1 Release published by 1 person

• v4.48.1 Patch release v4.48.1

  published Jan 20, 2025

65 Pull requests merged by 39 people

• Add Rocketknight1 to self-comment-ci.yml

  #35881 merged Jan 24, 2025

• add xpu device check in device_placement

  #35865 merged Jan 24, 2025

• use torch.testing.assertclose instead to get more details about error in cis

  #35659 merged Jan 24, 2025

• Fix Llava-NeXT / Llava-NeXT Video / Llava-OneVision's token unpadding mismatch

  #35779 merged Jan 24, 2025

• Fix test_pipelines_video_classification that was always failing

  #35842 merged Jan 23, 2025

• fix apply_chat_template() padding choice

  #35828 merged Jan 23, 2025

• Fix typo

  #35854 merged Jan 23, 2025

• [DOC] Fix contamination and missing paragraph in translation

  #35851 merged Jan 23, 2025

• Granite Vision Support

  #35579 merged Jan 23, 2025

• Fix more CI tests

  #35661 merged Jan 23, 2025

• Fix uploading processors/tokenizers to WandB on train end

  #35701 merged Jan 23, 2025

• Fix GA loss for Deepspeed

  #35808 merged Jan 23, 2025

• add qwen2.5vl

  #35569 merged Jan 23, 2025

• [Backend support] Allow num_logits_to_keep as Tensor and change it to logits_to_keep + add flag

  #35757 merged Jan 23, 2025

• [ tests] remove some flash attention class tests

  #35817 merged Jan 23, 2025

• Fix NoneType type as it requires py>=3.10

  #35843 merged Jan 22, 2025

• Add PyTorch version check for FA backend on AMD GPUs

  #35813 merged Jan 22, 2025

• Fix compatibility issues when using auto_gptq with these older versions

  #35830 merged Jan 22, 2025

• [chat] docs fix

  #35840 merged Jan 22, 2025

• Fix head_dim in config extracted from Gemma2 GGUF model

  #35818 merged Jan 22, 2025

• [Chat] Add Chat from TRL 🐈

  #35714 merged Jan 22, 2025

• Fix : Nemotron tokenizer for GGUF format

  #35836 merged Jan 22, 2025

• [pipeline] missing import regarding assisted generation

  #35752 merged Jan 22, 2025

• [gpt2] fix generation tests

  #35822 merged Jan 22, 2025

• Hotfix: missing working-directory in self-comment-ci.yml

  #35833 merged Jan 22, 2025

• Init cache on meta device

  #35164 merged Jan 22, 2025

• Another security patch for self-comment-ci.yml

  #35816 merged Jan 22, 2025

• Remove pyav pin to allow python 3.11 to be used

  #35823 merged Jan 21, 2025

• Remove old benchmark code

  #35730 merged Jan 21, 2025

• [Mimi] update test expected values for t4 runners

  #35696 merged Jan 21, 2025

• Improve modular documentation

  #35737 merged Jan 21, 2025

• add Qwen2-VL image processor fast

  #35733 merged Jan 21, 2025

• move fastspeech to audio models

  #35788 merged Jan 21, 2025

• [i18n-ar] Translated file: docs/source/ar/tasks/masked_language_modeling.md into Arabic

  #35198 merged Jan 21, 2025

• Optimized set_initialized_submodules.

  #35493 merged Jan 21, 2025

• Remove deprecated get_cached_models

  #35809 merged Jan 21, 2025

• Fixed typo in autoawq version number in an error message for IPEX backend requirements.

  #35815 merged Jan 21, 2025

• Fix : BLOOM tie_word_embeddings in GGUF

  #35812 merged Jan 21, 2025

• Auto-add timm tag to timm-wrapper models.

