1 Release published by 1 person

• v4.55.0 v4.55.0: New openai GPT OSS model!

  published Aug 5, 2025

82 Pull requests merged by 43 people

• unpin torchcodec==0.5.0 and use torch 2.8 on daily CI

  #40072 merged Aug 10, 2025

• Update HuBERT model card according to template

  #39742 merged Aug 10, 2025

• Revert "fix notification_service.py about time_spent"

  #40044 merged Aug 8, 2025

• GLM-4.5V Model Support

  #39805 merged Aug 8, 2025

• fix notification_service.py about time_spent

  #40037 merged Aug 8, 2025

• Bnb failling tests

  #40026 merged Aug 8, 2025

• Tie weights recursively on all submodels

  #39996 merged Aug 8, 2025

• [core] Refactor the Cache logic to make it simpler and more general

  #39797 merged Aug 8, 2025

• Fix missing None default values for Gemma3n model in get_placeholder_mask (#39991)

  #40024 merged Aug 8, 2025

• Harmonize past_key_value to past_key_valueS everywhere

  #39956 merged Aug 8, 2025

• Fix an annoying flaky test

  #40000 merged Aug 8, 2025

• Higgs modules_to_not_convert standardization

  #39989 merged Aug 8, 2025

• Fix broken image inference for Fuyu model

  #39915 merged Aug 8, 2025

• pin torchcodec==0.5.0 for now with torch 2.7.1 on daily CI

  #40013 merged Aug 7, 2025

• Update expected output values after #39885 (part 2)

  #40015 merged Aug 7, 2025

• Raising error when quantizing a quantized model

  #39998 merged Aug 7, 2025

• docs: fix duplication in 'en/optimizers.md'

  #40014 merged Aug 7, 2025

• unpin torch<2.8 on circleci

  #40012 merged Aug 7, 2025

• FA2 can continue generation from cache

  #39843 merged Aug 7, 2025

• Fix default values of getenv

  #39867 merged Aug 7, 2025

• Fix HGNetV2 Model Card and Image Classification Pipeline Usage Tips

  #39965 merged Aug 7, 2025

• fix: remove CHAT_TEMPLATE import in tests for deepseek-vl

  #40003 merged Aug 7, 2025

• Fix missing video inputs for PerceptionLM.

  #39971 merged Aug 7, 2025

• Fix int4 quantized model cannot work with cpu

  #39724 merged Aug 7, 2025

• Update expected output values after #39885 (part 1)

  #39990 merged Aug 7, 2025

• Fix consistency

  #39995 merged Aug 7, 2025

• Fix return typehint for decoder and annotate inv_freq

  #39610 merged Aug 7, 2025

• Bump transformers from 4.48.0 to 4.53.0 in /examples/tensorflow/language-modeling-tpu

  #39967 merged Aug 7, 2025

• Fix gemma3n feature extractor's incorrect squeeze

  #39919 merged Aug 7, 2025

• [Idefics] fix device mismatch

  #39981 merged Aug 7, 2025

• Various test fixes for AMD

  #39978 merged Aug 7, 2025

• Support input_embeds in torch exportable decoders

  #39836 merged Aug 7, 2025

• [superglue] Fixed the way batch mask was applied to the scores before match assignment computation

  #39968 merged Aug 7, 2025

• Gemma3 fixes

  #39960 merged Aug 7, 2025

• Modular fix: remove the model name in find_file_type

  #39897 merged Aug 6, 2025

• chore: update Deformable_Detr model card

  #39902 merged Aug 6, 2025

• [bugfix] fix flash_attention_2 unavailable error on Ascend NPU

  #39844 merged Aug 6, 2025

• Fix fix_and_overwrite mode of utils/check_docstring.py

  #39369 merged Aug 6, 2025

• remove triton_kernels dep with kernels instead

  #39926 merged Aug 6, 2025

• fix glm4v image process

  #39964 merged Aug 6, 2025

• fix typo

  #39936 merged Aug 6, 2025

• Fix grammatical error in MoE variable name: expert_hitted → expert_hit, hitted_experts → hit_experts

  #39959 merged Aug 6, 2025

• docs: fix typo in 'quantization-aware training'

  #39904 merged Aug 6, 2025

• Enable gpt-oss mxfp4 on older hardware (sm75+)

  #39940 merged Aug 6, 2025

• Fix MXFP4 quantizer validation to allow CPU inference with dequantize option

  #39953 merged Aug 6, 2025

• [docs] ko toc fix

  #39927 merged Aug 6, 2025

• circleci: pin torch 2.7.1 until torchcodec is updated

  #39951 merged Aug 6, 2025

• Fix CI: Tests failing on CPU due to torch.device('cpu').index being None

  #39933 merged Aug 6, 2025

• Avoid utils/check_bad_commit.py failing due to rate limit (requesting api.github.com)

