62 Pull requests merged by 38 people

• [FIX] Save speed metrics to logs

  #38136 merged May 15, 2025

• Omit creation of positional IDs within ESM if applicable

  #38089 merged May 15, 2025

• disable deepspeed when setting up fake trainer

  #38101 merged May 15, 2025

• enable trainer test cases on xpu

  #38138 merged May 15, 2025

• Hotfix: Flash Attention 2 support in Pixtral

  #38146 merged May 15, 2025

• [generate] Run custom generation code from the Hub

  #36405 merged May 15, 2025

• 🔴 Remove head mask in generative models

  #35786 merged May 15, 2025

• enable csm integration cases on xpu, all passed

  #38140 merged May 15, 2025

• [Qwen3] Qwen3 MoE add tp plan for expert mlps

  #38135 merged May 15, 2025

• Fix incorrect attention mask truncate in WhisperFlashAttention2

  #36477 merged May 14, 2025

• enable d_fine finetuning properly

  #37962 merged May 14, 2025

• Add my username to run_slow whitelist

  #38126 merged May 14, 2025

• [docs] add uv installation instructions for source builds

  #37968 merged May 14, 2025

• Update trainer.md

  #38113 merged May 14, 2025

• Add config validation and style tweaks

  #37589 merged May 14, 2025

• Fix auto batch size finder test

  #38125 merged May 14, 2025

• Fix temporal padding in Qwen2VLImageProcessor when the number of frames is not divisible by temporal_patch_size

  #38076 merged May 14, 2025

• [video processor] fix tests

  #38104 merged May 14, 2025

• enable finegrained_fp8 and granite_speech cases on XPU

  #38036 merged May 14, 2025

• Fix description and formatting errors in code docs

  #38074 merged May 13, 2025

• Add style bot

  #38102 merged May 13, 2025

• [CSM] update test for t4 runners

  #38110 merged May 13, 2025

• Add Fast Image Processor for vilt

  #37304 merged May 13, 2025

• Fix InternVL interpolate_pos_encoding and add to video_processing_auto

  #38092 merged May 13, 2025

• fix check_bad commit.py gives wrong results

  #38107 merged May 13, 2025

• [bug] fix llava processor to calculate unpadding size correctly

  #37988 merged May 13, 2025

• Fix past_key_values type hint in model output types

  #37953 merged May 13, 2025

• Fix bug in prefill_chunk_size that ignores disable_compile flag

  #38067 merged May 13, 2025

• [smolvlm] skip the test

  #38099 merged May 13, 2025

• Disable report callbacks for certain training tests

  #38088 merged May 13, 2025

• fix: Propagate lr_scheduler_kwargs options to create LR Scheduler when LayerWiseDummyOptimizer is used

  #34559 merged May 13, 2025

• add timeout for downloading the librispeech_asr dataset

  #38073 merged May 13, 2025

• update require_read_token

  #38093 merged May 13, 2025

• Refactor image processor phi4

  #36976 merged May 12, 2025

• uninstall kernels from docker images

  #38083 merged May 12, 2025

• update seed_worker to set seed based on worker_id and rank

  #37980 merged May 12, 2025

• Fix tot update in trainer

  #37923 merged May 12, 2025

• fix the inconsist docstring in apply_chat_template

  #38069 merged May 12, 2025

• chore(qwen2): display warning log only when sliding window attention …

  #36316 merged May 12, 2025

• Fix mt5 test on AMD devices

  #38081 merged May 12, 2025

• Add cuda graphs

  #38059 merged May 12, 2025

• docs: fix md style

  #38057 merged May 12, 2025

• Add AMD expectation to test_gpt2_sample

  #38079 merged May 12, 2025

• Fix OneFormer integration test

  #38016 merged May 12, 2025

• [chat] generate parameterization powered by GenerationConfig and UX-related changes

  #38047 merged May 12, 2025

• [VLM] fix loading issues

  #38051 merged May 12, 2025

• 🔴 Video processors as a separate class

  #35206 merged May 12, 2025

• fix(conversion): Fix size mismatch error during TF->PT model loading

  #38014 merged May 10, 2025

• enable generation fsdp/utils cases on XPU

  #38009 merged May 9, 2025

• Fix linalg.norm for CovnNextV2

  #38015 merged May 9, 2025

• Fix cache update!

