #36696 commented on Mar 22, 2025 • 9 new comments

• Create and Expose SamVisionModel as public for better accessibility

  #36493 commented on Mar 19, 2025 • 7 new comments

• Support batch size > 1 image-text inference

  #36682 commented on Mar 21, 2025 • 5 new comments

• Add FAST

  #35476 commented on Mar 21, 2025 • 4 new comments

• Add internlm3 dense

  #35694 commented on Mar 19, 2025 • 4 new comments

• Support `return_tensors` in audio chat templates

  #34601 commented on Mar 18, 2025 • 3 new comments

• Fixes DynamicCache export issues due to control flow and inplace modifications

  #36652 commented on Mar 20, 2025 • 3 new comments

• fix: support grad clipping for TP through replicating non-sharded modules

  #36132 commented on Mar 21, 2025 • 3 new comments

• Add index selection for `output_hidden_states`

  #33705 commented on Mar 18, 2025 • 3 new comments

• Add Distill Any Depth

  #36614 commented on Mar 21, 2025 • 2 new comments

• [Whisper] 🚨 Fix pipeline word timestamp: timestamp token is end of token time !!!

  #36632 commented on Mar 20, 2025 • 2 new comments

• Fix long lagging when streaming text without spaces and NJK chars

  #36708 commented on Mar 19, 2025 • 2 new comments

• Fix generation using flash-attention and static cache

  #36729 commented on Mar 17, 2025 • 2 new comments

• Integrate xlstm cleanly.

  #35377 commented on Mar 18, 2025 • 1 new comment

• Export T5 (encoder-decoder) to ExecuTorch

  #36486 commented on Mar 18, 2025 • 1 new comment

• Support QuestionAnswering Module for ModernBert based models.

  #35566 commented on Mar 22, 2025 • 1 new comment

• Add support for DeepseekAI's DeepseekVL

  #36248 commented on Mar 23, 2025 • 0 new comments

• Pipeline: fix unnecessary warnings

  #35753 commented on Mar 17, 2025 • 0 new comments

• Fix the eval_use_gather_object flag usage

  #36214 commented on Mar 18, 2025 • 0 new comments

• fix: condition bos_token_id and space as token

  #36211 commented on Mar 19, 2025 • 0 new comments

• (ugly) Use `parallelism=4` for `check_repository_consistency`

  #36197 commented on Mar 20, 2025 • 0 new comments

• Flash Attention v3

  #36190 commented on Mar 21, 2025 • 0 new comments

• fix immediate quantization of the first token in QuantizedCache

  #35760 commented on Mar 20, 2025 • 0 new comments

• [MLU] Fix FA2 check error, remove deepspeed-mlu deps.

  #36159 commented on Mar 20, 2025 • 0 new comments

• Remove head mask in generative models

  #35786 commented on Mar 19, 2025 • 0 new comments

• [WIP] add deepseek-v3

  #35926 commented on Mar 20, 2025 • 0 new comments

• Fix Mask2Former Weight Initialization Issues #35877

  #35904 commented on Mar 21, 2025 • 0 new comments

• Several fixes related to rotary position embeddings

  #35901 commented on Mar 19, 2025 • 0 new comments

• [WP] PagedAttention + Prefix Cache for FlashAttention2

  #36737 commented on Mar 21, 2025 • 0 new comments

• Convert _VALID_DICT_FIELDS to class attribute for shared dict parsing in subclasses

  #36736 commented on Mar 22, 2025 • 0 new comments

• Fix image processor speedup fixed

  #36732 commented on Mar 17, 2025 • 0 new comments

• Add CSM model

  #36719 commented on Mar 21, 2025 • 0 new comments

• fix whisper re-compile

  #36712 commented on Mar 21, 2025 • 0 new comments

• [WIP] Add DINO DETR Model to HuggingFace Transformers

  #36711 commented on Mar 17, 2025 • 0 new comments

• prune LM Head for USD

  #36695 commented on Mar 19, 2025 • 0 new comments

• don't pass NoneType for keep_in_fp32_modules

  #36675 commented on Mar 20, 2025 • 0 new comments

• Update quantizer_bnb_4bit.py

  #36669 commented on Mar 20, 2025 • 0 new comments

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

  #36649 commented on Mar 17, 2025 • 0 new comments

• Fix device issue in modeling_qwen2

  #36647 commented on Mar 21, 2025 • 0 new comments

• Refine parameter type annotations

  #36644 commented on Mar 20, 2025 • 0 new comments

• [WiP] Add Aimv2 model

  #36625 commented on Mar 22, 2025 • 0 new comments

• [WIP] Add support to load models with transforms

  #36621 commented on Mar 23, 2025 • 0 new comments

• Fixed 30s timestamp resets in Whisper long-form transcription

  #36612 commented on Mar 20, 2025 • 0 new comments

• Allow saving and loading multiple "raw" chat template files

  #36588 commented on Mar 20, 2025 • 0 new comments

• fix for loading gguf quantized model

  #36563 commented on Mar 20, 2025 • 0 new comments

• [Validation] First implementation of `@strict_dataclass` from `huggingface_hub`