  #39918 merged Aug 5, 2025

• [CI] post-GptOss fixes for green CI

  #39929 merged Aug 5, 2025

• gpt_oss last chat template changes

  #39925 merged Aug 5, 2025

• Add GPT OSS model from OpenAI

  #39923 merged Aug 5, 2025

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

  #39535 merged Aug 5, 2025

• Export SmolvLM

  #39614 merged Aug 5, 2025

• Update object_detection.md

  #39909 merged Aug 5, 2025

• run model debugging with forward arg

  #39905 merged Aug 5, 2025

• Revert "remove dtensors, not explicit (#39840)"

  #39912 merged Aug 5, 2025

• Fix aria tests

  #39879 merged Aug 5, 2025

• Fix eval thread fork bomb

  #39717 merged Aug 5, 2025

• Replace video_fps with fps in tests

  #39898 merged Aug 5, 2025

• Fix misleading WandB error when WANDB_DISABLED is set

  #39891 merged Aug 5, 2025

• Avoid aliasing in cond's branches for torch 2.8

  #39488 merged Aug 5, 2025

• Remove unnecessary CUDA sync in qwen2_5_vl

  #39870 merged Aug 5, 2025

• fix test_working_of_tp failure of accelerate ut

  #39828 merged Aug 5, 2025

• [Exaone4] Fixes the attn implementation!

  #39906 merged Aug 5, 2025

• Reorder serving docs

  #39634 merged Aug 5, 2025

• chore: update DETR model card

  #39822 merged Aug 4, 2025

• Add support for ModernBertForMultipleChoice

  #39232 merged Aug 4, 2025

• send some feedback when manually building doc via comment

  #39889 merged Aug 4, 2025

• Update cohere2 vision test

  #39888 merged Aug 4, 2025

• [DOCS] : Improved mimi model card

  #39824 merged Aug 4, 2025

• Fix link to models in README

  #39880 merged Aug 4, 2025

• Better return type hint for AutoModelForCausalLM and AutoModelForImageTextToText

  #39881 merged Aug 4, 2025

• Set torch.backends.cudnn.allow_tf32 = False for CI

  #39885 merged Aug 4, 2025

• Replace Tokenizer with PreTrainedTokenizerFast in ContinuousBatchProcessor

  #39858 merged Aug 4, 2025

• Rework add-new-model-like with modular and make test filenames coherent

  #39612 merged Aug 4, 2025

• Fix quant docker for fp-quant

  #39641 merged Aug 4, 2025

• Fix attn_implementation setter for models with backbone_config

  #39855 merged Aug 4, 2025

• Add support for including in-memory videos (not just files/urls) in apply_chat_template