  #35794 merged Jan 21, 2025

• Support adamw_torch_8bit

  #34993 merged Jan 21, 2025

• add a new flax example for Bert model inference

  #34794 merged Jan 21, 2025

• [Doc] Adding blog post to model doc for TimmWrapper

  #35744 merged Jan 21, 2025

• Byebye test_batching_equivalence's flakiness

  #35729 merged Jan 21, 2025

• Add LlavaImageProcessor

  #33191 merged Jan 21, 2025

• Update AMD Docker image

  #35804 merged Jan 21, 2025

• Fix "test_chat_template_dict" in video LLMs

  #35660 merged Jan 21, 2025

• Deterministic sorting in modular converter when adding new functions

  #35795 merged Jan 21, 2025

• modular_model_converter bugfix on assignments

  #35642 merged Jan 21, 2025

• Fixes, improvements to timm import behaviour

  #35800 merged Jan 20, 2025

• Tool calling: support more types

  #35776 merged Jan 20, 2025

• fix low-precision audio classification pipeline

  #35435 merged Jan 20, 2025

• Fix vits low-precision dtype

  #35418 merged Jan 20, 2025

• fix document qa bf16 pipeline

  #35456 merged Jan 20, 2025

• Don't import torch.distributed when it's not available

  #35777 merged Jan 20, 2025

• Patch moonshine

  #35731 merged Jan 20, 2025

• transformers.image_transforms.normalize wrong types

  #35773 merged Jan 20, 2025

• [fix] cannot import name 'Pop2PianoFeatureExtractor' from 'transformers'

  #35604 merged Jan 20, 2025

• Skip Falcon 7B GGML Test

  #35783 merged Jan 20, 2025

• remove code owners as it was generating too much noise BUT

  #35784 merged Jan 20, 2025

• [generate] update docstring of SequenceBiasLogitsProcessor

  #35699 merged Jan 20, 2025

• fix register_buffer in MimiEuclideanCodebook

  #35759 merged Jan 20, 2025

• Add SuperGlue model

  #29886 merged Jan 20, 2025

• [ViTPose] Convert more checkpoints

  #35638 merged Jan 20, 2025

• Security fix for self-comment-ci.yml

  #35548 merged Jan 20, 2025

• Fix CI for VLMs

  #35690 merged Jan 20, 2025

44 Pull requests opened by 29 people

• Fixed ViTMAE for non-square images

  #35769 opened Jan 19, 2025

• Fix Audio Classification Pipeline top_k Documentation Mismatch and Bug #35736

  #35771 opened Jan 19, 2025

• Add ColQwen2 to 🤗 transformers

  #35778 opened Jan 19, 2025

• feat: Add gradient testing for Flash Attention 2

  #35780 opened Jan 20, 2025

• [tests] further fix "Tester object has no attribute '_testMethodName'"

  #35781 opened Jan 20, 2025

• Remove head mask in generative models

  #35786 opened Jan 20, 2025

• Add StyleTTS 2

  #35790 opened Jan 20, 2025

• Fix is_causal being a tensor

  #35791 opened Jan 20, 2025

• Fix docstring for get_candidates return shape

  #35793 opened Jan 20, 2025

• Fix IsADirectoryError in notebook progress bar display methods. Fixes #34766

  #35796 opened Jan 20, 2025

• Fix StopStringCriteria to handle tokens above len(tokenizer)

  #35797 opened Jan 20, 2025

• Fix AutoProcessor import order issue with custom classes

  #35798 opened Jan 20, 2025

• Fix Noise Computation for NEFTune During Packed Training #34659

  #35799 opened Jan 20, 2025

• [generate] WIP vectorized beam search

  #35802 opened Jan 20, 2025

• Remove cache migration script

  #35810 opened Jan 21, 2025

• Exploring use of kwargs for timm model and transforms creation

  #35819 opened Jan 21, 2025

• [docs] uv install

  #35821 opened Jan 21, 2025

• Add MultipleChoice & QuestionAnswering heads to ModernBERT

  #35825 opened Jan 22, 2025

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

  #35831 opened Jan 22, 2025

• fix(FA): QKV not being casted to target_dtype for FA with dpo lora

  #35834 opened Jan 22, 2025

• Fix PretrainedTokenizerFast check

  #35835 opened Jan 22, 2025

• Optimize Qwen2VL vision model by precomputing cos/sin embeds before ViT blocks

  #35837 opened Jan 22, 2025

• Update modeling_attn_mask_utils.py

  #35841 opened Jan 22, 2025

• Update-tp

  #35844 opened Jan 22, 2025

• Nail in edge case of torch dtype being overriden permantly in the case of an error