  #38046 merged May 9, 2025

• Fix reduce-labels in BEIT Fast Image Processor

  #38042 merged May 9, 2025

• Re-Enable Trigger CircleCI via GitHub Actions when "ready for review" (#37885)

  #38041 merged May 9, 2025

• Support for version spec in requires & arbitrary mismatching depths across folders

  #37854 merged May 9, 2025

• Do not erase a cache_position passed explicitly to generate(), if there is one

  #37986 merged May 9, 2025

• Disable Trigger CircleCI via GitHub Actions when ready for review`

  #38038 merged May 9, 2025

• Trigger CircleCI via GitHub Actions when ready for review

  #37885 merged May 9, 2025

• [Temporary] Log some information in some pytest/pluggy internal places

  #37996 merged May 9, 2025

• enable utils test cases on XPU

  #38005 merged May 9, 2025

• make mistral3 pass on xpu

  #37882 merged May 9, 2025

• fix document masking for chunked attention

  #37429 merged May 9, 2025

• [AutoDocstring] Based on inspect parsing of the signature

  #33771 merged May 8, 2025

49 Pull requests opened by 38 people

• Update Loss Functions to Accept Tensor num_items_in_batch

  #38029 opened May 8, 2025

• Add `TemplateConstraint` and `OrdredConstraint` features (#27706)

  #38030 opened May 8, 2025

• check github actions 3

  #38044 opened May 9, 2025

• [fix] sliding window attention mask

  #38045 opened May 9, 2025

• Better pipeline type hints ✨

  #38049 opened May 9, 2025

• Handling Overlapping Annotations in Mask2Former by A Small Trick

  #38054 opened May 9, 2025

• SQuat cache implementation

  #38055 opened May 9, 2025

• [SAM-HQ] Update names in the docs

  #38058 opened May 10, 2025

• Improved cache docs

  #38060 opened May 10, 2025

• Fix broken example generation script for Llama3

  #38062 opened May 10, 2025

• Added scores in the streamer classes based on generation flag

  #38064 opened May 10, 2025

• Updated the Model docs - for the ALIGN model

  #38072 opened May 11, 2025

• Cache System Refactor: Layered Architecture

  #38077 opened May 12, 2025

• [gemma3] fix bidirectional attention mask

  #38080 opened May 12, 2025

• fix multi-image case for llava-onevision

  #38084 opened May 12, 2025

• Add CB

  #38085 opened May 12, 2025

• Refactor `MambaCache` to `modeling_mamba.py` (parity with Zamba)

  #38086 opened May 12, 2025

• Add optional RMSNorm support to BitNet quantization (config + layers)

  #38087 opened May 12, 2025

• Refactor `get_XXX_dataloader` from Trainer

  #38090 opened May 12, 2025

• In Llama4 fix wrongly inverted causal attention mask when using SDPA implementation

  #38094 opened May 12, 2025

• Fix amp deprecation issue

  #38100 opened May 13, 2025

• Fix incorrect batching audio index calculation for Phi-4-Multimodal

  #38103 opened May 13, 2025

• [video processors] support frame sampling within processors

  #38105 opened May 13, 2025

• Better typing in src/transformers/training_args.py

  #38106 opened May 13, 2025

• [`Attention`] Refactor Attention Interface for Bart-based Models and Enable Flex Attention

  #38108 opened May 13, 2025

• Support TP for save_pretrained()

  #38111 opened May 13, 2025

• Force real tensors and clone state_dict in src/transformers/modeling_utils.py

  #38114 opened May 13, 2025

• Minor llama4 fixes

  #38123 opened May 14, 2025

• [phi-4] add processor tests

  #38124 opened May 14, 2025

• [`compile`] re-enable for Qwen-VL models

  #38127 opened May 14, 2025

• 🚨🚨🚨 [pipelines] update defaults in pipelines that can `generate`

  #38129 opened May 14, 2025

• Make HF implementation match original OLMo 2 models for lower precisions

  #38131 opened May 14, 2025

• Add AMD MI300 CI caller leveraging self-hosted runner scale set workflow in hf-workflows

  #38132 opened May 14, 2025

• Skip non-selected experts for qwen3_moe

  #38133 opened May 15, 2025

• Fix error in calculating `cache_position` with past_length for Chatglm and Mamba model