  #39494 merged Aug 4, 2025

• Use comment to build doc on PRs

  #39846 merged Aug 4, 2025

• Refactor label name handling for PEFT models in Trainer class

  #39265 merged Aug 4, 2025

• Improve is_wandb_available function to verify WandB installation

  #39875 merged Aug 4, 2025

74 Pull requests opened by 56 people

• Remove deprecated max_size parameter from ConditionalDetrImageProcessor

  #39883 opened Aug 4, 2025

• added Textnet fast image processor

  #39884 opened Aug 4, 2025

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

  #39886 opened Aug 4, 2025

• Move old generation modes to the Hub 🧹🧹🧹🧽🧽

  #39887 opened Aug 4, 2025

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

  #39890 opened Aug 4, 2025

• [docs] Add reference to HF-maintained `custom_generate` collections

  #39894 opened Aug 4, 2025

• Add Videoprism

  #39895 opened Aug 4, 2025

• [model] Support MiniCPM-V 4.0

  #39899 opened Aug 5, 2025

• 🌐 [i18n-KO] Translated `fp_quant` to Korean

  #39901 opened Aug 5, 2025

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

  #39903 opened Aug 5, 2025

• Update dynamic attnt setter for multimodals

  #39908 opened Aug 5, 2025

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

  #39913 opened Aug 5, 2025

• 🌐 [i18n-KO] Updated ko/perf_train_cpu.md

  #39917 opened Aug 5, 2025

• 🌐 [i18n-KO] Updated ko/perf_train_special.md

  #39920 opened Aug 5, 2025

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

  #39922 opened Aug 5, 2025

• Add chat template tests

  #39924 opened Aug 5, 2025

• Fix hidden torchvision>=0.15 dependency issue

  #39928 opened Aug 5, 2025

• Add missing special token properties to MistralCommonTokenizer

  #39930 opened Aug 5, 2025

• Registers StaticCache serialization functions for torch.export.export

  #39931 opened Aug 5, 2025

• Fix whisper `return_language` with `return_timestamp=word`

  #39938 opened Aug 5, 2025

• fixing image_utils.py todo

  #39941 opened Aug 6, 2025

• fix llama issue

  #39942 opened Aug 6, 2025

• Add back `_tp_plan` attribute

  #39944 opened Aug 6, 2025

• Add pytest marker: `torch_compile_test` and `torch_export_test`

  #39950 opened Aug 6, 2025

• Use torch._check instead of a test to make the model Gemma3 exportable

  #39962 opened Aug 6, 2025

• Add Keypoint Matcher pipeline

  #39970 opened Aug 6, 2025

• Causal loss for `ForConditionalGeneration`

  #39973 opened Aug 7, 2025

• [bugfix] Fix tensor device in Idefics2, Idefics3, and SmolVLM

  #39975 opened Aug 7, 2025

• Fix Qwen3 MoE GGUF architecture mismatch

  #39976 opened Aug 7, 2025

• Fix cross-attention masking before residual connection

  #39979 opened Aug 7, 2025

• Fix setting attention for multimodal models

  #39984 opened Aug 7, 2025

• fix: resolve triton version check compatibility on windows

  #39986 opened Aug 7, 2025

• Add vggt-hf copied from vit

  #39987 opened Aug 7, 2025

• Update Glm4V processor and add tests

  #39988 opened Aug 7, 2025

• Default to dequantize if cpu in device_map for mxfp4

  #39993 opened Aug 7, 2025

• chore: Add type hints to import_utils.py module

  #39994 opened Aug 7, 2025

• make sure position_ids are passed in for causal mask creation for gpt-oss

  #39997 opened Aug 7, 2025

• allow TP to work in ND-parallel with fsdp cpu ram efficient loading

  #39999 opened Aug 7, 2025

• [`Flash Attention`] Fix flash attention integration

  #40002 opened Aug 7, 2025

• Fix PerceptionLM image preprocessing for non-tiled image input.

  #40006 opened Aug 7, 2025

• 🚨 Use lru_cache for sine pos embeddings MaskFormer

  #40007 opened Aug 7, 2025

• Fixes for EncoderDecoderCache

  #40008 opened Aug 7, 2025

• feat: extract rev in attn_implementation kernels via @

  #40009 opened Aug 7, 2025

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

  #40011 opened Aug 7, 2025

• [WIP] Fix naive for loops for MoE models resulting in sub 20% downstream MFU for training with trl, e.t.c (Qwen3, Deepseek V3, Ernie 4.5, GLM 4.5, Dots1)

  #40016 opened Aug 7, 2025

• Feat/add gpt oss sequence classification

  #40019 opened Aug 8, 2025

• [fix] batch inference for llava_onevision

  #40021 opened Aug 8, 2025

• fix: resolve dropout type error in DogeDecoder

  #40022 opened Aug 8, 2025

• Add support for SDPA for OWLViT and OWLv2

  #40023 opened Aug 8, 2025

• [GLM4V] fix vision placeholder mask

  #40025 opened Aug 8, 2025

• Add amd runners to run-slow command

  #40027 opened Aug 8, 2025

• Revert FA2 kwargs construction

  #40029 opened Aug 8, 2025

• Update boxes expectations for OWLViT test

  #40030 opened Aug 8, 2025

• Add model card for MobileViT

  #40033 opened Aug 8, 2025

• Remove deprecated cache-related objects

  #40035 opened Aug 8, 2025

• Fix error on importing unavailable torch.distributed

  #40038 opened Aug 8, 2025

• New DynamicSlidingWindow layer & caches

  #40039 opened Aug 8, 2025

• [`GPT Big Code`] Fix attention scaling

  #40041 opened Aug 8, 2025

• Add GptOssForSequenceClassification for GPT-OSS models

  #40043 opened Aug 8, 2025

• (small) fix conditional for input_ids and input_embeds in marian

  #40045 opened Aug 8, 2025

• Update wavlm.md to match new model card template

  #40047 opened Aug 8, 2025

• Standardize BARTpho model card: badges, new examples, fixed broken im…

  #40051 opened Aug 9, 2025

• Auto-log parallelism info to wandb.config using HF Accelerate

  #40055 opened Aug 9, 2025

• updated visualBERT modelcard

  #40057 opened Aug 9, 2025

• GGUF Qwen2VL

  #40058 opened Aug 9, 2025

• Fix Inefficient GELU implementation in GPT2

  #40059 opened Aug 9, 2025

• Avoid CUDA stream sync

  #40060 opened Aug 10, 2025

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

  #40061 opened Aug 10, 2025

• fix: move super().__init__ after vision_config init in Mistral3Config

  #40063 opened Aug 10, 2025

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

  #40064 opened Aug 10, 2025

• Delay float32 upcast in ForCausalLMLoss after filtering ignore_index

  #40065 opened Aug 10, 2025

• Change Qwen2RMSNorm to RMSNorm from PyTorch

  #40066 opened Aug 10, 2025

• Add missing arguments

  #40068 opened Aug 10, 2025

• Remove _prepare_flash_attention_from_position_ids

  #40069 opened Aug 10, 2025

43 Issues closed by 22 people

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

  #36183 closed Aug 10, 2025

• Gemma2 fall back to cpu execusion when attn_implementation='flash_attention_2'