  #35845 opened Jan 22, 2025

• Github action for auto-assigning reviewers

  #35846 opened Jan 22, 2025

• Fix Jitter Noise Passing to Experts in Switch Transformers #33969

  #35847 opened Jan 22, 2025

• 🚨🚨🚨 image-classification pipeline single-label and multi-label prob type squashing fns (sigmoid vs softmax) are backwards

  #35848 opened Jan 22, 2025

• Whisper: fix static cache CI

  #35852 opened Jan 23, 2025

• Add utility for Reload Transformers imports cache for development workflow #35508

  #35858 opened Jan 23, 2025

• Fix PaliGemma Pad Token Masking During Training #35855

  #35859 opened Jan 23, 2025

• Fix TP initialization

  #35860 opened Jan 23, 2025

• Add padding-free to bamba

  #35861 opened Jan 23, 2025

• [doctest] Fixes

  #35863 opened Jan 23, 2025

• Add Tensor Parallel support for Gemma

  #35864 opened Jan 23, 2025

• Fix device mismatch error in Whisper model during feature extraction

  #35866 opened Jan 24, 2025

• [docs] no hard coding cuda as bnb has multi-backend support

  #35867 opened Jan 24, 2025

• [docs] fix bugs in the bitsandbytes documentation

  #35868 opened Jan 24, 2025

• Add default TP plan for all models with backend support

  #35870 opened Jan 24, 2025

• fix gemma that needed kwargs

  #35871 opened Jan 24, 2025

• Fix lost loss values when using user-defined compute_loss_func in some cases

  #35872 opened Jan 24, 2025

• Make cache traceable

  #35873 opened Jan 24, 2025

• Fix model kwargs

  #35875 opened Jan 24, 2025

• Fix XGLM loss computation (PyTorch and TensorFlow)

  #35878 opened Jan 24, 2025

42 Issues closed by 16 people

• ModernBERT fails to work without FlashAttention !

  #35879 closed Jan 24, 2025

• Some Whisper beam search output (sequences_scores, etc.) is lost in _stack_split_outputs

  #32373 closed Jan 24, 2025

• LLaVA-OneVision image features and image tokens mismatch

  #35775 closed Jan 24, 2025

• [Question] Why doesn't `trainer.state.epoch` fall round after training?

  #35298 closed Jan 24, 2025

• DPO LoRA loss incorrect with Gradient Accumulation in 4.48.1

  #35856 closed Jan 23, 2025

• set_initialized_submodules too slow when loading big model like DeepSeekV3

  #35635 closed Jan 23, 2025

• How to load a model directly into the GPU memory？

  #35853 closed Jan 23, 2025

• A bug that may cause device inconsistency

  #31930 closed Jan 23, 2025

• Any plans to integrate GTE model natively into transformers

  #35568 closed Jan 23, 2025

• Significant Increase in Training Loss after Upgrading from Transformers 4.47.1 to 4.48.0

  #35787 closed Jan 23, 2025

• is it expected that the same token has two different token ids in T5TokenizerFast

  #35839 closed Jan 22, 2025

• Multi-GPU setup: indices should be either on cpu or on the same device as the indexed tensor (cuda:1)

  #33147 closed Jan 22, 2025

• (sort of) a bug with token offsets: some special tokens have (0, 0) offsets regardless of their position in the document

  #35125 closed Jan 22, 2025

• how to load the weight of decoder.embed_tokens.weight seperately from the shared weight?