  #38134 opened May 15, 2025

• Fix FSDP + llava-next/llava-onevision

  #38141 opened May 15, 2025

• [omni modality] support composite processor config

  #38142 opened May 15, 2025

• add dots1

  #38143 opened May 15, 2025

• [VLMs] add helpers for get/set embedding

  #38144 opened May 15, 2025

• remove unhandled parameter

  #38145 opened May 15, 2025

• [omni modality] support composite preprocessor config

  #38149 opened May 15, 2025

• Allow qwen/emu3 to process low res images

  #38150 opened May 15, 2025

• Fix Qwen2.5 Omni `SinusoidsPositionEmbedding` precision

  #38151 opened May 15, 2025

• Support for transformers explicit filename

  #38152 opened May 15, 2025

• [core] support tensor-valued _extra_state values in `from_pretrained`

  #38155 opened May 15, 2025

• Avoid incorrect generations for KV caches containing more than sliding_window tokens

  #38156 opened May 15, 2025

• Add Idefics2/3 and SmolVLM Fast image processors + improvements for fast image processors

  #38157 opened May 15, 2025

• [Examples] Add Comprehensive GPT2 vs DistilGPT2 Comparison with Perplexity and Benchmarks

  #38158 opened May 15, 2025

• Fix handling of slow/fast image processors in image_processing_auto.py

  #38161 opened May 15, 2025

48 Issues closed by 17 people

• Multiple processor classes have input side-effects

  #36865 closed May 15, 2025

• How to check the number of tokens processed or the load of each expert in the Qwen3 MoE model during inference?

  #38147 closed May 15, 2025

• transformers has no attribute TFFlorence2ForConditionalGeneration

  #37235 closed May 15, 2025

• Llama4TextExperts module implementation

  #37325 closed May 15, 2025

• BatchEncoding.to(device, dtype) could be worked!!

  #38096 closed May 15, 2025

• Tensor Parallelism with Quantized Models

  #38122 closed May 14, 2025

• Inconsistency in installation instructions for `venv` and `uv`

  #37956 closed May 14, 2025

• Potential bug in Qwen 2/2.5 VL Image Preprocessor

  #38003 closed May 14, 2025

• `output_hidden_states` only return part of hidden_state when setting `device_map="auto"`

  #36636 closed May 14, 2025

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

  #36845 closed May 14, 2025

• OSError: meta-llama/Llama-4-Scout-17B-16E-Instruct does not appear to have a file named X

  #37314 closed May 14, 2025

• while using trainer to train mnist model, 'ValueError: Found input variables with inconsistent numbers of samples: [10000, 8750]'

  #38024 closed May 14, 2025

• Versions greater than 4.49 are not compatible with Ascend NPU

  #37992 closed May 12, 2025

• Different DataLoader worker share the same seed and lost randomness

  #37932 closed May 12, 2025

• [Trainer] tot update steps is incorrect

  #37777 closed May 12, 2025

• transformers require torch >= 2.1.0 to run fp8 model, but im using 2.7.0

  #38034 closed May 12, 2025

• Add GPT-2-climate

  #20747 closed May 12, 2025

• Is there any plan to add kosmos-2 to the transformers.

  #24671 closed May 12, 2025

• Add MobileViT v2

  #22570 closed May 12, 2025

• [New model] RT-DETR

  #26742 closed May 12, 2025

• Typo in modeling_utils.py causing checkpoint loading error with Qwen2.5-VL

  #38070 closed May 12, 2025

• Qwen/Qwen2.5-VL-7B-Instruct not work [2025-05-10]

  #38056 closed May 12, 2025

• Video Processor as a separate class

  #33504 closed May 12, 2025

• Jitter Noise added to input being passed to experts in Switch Transformers

  #33969 closed May 12, 2025

• opencv imshow stuck forever when importing transformer

  #37239 closed May 12, 2025

• ed_video = input_tokens.index(video_token_id, st) ValueError: 151656 is not in list

  #37240 closed May 12, 2025

• TypeError: 'NoneType' object cannot be interpreted as an integer

  #37242 closed May 12, 2025

• Inconsistent results between torch and jax versions of DINOv2

  #37246 closed May 12, 2025

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

  #37248 closed May 12, 2025

• Same data but has large different while evaluating in the training stage vs evaluate it standalone from read the finetuned-model

  #37265 closed May 12, 2025

• RuntimeError when loading InternVL3-14B model: Embedding size mismatch

  #38033 closed May 12, 2025

• XLA FSDP V2 + TPU + T5 Family Models doesn't work

  #35142 closed May 11, 2025

• LayerDrop broken in various Flax models (Whisper/BART/more...)