  #39188 closed Aug 10, 2025

• Previous PRs introduced a bug on Accumulated Gradients Losses

  #40052 closed Aug 9, 2025

• Incorrect word timestamps and word repetitions with Whisper-Large-v3-turbo model

  #37248 closed Aug 9, 2025

• Pretrainedtokenizerfast Segmentation fault

  #39099 closed Aug 9, 2025

• New release 4.53.0 breaks HF trainer/model

  #39111 closed Aug 9, 2025

• Gradient accumulation steps for Vision Languge model

  #39123 closed Aug 9, 2025

• Not capable of exporting Mistral to ONNX format with the use of caching

  #39162 closed Aug 9, 2025

• Error when loading gguf file

  #40040 closed Aug 9, 2025

• Weights not tied when loading `from_pretrained` with a wrapped model

  #39900 closed Aug 8, 2025

• `TypeError: 'builtins.safe_open' object is not iterable` in `load_pytorch_state_dict_in_tf2_model `

  #40028 closed Aug 8, 2025

• Major issues with transformers version causing rubbish generations with Gemma3 family using vllm

  #40017 closed Aug 8, 2025

• AttributeError: 'BitsAndBytesConfig' object has no attribute 'get_loading_attributes' with transformers 4.55.0

  #39939 closed Aug 8, 2025

• Gemma3n get_placeholder_mask issue

  #39991 closed Aug 8, 2025

• flash-attn cannot perform deterministic computation

  #39982 closed Aug 8, 2025

• RoBERTa is not well implemented for tokenizers with pad_token_id != 1

  #34528 closed Aug 8, 2025

• [DeepSeek-V3] Different rotary embedding implementation between DeepSeek-AI and Transformers

  #39687 closed Aug 8, 2025

• ModernBertUnpaddedRotaryEmbedding __init__ error

  #39934 closed Aug 7, 2025

• video_inputs are not passed to perception_lm

  #40004 closed Aug 7, 2025

• Flash Attention fails with non aligned position_ids

  #39814 closed Aug 7, 2025

• `convert_deepseek_vl_weights_to_hf.py` not included in v4.55.0 release.

  #39966 closed Aug 7, 2025

• [Gemma3N] Audio processing issue

  #39911 closed Aug 7, 2025

• v4.55.0 Idefics3 RuntimeError Tensors on different devices

  #39947 closed Aug 7, 2025

• [gpt‑oss] eager_attention_forward not using sliding-window attention for GPT‑OSS models

  #39954 closed Aug 7, 2025

• Finetune `gpt-oss-20b` with `mxfp4` quantization

  #39969 closed Aug 6, 2025

• Fix grammatically incorrect variable name "expert_hitted" → "expert_hit" in MoE implementation

  #39955 closed Aug 6, 2025

• transformers serve doesn't handle OPTIONS http method

  #39932 closed Aug 6, 2025

• 454545

  #39864 closed Aug 6, 2025

• ImportError: cannot import name 'GenerationMixin' from 'transformers.generation'

  #38442 closed Aug 6, 2025

• Streaming mode support on HF vs kyutai-labs for the mimi model

  #38535 closed Aug 6, 2025

• enable GraniteMoeHybridIntegrationTest in UT

  #38542 closed Aug 6, 2025

• Llama4 inference encounter unsupported op in dynamo ?

  #38118 closed Aug 6, 2025

• Misleading WandB error when WANDB_DISABLED=True and report_to="wandb" are both set

  #39878 closed Aug 5, 2025

• pytorch_utils.py > isin_mps_friendly > RuntimeError: Expected elements.dtype() == test_elements.dtype() to be true, but got false.

  #37423 closed Aug 5, 2025

• Inefficient memory resharding in attention layer

  #39072 closed Aug 5, 2025

• Inefficient default GELU implementation in GPT2

  #39073 closed Aug 5, 2025

• facebook/dinov2-with-registers-giant does not appear to have a file named pytorch_model.bin, model.safetensors, tf_model.h5, model.ckpt or flax_model.msgpack.

  #39075 closed Aug 5, 2025

• AttributeError: 'HfTrainerDeepSpeedConfig' object has no attribute 'is_zero3'

  #39081 closed Aug 5, 2025

• Why `lm-head` weight still exists with `"tie_word_embeddings": true`

  #39812 closed Aug 4, 2025

• Iwin Transformer: Hierarchical Vision Transformer using Interleaved Windows

  #39704 closed Aug 4, 2025

• ValueError: Max cache length is not consistent across layers

  #39877 closed Aug 4, 2025

• Allow video objects (np array etc.) in apply_chat_template (not just paths or urls)

  #36560 closed Aug 4, 2025

• Exception while inference Qwen2VL and Qwen2VL, assert module.weight.shape[1] == 1

  #38665 closed Aug 4, 2025

38 Issues opened by 38 people

• Issue running model from ImageSegmentationPipeline

  #40071 opened Aug 10, 2025

• Transformer GGUF support philosophy / naive question

  #40070 opened Aug 10, 2025

• [BUG] No umt5 config for GGUF. This is not supported configuration.