  #35152 closed Jan 22, 2025

• Detokenization discrepancy with Llama3.1

  #35175 closed Jan 22, 2025

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

  #35233 closed Jan 22, 2025

• microsoft/Phi-3.5-mini-instruct not working with FA2 due to position_ids

  #35274 closed Jan 22, 2025

• Adding Audio-MAE Model for Underwater Audio to Hugging Face Transformers Library

  #35811 closed Jan 22, 2025

• Mismatch Between txt img_token and Image Count in Multimodal Processor Causes Debugging

  #35254 closed Jan 22, 2025

• Handle python 3.11 for huggingface["video"] and dependencies

  #35803 closed Jan 21, 2025

• image_transforms preprocess quite slow when run large image with qwen2vl

  #34272 closed Jan 21, 2025

• tokenizers v0.20 not supported

  #33528 closed Jan 21, 2025

• Links in release note are broken

  #35480 closed Jan 21, 2025

• Errors when returning the tensors with PyTorch using AutoImageProcessor

  #35806 closed Jan 21, 2025

• Mimic `adamw_torch_4bit` and have `adamw_torch_8bit`

  #34893 closed Jan 21, 2025

• Supporting Padding in llava processor

  #33175 closed Jan 21, 2025

• Issue with Idefics3 sample code

  #35369 closed Jan 21, 2025

• BarkProcessor voice_preset doesn't work

  #34634 closed Jan 21, 2025

• RuntimeError: "rshift_cuda" not implemented for 'Half'

  #35256 closed Jan 21, 2025

• transformers.image_transforms.normalize documents and checks for the wrong type for std and mean arguments

  #35772 closed Jan 20, 2025

• version 4.47.0 provides different generation results when using quantized awq model

  #35286 closed Jan 20, 2025

• inconsistent execution time

  #35265 closed Jan 20, 2025

• apply class transformers.SequenceBiasLogitsProcessor on Qwen model

  #35432 closed Jan 20, 2025

• Implement SuperPoint / SuperGlue

  #25489 closed Jan 20, 2025

• Aria processor does not work with images

  #35768 closed Jan 20, 2025

• Text Only input using LlaVa Next

  #35421 closed Jan 20, 2025

• Trying to train a model using automatic1111. Error - Exception training model: 'module 'transformers.integrations' has no attribute 'deepspeed''.

  #34427 closed Jan 20, 2025

• StopStringCriteria relies on `len(tokenizer)==model.config.vocab_size`, leading to index errors

  #35244 closed Jan 20, 2025

• is possible convert transforms tokenizers in sentence piece .model?

  #35538 closed Jan 20, 2025

• gradient calculation is not correct with gradient accumulation in LM training

  #35203 closed Jan 19, 2025

• run_mlm_flax on tpu v5-pods

  #35205 closed Jan 19, 2025

• How to convert my Mask2Former model (ResNet-50 backbone) to Hugging Face transformer

  #35186 closed Jan 18, 2025

27 Issues opened by 26 people

• Request to add Co-DETR

  #35882 opened Jan 24, 2025

• Mllama training via FSDP device and dtype misassignment

  #35880 opened Jan 24, 2025

• Mask2Former _init_weights

  #35877 opened Jan 24, 2025

• Support Shared Cache

  #35876 opened Jan 24, 2025

• ZeroShotClassificationArgumentHandler should be explicit it has a somewhat unsafe internal behaviour.

  #35874 opened Jan 24, 2025

• ERROR: Video features and Video Tokens do not match!!!

  #35869 opened Jan 24, 2025

• Adding special tokens by default

  #35862 opened Jan 23, 2025

• [DEV Testing] Issues with `test_modeling_common`

  #35857 opened Jan 23, 2025

• Paliegemma Pad Token not Masked

  #35855 opened Jan 23, 2025

• resume_from_checkpoint failed when using PEFT LORA

  #35850 opened Jan 23, 2025

• chane data

  #35849 opened Jan 23, 2025

• forward() got an unexpected keyword argument 'num_items_in_batch'

  #35838 opened Jan 22, 2025

• tokenizer_class: `LlamaTokenizerFast` becomes `LlamaTokenizer` after load + immediate save

  #35832 opened Jan 22, 2025

• ImportError: cannot import name 'NoneType' from 'types' on main in Python 3.9

  #35827 opened Jan 22, 2025

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

  #35826 opened Jan 22, 2025

• multi-gpu: test_model_parallel_beam_search tests fail with "IndexError: list index out of range"