  #35468 closed May 11, 2025

• llama code break with torch compile

  #36484 closed May 11, 2025

• a logic error in _preprocess function of Qwen2VLImageProcessor Class

  #37064 closed May 11, 2025

• Whether transformers Trainer support pipeline parallelism?

  #37129 closed May 11, 2025

• Quen FSDP model training hangs when some batches do not contain images

  #37186 closed May 11, 2025

• Bug when using StaticCache in Qwen2.5 Inference with custom inputs_embeds and attention_masks

  #37189 closed May 11, 2025

• Gemma3 Gradient Accumulation loss

  #37197 closed May 11, 2025

• torch.compile graph break when tuning llama with FA2

  #37199 closed May 11, 2025

• RWKV6-Finch-7B-HF crashes during inference

  #37221 closed May 11, 2025

• Why does `transformers` load FA2 when it's not asked to do so?

  #37227 closed May 11, 2025

• Request to add D-FINE

  #35283 closed May 11, 2025

• Loading a Pytorch model from a Tensorflow saved model doesn't work

  #37786 closed May 10, 2025

• Removing GenerateMixin inheritance from PreTrainedModel class results in Phi4 load fail

  #38050 closed May 10, 2025

• Performance degradation on certain vision models from v4.51.*

  #37748 closed May 9, 2025

• Swinv2Model reports an error when using the parameter use_obsolute_embeddings

  #37161 closed May 9, 2025

• qwen3-moe attention module is defined repeatedly.

  #37813 closed May 9, 2025

28 Issues opened by 25 people

• `tie_word_embeddings` not saved on customized model

  #38160 opened May 15, 2025

• Support `extra_state` attributes in from_pretrained

  #38154 opened May 15, 2025

• Have to import cv2 and pop up window frist, or else it stuck forever

  #38139 opened May 15, 2025

• Speed metrics are not logged

  #38137 opened May 15, 2025

• eval_loss not found when training a peft model using trainer.py / losses not retrieved from base model where appropriate

  #38130 opened May 14, 2025

• ImportError: cannot import name 'DataCollatorForCTCWithPadding' on macOS ARM64 with transformers==4.51.3 (Python 3.10/3.11) - Class missing from installed data_collator.py

  #38128 opened May 14, 2025

• Emu3 precision regression

  #38121 opened May 14, 2025

• phi-4-mm HF format

  #38120 opened May 14, 2025

• Unable to quantize a pretrained SegFormer-B0 to int8 using Quanto

  #38119 opened May 14, 2025

• Llama4 inference encounter unsupported op in dynamo ?

  #38118 opened May 14, 2025

• OLMo and OLMo 2 models do not match original models for low precisions

  #38117 opened May 14, 2025

• [Bug in Generate] 4.51.2 vs 4.46 Beam search results are sometimes different, not sure if beam search or T5 model change is the reason?

  #38116 opened May 13, 2025

• support static kv cache with torch.compile for qwen2vl

  #38115 opened May 13, 2025

• [Bug] Phi-4-multimodal audio processor failed to process multiple audios with close length

  #38098 opened May 13, 2025

• ImportError: cannot import name 'amp' from 'apex'

  #38095 opened May 13, 2025

• autoawq has been deprecated. Is it possible to support the use of llm-compresser as an alternative to autoawq

  #38078 opened May 12, 2025

• transformers showing decoder model architecture detected so padding should be left

  #38071 opened May 11, 2025

• `AutoModel.from_pretrained(...)` (with explicit `device_map` unset) fails under `with torch.device("meta")` with PyTorch 2.6.0 and 2.7.0

  #38066 opened May 10, 2025

• [minor] Protect against broken/missing torchvision installations and do not hard-fail at timm/torchvision import (many text models don't need any timm/torchvision as hard dependencies)

  #38065 opened May 10, 2025

• Adding native support to load GGUF models using transformers

  #38063 opened May 10, 2025

• Weights not initialized correctly when instantiating model with a pretrained backbone

  #38061 opened May 10, 2025

• Attention mask for multi-image input in gemma3

  #38053 opened May 9, 2025

• `.to` on a `PreTrainedModel` throws a Pyright type check error. What is the correct way to put a model to the device that does not throw type check errors?