  #36534 commented on Mar 17, 2025 • 0 new comments

• Add an event related to forward in the TrainerCallback

  #36496 commented on Mar 18, 2025 • 0 new comments

• Add NVIDIA Cosmos

  #36476 commented on Mar 19, 2025 • 0 new comments

• Customize docstrings fast image processor

  #36466 commented on Mar 17, 2025 • 0 new comments

• Add PlainDETR

  #36437 commented on Mar 18, 2025 • 0 new comments

• Fix: Use config.use_sliding_window instead of config.sliding_window

  #36377 commented on Mar 21, 2025 • 0 new comments

• Add EfficientLoFTR model

  #36355 commented on Mar 19, 2025 • 0 new comments

• 🚨Deprecate legacy argument for image-text-to-text models and adopt new behavior by default

  #36307 commented on Mar 20, 2025 • 0 new comments

• [`ModernBERT`] Never save 'reference_compile' config; should be set based on end user

  #36305 commented on Mar 20, 2025 • 0 new comments

• enable tp on CPU

  #36299 commented on Mar 19, 2025 • 0 new comments

• Update composition flag usage

  #36263 commented on Mar 19, 2025 • 0 new comments

• Add D-FINE Model into Transformers

  #36261 commented on Mar 20, 2025 • 0 new comments

• [docs] add return_timestamps=True for Whisper long-form transcription

  #35633 commented on Mar 20, 2025 • 0 new comments

• CVE-2024-11392 - AWS Scanner and Trivy Flagging Transformers 4.48.1 as Vulnerable

  #36041 commented on Mar 20, 2025 • 0 new comments

• Incompatibility in flash_attention_2 + Llama + Transformers>=4.43 + Autocast to fp16

  #36224 commented on Mar 20, 2025 • 0 new comments

• IsADirectoryError when training with tqdm enabled for trainer

  #34766 commented on Mar 20, 2025 • 0 new comments

• `AutoModelForCasualLM.from_pretrained()` exits without warning/error

  #36245 commented on Mar 19, 2025 • 0 new comments

• Marian RNN conversion support

  #36651 commented on Mar 19, 2025 • 0 new comments

• Stop output to stdout in streamers.py methods

  #36562 commented on Mar 19, 2025 • 0 new comments

• Inconsistent output lengths when `max_length=20` is set implicitly vs explicitly in `generate()`

  #35765 commented on Mar 19, 2025 • 0 new comments

• A warning message showing that `MultiScaleDeformableAttention.so` is not found in `/root/.cache/torch_extensions` if `ninja` is installed with `transformers`

  #35349 commented on Mar 19, 2025 • 0 new comments

• Difficulties with multi-GPU Inferencing

  #36634 commented on Mar 19, 2025 • 0 new comments

• Cryptic error when using AutoTokenizer with SentencePiece tokenizers without sentencepiece installed

  #36291 commented on Mar 19, 2025 • 0 new comments

• Model Card to include key information (e.g. max_sequence_length, etc.)

  #36743 commented on Mar 19, 2025 • 0 new comments

• Unable to deploy Gemma 3 on AWS SageMaker due to lack of support in tranfomers release

  #36738 commented on Mar 19, 2025 • 0 new comments

• Error when tokenizer is set to string: `AttributeError: 'str' object has no attribute 'pad_token_id'`

  #36731 commented on Mar 19, 2025 • 0 new comments

• Custom 4D tensor caused shape mismatch error

  #35290 commented on Mar 18, 2025 • 0 new comments

• Gemma 3 1B - TypeError: 'NoneType' object is not callable

  #36745 commented on Mar 18, 2025 • 0 new comments

• Tensor size mismatch when trying to run RT-DETR on multiple gpus

  #33165 commented on Mar 18, 2025 • 0 new comments

• Issue with Progressive Generation Using inputs_embeds and past_key_values

  #35707 commented on Mar 18, 2025 • 0 new comments

• RWKV CUDA error: an illegal memory access was encountered during training from scratch

  #35805 commented on Mar 18, 2025 • 0 new comments

• Whisper `.generate()` function not respecting `max_new_tokens` or `max_length`

  #36183 commented on Mar 18, 2025 • 0 new comments

• Token healing throws error with "Qwen/Qwen2.5-Coder-7B-Instruct"

  #36210 commented on Mar 18, 2025 • 0 new comments

• [bug] use_gather_object is not respected after the first eval in trainer

  #36213 commented on Mar 18, 2025 • 0 new comments

• Model trained with Flash Attention 2.0 raises "RuntimeError: query and key must have the same dtype" when generating

  #30019 commented on Mar 18, 2025 • 0 new comments

• lm_head parameters missing from named_parameters() in Qwen2.5-VL-3B-Instruct model

  #36598 commented on Mar 17, 2025 • 0 new comments

• SDPA `is_causal=False` has no effect due to `LlamaModel._prepare_4d_causal_attention_mask_with_cache_position`

  #36150 commented on Mar 17, 2025 • 0 new comments

• model.generate function is not compatible with custom position_ids

  #36510 commented on Mar 17, 2025 • 0 new comments

• MultiTask Classification and label_names on Trainer

  #33193 commented on Mar 17, 2025 • 0 new comments

• The parameter 'text' may be None as the comments says, there is a confuse.