  #35824 opened Jan 21, 2025

• convert_llama_weight_to_hf.py

  #35820 opened Jan 21, 2025

• [Feature Request] Support register customize quantization method out-of-tree

  #35814 opened Jan 21, 2025

• How to change data

  #35807 opened Jan 21, 2025

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

  #35805 opened Jan 21, 2025

• Ascend：Training not loaded into NPU

  #35785 opened Jan 20, 2025

• Auto-resume from checkpoint throws error if last checkpoint is incomplete

  #35782 opened Jan 20, 2025

• TPU Initialization Error with Transformers in Kaggle TPU VM v3-8

  #35774 opened Jan 19, 2025

• Mamba2 doesn't support Multi-GPU training (fast path)

  #35770 opened Jan 19, 2025

• Issue: Error with _eos_token_tensor when using Generator with GenerationMixin

  #35767 opened Jan 18, 2025

• Defining LLM Dataset types in Trainers or during Training Workflow

  #35766 opened Jan 18, 2025

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

  #35765 opened Jan 18, 2025

132 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 Zamba2

  #34517 commented on Jan 24, 2025 • 39 new comments

• Add support for Apple's Depth-Pro

  #34583 commented on Jan 24, 2025 • 20 new comments

• Adding RTDETRv2

  #34773 commented on Jan 22, 2025 • 17 new comments

• Add GOT-OCR 2.0 to Transformers

  #34721 commented on Jan 24, 2025 • 16 new comments

• Add LightGlue model

  #31718 commented on Jan 24, 2025 • 15 new comments

• support telechat2

  #35415 commented on Jan 23, 2025 • 13 new comments

• Pixtral: vectorize patch embeddings and enable tests

  #35122 commented on Jan 21, 2025 • 12 new comments

• Trainer Refactor: Part 1

  #35567 commented on Jan 23, 2025 • 11 new comments

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

  #35627 commented on Jan 22, 2025 • 10 new comments

• Universal Speculative Decoding `CandidateGenerator`

  #35029 commented on Jan 23, 2025 • 9 new comments

• Uniformize OwlViT and Owlv2 processors

  #35700 commented on Jan 24, 2025 • 8 new comments

• Save checkpoint to temporary directory to handle partial saves during failures

  #35580 commented on Jan 24, 2025 • 7 new comments

• Mask2former & Maskformer Fast Image Processor

  #35685 commented on Jan 23, 2025 • 7 new comments

• Add Prompt Depth Anything Model

  #35401 commented on Jan 22, 2025 • 7 new comments

• Add internlm3 dense

  #35694 commented on Jan 22, 2025 • 6 new comments

• Several fixes related to rotary position embeddings

  #35376 commented on Jan 22, 2025 • 6 new comments

• switch from `training_args.bin` `training_args.json`

  #35010 commented on Jan 21, 2025 • 6 new comments

• Integrate xlstm cleanly.

  #35377 commented on Jan 20, 2025 • 3 new comments

• Refactoring of ImageProcessorFast

  #35069 commented on Jan 22, 2025 • 3 new comments

• Bart: new cache format

  #35314 commented on Jan 23, 2025 • 3 new comments

• Remove _supports_static_cache = True for some model classes

  #34975 commented on Jan 23, 2025 • 3 new comments

• Continuous batching

  #35727 commented on Jan 23, 2025 • 2 new comments

• Make `output_dir` Optional in `TrainingArguments` #27866

  #35735 commented on Jan 20, 2025 • 2 new comments

• Add DAB-DETR Object detection/segmentation model

  #30803 commented on Jan 23, 2025 • 2 new comments

• [WiP] `GPT2Model` StaticCache support

  #35761 commented on Jan 23, 2025 • 2 new comments

• ModernBERT FlexAttention

  #35423 commented on Jan 21, 2025 • 1 new comment

• Fix hardcoded `float` dtypes in DeBERTa model, which caused multiple RuntimeErrors in `bfloat16`

  #35336 commented on Jan 24, 2025 • 1 new comment

• Adding FlexAttention Support for Qwen2 models

  #35155 commented on Jan 24, 2025 • 1 new comment

• Fix mask slicing for models with HybridCache

  #35681 commented on Jan 21, 2025 • 1 new comment

• Test: generate with `torch.compile(model.forward)` as a fast test

  #34544 commented on Jan 23, 2025 • 1 new comment

• fix immediate quantization of the first token in QuantizedCache

  #35760 commented on Jan 20, 2025 • 1 new comment

• add RAdamScheduleFree optimizer

  #35313 commented on Jan 20, 2025 • 0 new comments

• Add support for DeepSpeed sequence parallelism (Ulysses)