  #38052 opened May 9, 2025

• Modernbert 3D attention mask

  #38040 opened May 9, 2025

• Trainer API doesnt stop after the training has been completed

  #38039 opened May 9, 2025

• [BUG] Setting average_tokens_across_devices to True caused an error because it attempted to gather CPU tensors using NCCL.

  #38035 opened May 9, 2025

• Removing the modification of loss value due to rounding off to 4 digits

  #38032 opened May 9, 2025

• bug in new prefill_chunk_size implementation

  #38028 opened May 8, 2025

112 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 Magma Agentic Model from Microsoft

  #37267 commented on May 13, 2025 • 64 new comments

• New cache tests and modular Hybrid Cache

  #37972 commented on May 14, 2025 • 25 new comments

• [core] Completely rewrite the masking logic for all attentions

  #37866 commented on May 15, 2025 • 18 new comments

• Add ColQwen2 to 🤗 transformers

  #35778 commented on May 13, 2025 • 14 new comments

• Bart: new cache format

  #35314 commented on May 15, 2025 • 12 new comments

• Superpoint fast image processor

  #37804 commented on May 15, 2025 • 8 new comments

• Add Fast Image Processor for mobileViT

  #37143 commented on May 12, 2025 • 8 new comments

• 36978 | Fast image processor for DPT model

  #37481 commented on May 15, 2025 • 7 new comments

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

  #36534 commented on May 15, 2025 • 6 new comments

• add profiler to trainer

  #37889 commented on May 15, 2025 • 6 new comments

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

  #36835 commented on May 15, 2025 • 4 new comments

• Support Kosmos-2.5

  #31711 commented on May 15, 2025 • 4 new comments

• Add args support for fast image processors

  #37018 commented on May 13, 2025 • 3 new comments

• Feat: save_pretrained for tensor parallel (and other parallelisms) models

  #37919 commented on May 15, 2025 • 3 new comments

• Fix Float64 RuntimeError on Integrated Graphics when using DirectML

  #37735 commented on May 12, 2025 • 2 new comments

• parallelism goes brrr

  #37877 commented on May 15, 2025 • 2 new comments

• Feat: add warnings for unused keys and rules in tensor parallel

  #37893 commented on May 15, 2025 • 2 new comments

• Adds use_repr to model_addition_debugger_context

  #37984 commented on May 12, 2025 • 2 new comments

• feat: support indivisible shards for TP model loading and TPlizing.

  #37220 commented on May 12, 2025 • 1 new comment

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

  #36132 commented on May 15, 2025 • 1 new comment

• add fast image processor nougat

  #37661 commented on May 13, 2025 • 1 new comment

• Translating model_doc/bert.md to Chinese

  #37806 commented on May 14, 2025 • 1 new comment

• make Llama4TextMoe forward more readable

  #37529 commented on May 12, 2025 • 0 new comments

• internalize build_inputs_with_special_tokens and prepare_for_model

  #37522 commented on May 15, 2025 • 0 new comments

• Docs: fix docstrings for Gemma3 modeling

  #37534 commented on May 9, 2025 • 0 new comments

• Add callback to monitor progress in whisper transcription

  #37483 commented on May 14, 2025 • 0 new comments

• Mllama fast image processor

  #37539 commented on May 15, 2025 • 0 new comments

• Inherited CausalLM Tests

  #37590 commented on May 15, 2025 • 0 new comments

• Fix interpolation of convnext image processor

  #37460 commented on May 14, 2025 • 0 new comments

• [Cache] Support compilable cache reuse with smaller batch sizes

  #37394 commented on May 15, 2025 • 0 new comments

• Add `segmentation_maps` support to MobileNetV2ImageProcessor

  #37312 commented on May 9, 2025 • 0 new comments

• Add Fast Image Processor for Chameleon

  #37140 commented on May 13, 2025 • 0 new comments

• Improve typing in TrainingArgument

  #36944 commented on May 13, 2025 • 0 new comments

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

  #36809 commented on May 13, 2025 • 0 new comments

• RuntimeError when converting and saving Flax ViT model to PyTorch

  #37999 commented on May 12, 2025 • 0 new comments

• Pass `eps` to `Mistral3RMSNorm`

  #38026 commented on May 13, 2025 • 0 new comments

• [ESM] Add flash-attention-2 backend for ESM-2

  #38023 commented on May 15, 2025 • 0 new comments

• Qwen2.5-Omni: Update modeling_qwen2_5_omni.py to fix error when loading quantized weights with AutoAWQ.