  #36667 commented on Mar 17, 2025 • 0 new comments

• ../aten/src/ATen/native/cuda/Indexing.cu:1289: indexSelectLargeIndex: block: [267,0,0], thread: [25,0,0] Assertion `srcIndex < srcSelectDimSize` failed.

  #33985 commented on Mar 17, 2025 • 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 Mar 17, 2025 • 0 new comments

• Offline mode doesn't work with models that require `trust_remote_code=True`

  #34855 commented on Mar 16, 2025 • 0 new comments

• Fix Batch Size Mismatch When Using `crops_n_layers` in `mask-generation` Pipeline #35530

  #35627 commented on Mar 20, 2025 • 0 new comments

• Process inputs directly in apply_chat_template in image-text-to-text pipeline

  #35616 commented on Mar 19, 2025 • 0 new comments

• Add Relation DETR

  #34900 commented on Mar 20, 2025 • 0 new comments

• Bye bye env vars, keep everything as configs

  #34886 commented on Mar 19, 2025 • 0 new comments

• Add Molmo (7B-D, 7B-O, 70B)

  #33962 commented on Mar 21, 2025 • 0 new comments

• [`AutoDocstring`] Based on inspect parsing of the signature

  #33771 commented on Mar 17, 2025 • 0 new comments

• Trainer: add predict with generate

  #32346 commented on Mar 21, 2025 • 0 new comments

• Improve support for image generation with Chameleon & Anole

  #32013 commented on Mar 19, 2025 • 0 new comments

• safetensor/mmap memory leak when per-layer weights are converted do other dtypes

  #34366 commented on Mar 23, 2025 • 0 new comments

• tensor parallel training bug

  #36296 commented on Mar 23, 2025 • 0 new comments

• Bug about num_update_steps_per_epoch in function _inner_training_loop

  #36297 commented on Mar 23, 2025 • 0 new comments

• DS3 zero3_save_16bit_model is not compatible with resume_from_checkpoint

  #36317 commented on Mar 23, 2025 • 0 new comments

• PruneTrain: Fast Neural Network Training by Dynamic Sparse Model Reconfiguration

  #10105 commented on Mar 23, 2025 • 0 new comments

• Error during processing: MllamaForCausalLM does not support Flash Attention 2.0 yet.

  #36557 commented on Mar 22, 2025 • 0 new comments

• flash_attention_2 2.7.2.post1 seems to crash when using `torch.compile` and `DataCollatorWithFlattening`

  #35588 commented on Mar 22, 2025 • 0 new comments

• rework `test_multi_gpu_data_parallel_forward`

  #31087 commented on Mar 22, 2025 • 0 new comments

• model.gradient_checkpointing_enable() makes loss.requires_grad be False

  #35826 commented on Mar 22, 2025 • 0 new comments

• Torch -> ONNX doesn't work after upgrading transformers to 4.49.0

  #36276 commented on Mar 22, 2025 • 0 new comments

• AttributeError: 'dict' object has no attribute 'to_dict'; for Inferencing Lora Merged Qwen/Qwen2.5-VL-3B-Instruct

  #36281 commented on Mar 22, 2025 • 0 new comments

• past_key_value(s) name inconsistency causing problems

  #36290 commented on Mar 22, 2025 • 0 new comments

• [Bugs] RuntimeError: No CUDA GPUs are available in transformers v4.48.0 or above when running Ray RLHF example

  #36295 commented on Mar 22, 2025 • 0 new comments

• Add Magma from Microsoft to Transformers

  #36629 commented on Mar 21, 2025 • 0 new comments

• AutoModel from_pretrained does not recursively download relative imports

  #36653 commented on Mar 21, 2025 • 0 new comments

• Qwen2VLForConditionalGeneration doesn't work with MPS devices

  #36413 commented on Mar 21, 2025 • 0 new comments

• Mask2Former _init_weights

  #35877 commented on Mar 21, 2025 • 0 new comments

• Support Flex Attention for encoder only models (XLMRoberta, ModernBERT etc...)

  #36697 commented on Mar 21, 2025 • 0 new comments

• Community contribution: Adding GGUF support for more architectures

  #33260 commented on Mar 21, 2025 • 0 new comments

• cannot import name 'is_timm_config_dict' from 'transformers.utils.generic'

  #36068 commented on Mar 21, 2025 • 0 new comments

• ValueError: Unrecognized image processor in Qwen/Qwen2.5-VL-3B-Instruct.

  #36193 commented on Mar 21, 2025 • 0 new comments

• modeling_phi3 errors with AttributeError: 'DynamicCache' object has no attribute 'get_max_length'

  #36071 commented on Mar 20, 2025 • 0 new comments