  #35301 commented on Jan 24, 2025 • 0 new comments

• 🔴 Video processors as a separate class

  #35206 commented on Jan 24, 2025 • 0 new comments

• Create Zero-Delay_QKV_Compression.md

  #35324 commented on Jan 23, 2025 • 0 new comments

• Create zero-delay-qkv-compression.py

  #35328 commented on Jan 23, 2025 • 0 new comments

• Create zero_delay_qkv_benchmark.py

  #35329 commented on Jan 23, 2025 • 0 new comments

• Samhq model addition

  #35147 commented on Jan 24, 2025 • 0 new comments

• Commont bot CI for other jobs (`generation` / `quantization`)

  #35341 commented on Jan 22, 2025 • 0 new comments

• Output dicts support in text generation pipeline

  #35092 commented on Jan 21, 2025 • 0 new comments

• Replace all torch.FloatTensor by torch.Tensor

  #35004 commented on Jan 24, 2025 • 0 new comments

• Fix Gemma2 dtype issue when storing weights in float16 precision

  #35398 commented on Jan 21, 2025 • 0 new comments

• Add timm_wrapper support to AutoFeatureExtractor

  #35764 commented on Jan 21, 2025 • 0 new comments

• multi-gpu: fix tensor device placements for various models

  #35763 commented on Jan 22, 2025 • 0 new comments

• Pipeline: fix unnecessary warnings

  #35753 commented on Jan 20, 2025 • 0 new comments

• [Whisper] Pipeline: handle long form generation

  #35750 commented on Jan 20, 2025 • 0 new comments

• More tensor parallel

  #35748 commented on Jan 23, 2025 • 0 new comments

• Fix multi gpu loss sync condition, add doc and test

  #35743 commented on Jan 21, 2025 • 0 new comments

• Fix: loading DBRX back from saved path

  #35728 commented on Jan 24, 2025 • 0 new comments

• VLM: compile compatibility

  #35724 commented on Jan 23, 2025 • 0 new comments

• Add mean_resizing for every VLMs' resizing_token_embeddings()

  #35717 commented on Jan 20, 2025 • 0 new comments

• Fix Aria CI and testing

  #35674 commented on Jan 20, 2025 • 0 new comments

• Add more rigerous non-slow grad accum tests

  #35668 commented on Jan 21, 2025 • 0 new comments

• Decompose chat template docs

  #35657 commented on Jan 22, 2025 • 0 new comments

• Fix tests for vision models

  #35654 commented on Jan 22, 2025 • 0 new comments

• Add XPU type for work-around -inf mask causing sdpa NaN issue in modeling files

  #35647 commented on Jan 24, 2025 • 0 new comments

• Guard against unset resolved_archive_file

  #35628 commented on Jan 23, 2025 • 0 new comments

• Uniformize LlavaNextVideoProcessor kwargs

  #35613 commented on Jan 20, 2025 • 0 new comments

• Fix the config class comparison for remote code models

  #35592 commented on Jan 23, 2025 • 0 new comments

• Support QuestionAnswering Module for ModernBert based models.

  #35566 commented on Jan 23, 2025 • 0 new comments

• BLIPs clean-up

  #35560 commented on Jan 20, 2025 • 0 new comments

• Add support for nested images to LLava and VipLLava

  #35558 commented on Jan 21, 2025 • 0 new comments

• Add support for 4D custom attention masks in GPT-2

  #35517 commented on Jan 19, 2025 • 0 new comments

• [tokenizer] fix llama tokenizer (slow)

  #35488 commented on Jan 23, 2025 • 0 new comments

• Add FAST

  #35476 commented on Jan 22, 2025 • 0 new comments

• Add SDPA support for LayoutLMv3 model

  #35469 commented on Jan 22, 2025 • 0 new comments

• Fix #35447 Tokenizer does not split text according to newly added input tokens

  #35455 commented on Jan 20, 2025 • 0 new comments

• Support constant lr with cooldown

  #35453 commented on Jan 24, 2025 • 0 new comments

• [WIP] Add support for flex attention (paged attention)