  #38013 commented on May 12, 2025 • 0 new comments

• proof of concept for using dataset of test cases for tokenizer tests

  #37994 commented on May 13, 2025 • 0 new comments

• update loss computation in modeling code

  #37993 commented on May 15, 2025 • 0 new comments

• CI result inspector util

  #37976 commented on May 14, 2025 • 0 new comments

• Include output embedding as well with `include_embedding` flag

  #37935 commented on May 13, 2025 • 0 new comments

• Fix wrong example in grounding dino

  #37921 commented on May 10, 2025 • 0 new comments

• support MiniCPM-o2.6

  #37917 commented on May 15, 2025 • 0 new comments

• Feat: Add class_proba option to semantic segmentation post-processing

  #37904 commented on May 13, 2025 • 0 new comments

• Get our efficiency back

  #37884 commented on May 9, 2025 • 0 new comments

• [WIP] Perception lm

  #37878 commented on May 14, 2025 • 0 new comments

• New bart model card

  #37858 commented on May 14, 2025 • 0 new comments

• Add z-loss to Bamba for v2

  #37842 commented on May 9, 2025 • 0 new comments

• Added False case implementation for config.do_stable_layer_norm in FlaxWav2vec2Models

  #37822 commented on May 15, 2025 • 0 new comments

• general spm converter

  #37763 commented on May 15, 2025 • 0 new comments

• Stop autoconverting custom code checkpoints

  #37751 commented on May 9, 2025 • 0 new comments

• [VLMs] add helpers to get multimodal encodings

  #37743 commented on May 9, 2025 • 0 new comments

• refactor can_save_slow_tokenizer

  #37722 commented on May 9, 2025 • 0 new comments

• :rotating_light: :rotating_light: Fix custom code saving

  #37716 commented on May 15, 2025 • 0 new comments

• Add support for manually setting `head_dim` in Qwen2 MoE

  #37643 commented on May 9, 2025 • 0 new comments

• Add time-based evaluation strategy to Trainer

  #37642 commented on May 9, 2025 • 0 new comments

• Add PLM Model

  #37634 commented on May 9, 2025 • 0 new comments

• [WiP] Add EoMT Model

  #37610 commented on May 12, 2025 • 0 new comments

• Trainer Stuck at 0% Progress during Training on Multi-GPU Setup

  #38008 commented on May 12, 2025 • 0 new comments

• Trainer.training_step incorrectly normalizes mean token loss when n_gpu > 1

  #37474 commented on May 12, 2025 • 0 new comments

• Community contribution: Adding GGUF support for more architectures

  #33260 commented on May 12, 2025 • 0 new comments

• How to solve the error of converting Qwen onnx_model to tensorRT_model?

  #37408 commented on May 12, 2025 • 0 new comments

• Loading HQQ quantized models is broken since #35926

  #37263 commented on May 12, 2025 • 0 new comments

• Support multimodal models in vLLM with transformers backend

  #37780 commented on May 12, 2025 • 0 new comments

• Model implementation with Transformers and Hugging face hub.

  #27532 commented on May 12, 2025 • 0 new comments

• how to fine tune TrOCR on specifique langage guide.

  #33106 commented on May 12, 2025 • 0 new comments

• Patches for different modalities

  #34585 commented on May 12, 2025 • 0 new comments

• Refactor bert-based models to use global attention function

  #37495 commented on May 12, 2025 • 0 new comments

• FileNotFoundError when using SentenceTransformerTrainingArguments(load_best_model_at_end=True) and Peft

  #34747 commented on May 12, 2025 • 0 new comments

• Inconsistent Documentation for `⁠dataset_index` Requirement Across ViTPose Models

  #36773 commented on May 12, 2025 • 0 new comments

• Since 4.50.0, saving and loading a Whisper model causes an error

  #37172 commented on May 12, 2025 • 0 new comments

• Issue: Unexpected Shape of logits: When Using generate() with num_return_sequences > 1

  #37378 commented on May 11, 2025 • 0 new comments

• ImportError: cannot import name '_flash_supports_window_size' from 'transformers.modeling_flash_attention_utils'

  #37428 commented on May 11, 2025 • 0 new comments

• facebook/opt-30b Cuda Allocation Error with version >= 4.50.0 code

  #37436 commented on May 11, 2025 • 0 new comments

• Processor multiprocessing error when load custom processor

  #37637 commented on May 10, 2025 • 0 new comments

• Make `argmax` in `post_process_semantic_segmentation` optional

  #37715 commented on May 10, 2025 • 0 new comments

• FP8 tensors not saved correctly

  #37250 commented on May 10, 2025 • 0 new comments

• TimeSformer assumes a fixed number of frames in its layers even though it interpolates temporal embeddings based on the input

  #38027 commented on May 10, 2025 • 0 new comments

• clarify the label shifting behavior of llama models when `labels` is given.