  #40067 opened Aug 10, 2025

• [Mistral3] attn_implementation not applied to vision_tower.config in Mistral3Config due to init order

  #40062 opened Aug 10, 2025

• Question: How to write a custome tokenizer form scratch

  #40056 opened Aug 9, 2025

• Whisper transcription accuracy improves when last 1600 samples of input audio are muted

  #40054 opened Aug 9, 2025

• Support text classification with GPT-OSS models

  #40050 opened Aug 9, 2025

• Please support loading Qwen 2.5 VL from GGUF

  #40049 opened Aug 9, 2025

• Recent releases break backwards-compatibility with key_cache

  #40046 opened Aug 8, 2025

• Support loading glm4moe GGUF

  #40042 opened Aug 8, 2025

• `plamo-2-1b` broken on latest main

  #40034 opened Aug 8, 2025

• Add Padding Strategy to DataCollatorForLanguageModeling

  #40032 opened Aug 8, 2025

• [gpt-oss] MoE routing bug in the mxfp4 implementation (in distributed setting)

  #40031 opened Aug 8, 2025

• accelerate==1.10.0 and safetensors==0.6.1 are incompatible with transformers==4.53.1

  #40020 opened Aug 8, 2025

• need GptOssForSequenceClassification

  #40018 opened Aug 8, 2025

• Customizable Logit Warping Strategies for Generation

  #40010 opened Aug 7, 2025

• Possible wrong init call

  #40001 opened Aug 7, 2025

• [gpt-oss] Transform checkpoint from safetensors to state dict

  #39992 opened Aug 7, 2025

• Triton version check compatibility on windows

  #39985 opened Aug 7, 2025

• CVE fix for v4.37.2 and v4.38.0

  #39983 opened Aug 7, 2025

• FSDP2 not compatible with transformers >= 4.54.0 GenericForTokenClassification

  #39977 opened Aug 7, 2025

• bug in new transformers: 'Florence2ForConditionalGeneration' object has no attribute '_supports_sdpa'

  #39974 opened Aug 7, 2025

• Gemma3 with fp16 in inference (I don't know if this change is working in fine-tune) #BUG FIX

  #39972 opened Aug 6, 2025

• change `dataloader_persistent_workers` default value to `True`

  #39963 opened Aug 6, 2025

• Calling `trainer.evaluate()` before `trainer.train()` with FSDP 2 raises `ValueError: When using FSDP2, a model and optimizer must be passed together to `Accelerator.prepare()...`

  #39961 opened Aug 6, 2025

• TypeError: Received a NoneType for argument video_processor, but a BaseVideoProcessor was expected.(this issue im getting when using doc-ocr)

  #39958 opened Aug 6, 2025

• Retaining computational graph after using AutoImageProcessor

  #39946 opened Aug 6, 2025

• GPT-OSS mxfp4 with triton_kernel: make_default_matmul_mxfp4_w_layout not found

  #39945 opened Aug 6, 2025

• Breaking change in unset `_tp_plan` attribute

  #39943 opened Aug 6, 2025

• Still getting "fp16 mixed precision requires a GPU (not 'mps')." error

  #39935 opened Aug 5, 2025

• [Gemma3N] Not able to add new special tokens to model/tokenizer due to projection error

  #39921 opened Aug 5, 2025

• When using batch_eval_metrics, inputs are not gathered from different device, which is wrong behavior

  #39916 opened Aug 5, 2025

• Idefics 3: shape mismatch: value tensor of shape [256, 576] cannot be broadcast to indexing result of shape [192, 576]

  #39914 opened Aug 5, 2025

• Question: Llama4 weight reshaping

  #39910 opened Aug 5, 2025

• Hidden torchvision>=0.19.0 dependency results in quiet import failures of e.g. PreTrainedModel

  #39907 opened Aug 5, 2025

• v4.54.1 average_tokens_across_devices=True would cause "ValueError: Tensors must be CUDA and dense" when gathering num_items_in_batch

  #39896 opened Aug 4, 2025

• Add VideoPrism

  #39893 opened Aug 4, 2025

• [Feature Request] Automatically log parallelism configuration from Accelerate to W&B

  #39882 opened Aug 4, 2025

121 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 Segment Anything 2 (SAM2)