  #35419 commented on Jan 21, 2025 • 0 new comments

• Add D-FINE Model into Transformers

  #35400 commented on Jan 22, 2025 • 0 new comments

• Add support for MiniMax-Text-01 and MiniMax-VL-01 from MiniMaxAI

  #35710 commented on Jan 22, 2025 • 0 new comments

• multi-gpu: test_model_parallel_beam_search tests fail with "RuntimeError: Expected all tensors to be on the same device"

  #35762 commented on Jan 22, 2025 • 0 new comments

• Incorrect Whisper long-form decoding timestamps

  #31942 commented on Jan 22, 2025 • 0 new comments

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

  #34855 commented on Jan 22, 2025 • 0 new comments

• Unknown quantization type, got fp8

  #35471 commented on Jan 22, 2025 • 0 new comments

• Gemma2: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:7 and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

  #34706 commented on Jan 22, 2025 • 0 new comments

• Multiple training runs not working with deepspeed

  #35073 commented on Jan 22, 2025 • 0 new comments

• Add CLIP-ViP

  #22829 commented on Jan 21, 2025 • 0 new comments

• LlavaNextVideoProcessor -> TypeError: LlavaNextVideoProcessor.__call__() got an unexpected keyword argument 'legacy' (I have the fix)

  #35602 commented on Jan 21, 2025 • 0 new comments

• No use `no_sync` context manager when using gradient accumulation w/ deepspeed's zero stage 2 or 3 via `accelerate`

  #34984 commented on Jan 21, 2025 • 0 new comments

• `RuntimeError: self and mat2 must have the same dtype, but got Float and BFloat16` when training with `torch_compile`

  #35382 commented on Jan 21, 2025 • 0 new comments

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

  #35414 commented on Jan 21, 2025 • 0 new comments

• TypeError: Accelerator.__init__() got an unexpected keyword argument 'dispatch_batches'

  #34714 commented on Jan 21, 2025 • 0 new comments

• redirect logging output to `stdout` instead of `stderr`

  #34613 commented on Jan 21, 2025 • 0 new comments

• When set num_beams in GenerationConfig, stop_strings parameter has no effect

  #34843 commented on Jan 20, 2025 • 0 new comments

• Neftune computation is probably wrong with packed training

  #34659 commented on Jan 20, 2025 • 0 new comments

• Default value for mean_resizing in resize_token_embeddings should be False

  #35357 commented on Jan 20, 2025 • 0 new comments

• Qwen2VL exhibits significant performance differences under different attention implementations.

  #35749 commented on Jan 20, 2025 • 0 new comments

• Help Understanding Beam Search Scores in Hugging Face (LLaMA + LoRA)

  #35618 commented on Jan 20, 2025 • 0 new comments

• tokenizer decode decode with timestamp fails for extended vocabulary

  #35330 commented on Jan 20, 2025 • 0 new comments

• [Whisper] TypeError: '<=' not supported between instances of 'NoneType' and 'float'

  #33552 commented on Jan 20, 2025 • 0 new comments

• Model loaded with `PretrainedModel.from_pretrained` and `with torch.device("cuda"):` decorator leads to unexpected errors compared to `.to("cuda")`

  #35371 commented on Jan 20, 2025 • 0 new comments

• MultiModalityCausalLM does not support Flash Attention 2.0 yet

  #35383 commented on Jan 20, 2025 • 0 new comments

• FA2 support for Aria

  #35670 commented on Jan 20, 2025 • 0 new comments

• Object Detection Pipeline only outputs first element when batching

  #31356 commented on Jan 19, 2025 • 0 new comments

• Audio-Classification Pipeline top_k Documentation mismatch and bug (possibly generalizes to any classification pipelines)

  #35736 commented on Jan 19, 2025 • 0 new comments

• Maybe the way SequenceClassification Model calculates the last non-pad token is not reasonable.