  #32944 commented on May 10, 2025 • 0 new comments

• A shallow copy in groundingdino

  #37333 commented on May 9, 2025 • 0 new comments

• Image Processor fails to process void segmentation maps

  #30064 commented on May 9, 2025 • 0 new comments

• Are there any plans to provide some performance analysis tools for transformers?

  #36360 commented on May 9, 2025 • 0 new comments

• Can't load Llama4 Processor

  #37375 commented on May 9, 2025 • 0 new comments

• Does Qwen_2_5_VL support variable length attention computation?

  #38007 commented on May 9, 2025 • 0 new comments

• [WIP] Add DINO DETR Model to HuggingFace Transformers

  #36711 commented on May 11, 2025 • 0 new comments

• Add Aimv2 model

  #36625 commented on May 13, 2025 • 0 new comments

• Add evolla rebase main

  #36232 commented on May 12, 2025 • 0 new comments

• [WIP] Add a dedicated tokenizer for byte level transformers

  #36216 commented on May 12, 2025 • 0 new comments

• [ModernBERT] Add CausalLM functionality to ModernBERT

  #35946 commented on May 14, 2025 • 0 new comments

• Add padding-free to bamba

  #35861 commented on May 12, 2025 • 0 new comments

• [Whisper] Pipeline: handle long form generation

  #35750 commented on May 9, 2025 • 0 new comments

• Integrate xlstm cleanly.

  #35377 commented on May 15, 2025 • 0 new comments

• [`ESM`] Add support for sdpa.

  #34954 commented on May 13, 2025 • 0 new comments

• Add Molmo (7B-D, 7B-O, 70B)

  #33962 commented on May 12, 2025 • 0 new comments

• Custom beam search scorer argument in generate function

  #32097 commented on May 14, 2025 • 0 new comments

• [Contributions Welcome] Add Fast Image Processors

  #36978 commented on May 15, 2025 • 0 new comments

• Cannot run backward with tensor parallel

  #36657 commented on May 15, 2025 • 0 new comments

• Why can't InternVL3-8B start vLLM after being converted to the Hugging Face format? It shows the error: `ValueError: 'limit_mm_per_prompt' is only supported for multimodal models.'

  #38000 commented on May 15, 2025 • 0 new comments

• Incorrect installation instructions

  #37476 commented on May 15, 2025 • 0 new comments

• 4.51.3 is much faster than prevous version - do you see the same?

  #37504 commented on May 15, 2025 • 0 new comments

• Trainer num_tokens() function seem to be outdated and not correct

  #37510 commented on May 15, 2025 • 0 new comments

• Wrong KV cache update for sliding-window attention (SWA) layers when total sequence length reaches window size

  #37574 commented on May 15, 2025 • 0 new comments

• [Community contributions] Model cards

  #36979 commented on May 15, 2025 • 0 new comments

• Qwen2vl support for GGUF

  #35282 commented on May 14, 2025 • 0 new comments

• The "force_words_ids" does not seem to be available on llama4

  #37478 commented on May 14, 2025 • 0 new comments

• Convnext image preprocessor raises an AssertionError when comparing logits

  #37461 commented on May 13, 2025 • 0 new comments

• `last_cache_position` definition issue in hybrid SWA models

  #37706 commented on May 13, 2025 • 0 new comments

• FSDP Torch XLA vs. FSDPv2 (SMPD) Torch XLA checkpoint saving bug

  #36004 commented on May 13, 2025 • 0 new comments

• Whisper word-level timestamp extraction fails with beam search

  #36093 commented on May 13, 2025 • 0 new comments

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

  #37423 commented on May 13, 2025 • 0 new comments

• Broken phi4 model

  #37464 commented on May 13, 2025 • 0 new comments