  #32317 commented on Aug 8, 2025 • 43 new comments

• Add support for Florence-2

  #38188 commented on Aug 7, 2025 • 27 new comments

• [Feat] Adding Intern-S1

  #39722 commented on Aug 7, 2025 • 22 new comments

• blt wip

  #38579 commented on Aug 8, 2025 • 21 new comments

• [video processors] decode only sampled videos -> less RAM and faster processing

  #39600 commented on Aug 8, 2025 • 16 new comments

• HunYuan opensource

  #39606 commented on Aug 8, 2025 • 13 new comments

• Update model card for gpt neox japanese

  #39862 commented on Aug 5, 2025 • 12 new comments

• [WIP] Computer vision util: vision visualizer

  #36892 commented on Aug 10, 2025 • 9 new comments

• Mistral: Add support for interleaved attention

  #39799 commented on Aug 8, 2025 • 9 new comments

• Fix missing initializations for models created in 2022

  #39772 commented on Aug 8, 2025 • 6 new comments

• support MiniCPM-o2.6

  #37917 commented on Aug 7, 2025 • 6 new comments

• Improve Gemma3n model and tests

  #39764 commented on Aug 6, 2025 • 5 new comments

• [WIP] RoPE refactor

  #39847 commented on Aug 8, 2025 • 4 new comments

• Refactor vit-like models

  #39816 commented on Aug 5, 2025 • 3 new comments

• README: Update Bert Japanese model card

  #39466 commented on Aug 6, 2025 • 3 new comments

• 🌐 [i18n-KO] Translated grounding-dino.md to Korean

  #39861 commented on Aug 9, 2025 • 3 new comments

• Fix issue #39191 respect accelerate config to disable torch.dynamo compilation

  #39683 commented on Aug 6, 2025 • 2 new comments

• Fix rope_deltas corruption in Qwen2.5VL during CFG generation

  #39756 commented on Aug 4, 2025 • 2 new comments

• 🌐 [i18n-KO] Translated `main_classes/processors.md` to Korean

  #39519 commented on Aug 10, 2025 • 2 new comments

• Add Whole Word Masking and Padding Strategy to DataCollatorForLanguageModeling

  #39485 commented on Aug 8, 2025 • 2 new comments

• 🌐 [i18n-KO] Translated `main_classes/backbones.md` to Korean

  #39714 commented on Aug 9, 2025 • 1 new comment

• [Tests] [Bugfix] Make weights tied for `dynamic_tied_weights` test

  #39740 commented on Aug 5, 2025 • 1 new comment

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

  #39577 commented on Aug 5, 2025 • 1 new comment

• Feat: add Kwai-Keye transformers

  #39292 commented on Aug 10, 2025 • 1 new comment

• fix bug when using DP in trl, the batch size of input and output dism…

  #38938 commented on Aug 8, 2025 • 1 new comment

• Enable SIM rules

  #39806 commented on Aug 6, 2025 • 1 new comment

• 🌐 [i18n-KO] Translated `gemma3.md` to Korean

  #39865 commented on Aug 6, 2025 • 1 new comment

• [Draft] Add Llasa TTS family of models

  #39760 commented on Aug 4, 2025 • 1 new comment

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

  #39863 commented on Aug 10, 2025 • 1 new comment

• handle multimodal models with tp_plan on the text_config

  #39735 commented on Aug 4, 2025 • 0 new comments

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

  #39518 commented on Aug 4, 2025 • 0 new comments

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

  #39517 commented on Aug 5, 2025 • 0 new comments

• Add model arcinstitute state

  #39480 commented on Aug 7, 2025 • 0 new comments

• refactor(modeling_llama): make RotaryEmbedding default path explicit

  #39831 commented on Aug 4, 2025 • 0 new comments

• Add a unit test for BartModel to compare eager, sdpa on one particular set of inputs

  #39435 commented on Aug 7, 2025 • 0 new comments

• Fix inconsistency in SeamlessM4T and SeamlessM4Tv2 docs

  #39364 commented on Aug 7, 2025 • 0 new comments

• Fix DeepSpeed mixed precision precedence over Accelerate defaults

  #39856 commented on Aug 10, 2025 • 0 new comments

• WIP: Initial support for bnb 4bit on any nn.Parameter

  #39859 commented on Aug 7, 2025 • 0 new comments

• Fix the issue that csm model cannot work with pipeline mode.