  #35352 commented on Jan 19, 2025 • 0 new comments

• Tranformers documentation translation to Italian

  #17459 commented on Jan 18, 2025 • 0 new comments

• unable to convert llama 3.3 weights to hf.py

  #35326 commented on Jan 18, 2025 • 0 new comments

• MPI environment variables are not set.

  #35331 commented on Jan 18, 2025 • 0 new comments

• Default arguments in `DebertaConfig` disable relative attention, contrary to the docs and `deberta-base`

  #35335 commented on Jan 18, 2025 • 0 new comments

• Move `DataCollatorForMultipleChoice` from the docs to the package

  #34763 commented on Jan 24, 2025 • 0 new comments

• Support `return_tensors` in audio chat templates

  #34601 commented on Jan 20, 2025 • 0 new comments

• Handle num_items_in_batch in Mistral's forward

  #34576 commented on Jan 24, 2025 • 0 new comments

• fix: DataCollatorWithFlattening incompatible with Tensor input ids

  #34267 commented on Jan 23, 2025 • 0 new comments

• LLaVA-NeXT: add new model checkpoints

  #34195 commented on Jan 21, 2025 • 0 new comments

• feat: add support for tensor parallel using Pytorch

  #34194 commented on Jan 23, 2025 • 0 new comments

• Add TimesFM Time Series Forecasting Model

  #34082 commented on Jan 21, 2025 • 0 new comments

• fix: Updated BridgeTower Image processor

  #32384 commented on Jan 23, 2025 • 0 new comments

• Add Segment Anything 2 (SAM2)

  #32317 commented on Jan 21, 2025 • 0 new comments

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

  #32085 commented on Jan 22, 2025 • 0 new comments

• [docs] Redesign

  #31757 commented on Jan 23, 2025 • 0 new comments

• Support Kosmos-2.5

  #31711 commented on Jan 24, 2025 • 0 new comments

• RLE of SAM can't handle masks with no change

  #35664 commented on Jan 25, 2025 • 0 new comments

• AttributeError: 'Config' object has no attribute '_get_non_default_generation_parameters'

  #35543 commented on Jan 24, 2025 • 0 new comments

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

  #33969 commented on Jan 24, 2025 • 0 new comments

• BatchEncoding.to throws away columns silently, thus no way to pass non-tensor columns such as String in Trainer metric computation

  #34983 commented on Jan 24, 2025 • 0 new comments

• Allow passing 2D attention mask

  #27640 commented on Jan 24, 2025 • 0 new comments

• Supporting Selective Activation Checkpointing and CPU Offloading Option.

  #29648 commented on Jan 24, 2025 • 0 new comments

• Reload Transformers imports

  #35508 commented on Jan 23, 2025 • 0 new comments

• hyperparameter_serch() does not consider LoRA parameters like r to be finetuned.

  #29391 commented on Jan 23, 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 Jan 23, 2025 • 0 new comments

• Memory leak on python 3.10.*

  #35434 commented on Jan 23, 2025 • 0 new comments

• AttributeError in automatic_speech_recognition.py when return_segments and return_timestamps are both True

  #35713 commented on Jan 23, 2025 • 0 new comments

• Recomputed tensor size does not match when using activation checkpointing when using FSDP and accelerate

  #34928 commented on Jan 23, 2025 • 0 new comments

• Vision models don't work for non-square object

  #35280 commented on Jan 23, 2025 • 0 new comments

• resizing token embeddings causes output embedding to be reinitialized in `post_init` when `tie_word_embedding` is False

  #35141 commented on Jan 23, 2025 • 0 new comments

• DeBERTa's `DisentangledSelfAttention` hardcodes `float` dtype, which causes `bfloat16` overflow error

  #35332 commented on Jan 23, 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 Jan 23, 2025 • 0 new comments

• Accelerate x Trainer issue tracker:

  #33345 commented on Jan 23, 2025 • 0 new comments

• When gradient checkpointing is enabled, flash_attn_kwargs cannot be passed into the decoder_layer

  #35509 commented on Jan 23, 2025 • 0 new comments

• AttributeError: 'SegformerFeatureExtractor' object has no attribute 'reduce_labels' still has no clear guide around

  #35402 commented on Jan 23, 2025 • 0 new comments