  #39349 commented on Aug 8, 2025 • 0 new comments

• feat(trainer): emergency checkpointing on crashes & SIGTERM/SIGINT

  #39140 commented on Aug 5, 2025 • 0 new comments

• Update Dockerfiles to install packages inside a virtual environment

  #39098 commented on Aug 4, 2025 • 0 new comments

• Update README.md

  #39869 commented on Aug 6, 2025 • 0 new comments

• Allow compression on meta device

  #39039 commented on Aug 8, 2025 • 0 new comments

• fix: Catch correct ConnectionError for additional_chat_templates

  #39874 commented on Aug 9, 2025 • 0 new comments

• FP-Quant NVFP4 and Python 3.9 support

  #39876 commented on Aug 9, 2025 • 0 new comments

• [qwen-vl] fix beam search with videos

  #39726 commented on Aug 4, 2025 • 0 new comments

• Fix HfArgumentParser to filter out dict types from Union

  #39741 commented on Aug 5, 2025 • 0 new comments

• Audio encodings now match conv2d weight dtype in Gemma3nAudioSSCPConvBlock

  #39743 commented on Aug 5, 2025 • 0 new comments

• 🌐 [i18n-KO] Translated `main_classes/optimizer_schedules.md` to Korean

  #39713 commented on Aug 10, 2025 • 0 new comments

• use untyped storage for dtensors due to deprecation

  #39697 commented on Aug 4, 2025 • 0 new comments

• 🌐 [i18n-KO] Translated `deepseek_v3.md` to Korean

  #39649 commented on Aug 10, 2025 • 0 new comments

• fix tensor device when loading state dict

  #39623 commented on Aug 8, 2025 • 0 new comments

• Stop using `from_legacy_cache` as Cache initialization

  #39765 commented on Aug 6, 2025 • 0 new comments

• feat: add `is_fast` to ImageProcessor

  #39603 commented on Aug 6, 2025 • 0 new comments

• [gemma3] update conversion key mapping

  #39778 commented on Aug 4, 2025 • 0 new comments

• Add Fast Image Processor for ImageGPT

  #39592 commented on Aug 7, 2025 • 0 new comments

• Use `dtype` instead of `torch_dtype` everywhere!

  #39782 commented on Aug 7, 2025 • 0 new comments

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

  #39571 commented on Aug 9, 2025 • 0 new comments

• 🌐 [i18n-KO] Translated `vision-encoder-decoder.md` to Korean

  #39563 commented on Aug 8, 2025 • 0 new comments

• 🌐 [i18n-KO] Translated `main_classes/deepspeed.md` to Korean

  #39559 commented on Aug 10, 2025 • 0 new comments

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

  #39557 commented on Aug 10, 2025 • 0 new comments

• Add Muon optimizer implementation and integration

  #39541 commented on Aug 8, 2025 • 0 new comments

• Served models handle with nested content

  #39792 commented on Aug 5, 2025 • 0 new comments

• hangs during training using deepspeed

  #39275 commented on Aug 8, 2025 • 0 new comments

• Support for context-free-grammars (CFG) to constrain model output

  #25778 commented on Aug 7, 2025 • 0 new comments

• [Contributions Welcome] Add Fast Image Processors

  #36978 commented on Aug 7, 2025 • 0 new comments

• [RFC] Updating pipeline models

  #26690 commented on Aug 7, 2025 • 0 new comments

• Loading audio in video from video URLs fail with chat template

  #39076 commented on Aug 7, 2025 • 0 new comments

• ImportError: cannot import name 'pipeline' from 'transformers'

  #39137 commented on Aug 6, 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 Aug 6, 2025 • 0 new comments

• transformers env fails with: ModuleNotFoundError: No module named 'PIL'

  #39779 commented on Aug 5, 2025 • 0 new comments

• We now require users to upgrade torch to at least v2.6 in order to use the function.

  #38464 commented on Aug 5, 2025 • 0 new comments

• MistralCommonTokenizer does not match PreTrainedTokenizer

  #39841 commented on Aug 5, 2025 • 0 new comments

• Support topNSigma sampling in `generate`

  #39850 commented on Aug 5, 2025 • 0 new comments

• ValueError: Number of image placeholders in the prompt does not match the number of images. internVL3

  #39703 commented on Aug 5, 2025 • 0 new comments

• InternVL, PerceptionLM inference freeze in 4.54.1

  #39872 commented on Aug 5, 2025 • 0 new comments

• [FEAT] [non-CUDA]: Support alternative implementation for `constraints.positive_definite.check`

  #36660 commented on Aug 5, 2025 • 0 new comments

• Torch patches tracker for HPU/Gaudi

  #39175 commented on Aug 5, 2025 • 0 new comments

• Missing einops dependency causing ModuleNotFoundError

  #39811 commented on Aug 5, 2025 • 0 new comments

• Regression - High memory usage when using transformers model with FSDP + LoRA

  #39795 commented on Aug 5, 2025 • 0 new comments

• Accelerate beam search decoding via tree attention

  #39682 commented on Aug 5, 2025 • 0 new comments

• Failed to export PyTorch traced graph of Mixtral-8x7B-Instruct-v0.1 due to the PR #32429

  #38518 commented on Aug 4, 2025 • 0 new comments

• Support for per-token latency tracking in `generate()` (suggested options: using callback, profiler class, or using a config flag)

  #39437 commented on Aug 4, 2025 • 0 new comments

• T5Gemma failing on provided example

  #39522 commented on Aug 4, 2025 • 0 new comments

• Tensor parallelism for GLM-4.5

  #39868 commented on Aug 4, 2025 • 0 new comments

• Florence2ForConditionalGeneration does not support Flash Attention 2.0 yet ?...

  #39860 commented on Aug 4, 2025 • 0 new comments

• Handling of full_text_row_masked_out_mask in mllama is incorrect.

  #39379 commented on Aug 4, 2025 • 0 new comments

• transformers: FlaubertTokenizer: do_lowercase_and_remove_accent: make the logger warning actionable (don't only tell what's wrong, rather suggest what could be done about that)

  #39224 commented on Aug 4, 2025 • 0 new comments

• Exporting Llava decoder into ONNX format

  #38924 commented on Aug 4, 2025 • 0 new comments

• torch fake_tensor load hf model failed

  #39217 commented on Aug 4, 2025 • 0 new comments

• v4.53.0 - Qwen 2.5 VL Flash Attention error - object has no attribute is_causal

  #39231 commented on Aug 4, 2025 • 0 new comments

• Memory leak occurred during training qwen-2.5-vl

  #39803 commented on Aug 4, 2025 • 0 new comments

• Adds Universal Intelligence to awesome transformers documentation

  #38641 commented on Aug 7, 2025 • 0 new comments

• Add Bagel

  #38569 commented on Aug 10, 2025 • 0 new comments

• [omni modality] support composite processor config

  #38142 commented on Aug 5, 2025 • 0 new comments

• fix: qwen2.5 omni apply_chat_template system content check

  #37511 commented on Aug 8, 2025 • 0 new comments

• Add Plain-DETR

  #37096 commented on Aug 10, 2025 • 0 new comments

• Add Ovis2 model and processor implementation

  #37088 commented on Aug 8, 2025 • 0 new comments

• [Community contributions] Model cards

  #36979 commented on Aug 10, 2025 • 0 new comments

• Please support GGUF format for UMT5EncoderModel

  #36774 commented on Aug 10, 2025 • 0 new comments

• "pipeline" is not exported from module "transformers"

  #37646 commented on Aug 10, 2025 • 0 new comments

• YaRN: factor is not effective with original_max_position_embeddings

  #38224 commented on Aug 10, 2025 • 0 new comments

• Attention refactor in #35235 adds a `__getitem__` into the forward pass, which causes errors with torch dynamo.

  #38271 commented on Aug 10, 2025 • 0 new comments

• Potential Memory Leak or Caching in Fast Image Processor

  #38656 commented on Aug 10, 2025 • 0 new comments

• CPMANT Model Fails to Run Following Official Tutorial

  #39026 commented on Aug 10, 2025 • 0 new comments

• AutoModelForCausalLM.from_pretrained(..., device_map=...) ignore `Tensor.retain_grad()` in Multi-GPUs setting

  #39036 commented on Aug 10, 2025 • 0 new comments

• QWEN2VLProcessor missing video_token_id in mm_token_type_ids

  #39112 commented on Aug 10, 2025 • 0 new comments

• bf16_full_eval=True moves model to device before FSDP application and causes cuda OOM

  #39136 commented on Aug 10, 2025 • 0 new comments

• apply_rotary_pos_emb_flashatt failed during triton jit compilation 'constexpr' object has no attribute 'bit_length'

  #39167 commented on Aug 10, 2025 • 0 new comments

• [Trainer] Eval loss depends on batch size (with solution)

  #39241 commented on Aug 10, 2025 • 0 new comments

• TypeError: GenerationMixin._extract_past_from_model_output() got an unexpected keyword argument 'standardize_cache_format'

  #39336 commented on Aug 10, 2025 • 0 new comments

• `MoshiIntegrationTests` started to fail after #34464

  #38725 commented on Aug 9, 2025 • 0 new comments

• CUDA OOM when running meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8

  #37532 commented on Aug 9, 2025 • 0 new comments

• Improve CI/CD by completing migration from setup.py to pyproject.toml

  #38928 commented on Aug 9, 2025 • 0 new comments

• Qwen3 MOE models w/non-empty `mlp_only_layers` fail when `output_router_logits=True`

  #39203 commented on Aug 9, 2025 • 0 new comments

• Inference with model.generate( ) using a quantized model leads to assertion error

  #39311 commented on Aug 9, 2025 • 0 new comments

• Whisper demo code for model + processor API is broken

  #39318 commented on Aug 9, 2025 • 0 new comments

• Off-by-one error when using flash_attention with a sliding window

  #39408 commented on Aug 9, 2025 • 0 new comments

• Support `StaticCache` in assisted generation

  #32946 commented on Aug 8, 2025 • 0 new comments

• ValueError: You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time

  #39542 commented on Aug 8, 2025 • 0 new comments

• Please help i am trying to run model but issue

  #39260 commented on Aug 8, 2025 • 0 new comments