1 Release published by 1 person

• v4.48.0 v4.48.0: ModernBERT, Aria, TimmWrapper, ColPali, Falcon3, Bamba, VitPose, DinoV2 w/ Registers, Emu3, Cohere v2, TextNet, DiffLlama, PixtralLarge, Moonshine

  published Jan 10, 2025

84 Pull requests merged by 51 people

• Update codeowners with individual model owners

  #35595 merged Jan 10, 2025

• Skip MobileNetV1ModelTest::test_batching_equivalence for now

  #35614 merged Jan 10, 2025

• Fix flaky test_beam_search_low_memory

  #35611 merged Jan 10, 2025

• Let EarlyStoppingCallback not require load_best_model_at_end

  #35101 merged Jan 10, 2025

• Added error when sequence length is bigger than max_position_embeddings

  #32156 merged Jan 10, 2025

• Use inherit tempdir makers for tests fix failing DS tests

  #35600 merged Jan 10, 2025

• Fix flaky test_custom_4d_attention_mask

  #35606 merged Jan 10, 2025

• [WIP] Emu3: add model

  #33770 merged Jan 10, 2025

• Fix flex_attention in training mode

  #35605 merged Jan 10, 2025

• Chat template: return vectorized output in processors

  #34275 merged Jan 10, 2025

• Add Moonshine

  #34784 merged Jan 10, 2025

• Skip torchscript tests if we see a cache object produced in a test

  #35596 merged Jan 10, 2025

• ModernBert: reuse GemmaRotaryEmbedding via modular Integration tests

  #35459 merged Jan 10, 2025

• Add flex_attn to diffllama

  #35601 merged Jan 9, 2025

• ModernBERT bug fixes

  #35404 merged Jan 9, 2025

• add _supports_flex_attn = True for models that do support it

  #35598 merged Jan 9, 2025

• [doc] deepspeed universal checkpoint

  #35015 merged Jan 9, 2025

• Refactor/fix Cohere2

  #35863 merged Jan 9, 2025

• [tokenizers] Ensure that add_prefix_space is propagated to backend_tokenizer.pre_tokenizer

  #35593 merged Jan 9, 2025

• Fix modular edge case modular sorting order

  #35562 merged Jan 9, 2025

• PR for Issue #22694: Fixed Training Evaluation table display for VSCode

  #35557 merged Jan 9, 2025

• Small fix rope kwargs

  #35589 merged Jan 9, 2025

• Fix flaky SwitchTransformersModelTest::test_training_gradient

  #35587 merged Jan 9, 2025

• tokenizer train from iterator without pre_tokenizers

  #35396 merged Jan 9, 2025

• feat: add TP plan for granite

  #35573 merged Jan 9, 2025

• [Idefics3] Move image features to same device as input embeds

  #35100 merged Jan 9, 2025

• update modular_modernbert -- add inputs_embeds param to ModernBertModel

  #35373 merged Jan 9, 2025

• Fix flaky test_batching_equivalence

  #35564 merged Jan 9, 2025

• Setup loss_type in config at model init time

  #34616 merged Jan 9, 2025

• Re-add missing __all__ for Cohere and Phi3

  #35578 merged Jan 9, 2025

• Minor fix in video text 2 text docs

  #35546 merged Jan 9, 2025

• More model refactoring!

  #35359 merged Jan 9, 2025

• Don't show warning for inv_freq buffers

  #35255 merged Jan 9, 2025

• Fix multi-gpu loss

  #35395 merged Jan 9, 2025

• update code owners

  #35576 merged Jan 9, 2025

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

  #35199 merged Jan 8, 2025

• Fix all output_dir in test_trainer.py to use tmp_dir

  #35266 merged Jan 8, 2025

• Pipeline: simple API for assisted generation

  #34504 merged Jan 8, 2025

• [PixtralLarge] Update Pixtral conversion script to support large format!

  #34801 merged Jan 8, 2025

• [docs] Remove Hiera from AUDIO MODELS in docs

  #35544 merged Jan 8, 2025

• ovewrite top_k when crate audio classification pipeline

  #35541 merged Jan 8, 2025

• add code owners

  #35528 merged Jan 8, 2025

• Add ViTPose

  #30530 merged Jan 8, 2025

• fix: Qwen2-VL generate with inputs_embeds

  #35466 merged Jan 8, 2025

• Update doc for metric_for_best_model when save_strategy="best".

  #35389 merged Jan 8, 2025

• Add: num_additional_image_tokens to models

  #35052 merged Jan 8, 2025

• Enable auto task for timm models in pipeline

  #35531 merged Jan 8, 2025

• Bump torch requirement to 2

  #35479 merged Jan 8, 2025

• Timm wrapper label names

  #35553 merged Jan 8, 2025

• Update missing model error message

  #35370 merged Jan 8, 2025

• Update doc and default value of TextNetImageProcessor

  #35563 merged Jan 8, 2025

• Add support for modular with fast image processors

  #35379 merged Jan 8, 2025

• [Docs] links to logits-processor-zoo

  #35552 merged Jan 8, 2025

• Fix Qwen2VL processor to handle odd number of frames

  #35431 merged Jan 8, 2025

• support chat generator as input of TextGenerationPipeline

  #35551 merged Jan 8, 2025

• Pass correct num_items_in_batch value into the training_step function

  #35438 merged Jan 8, 2025

• MODERNBERT_INPUTS_DOCSTRING: past_key_values are ignored

  #35513 merged Jan 8, 2025

• VLMs: major clean up 🧼

  #34502 merged Jan 8, 2025

• Add TextNet

  #34979 merged Jan 8, 2025

• [docs] Remove sortish_sampler

  #35539 merged Jan 7, 2025

• Correctly list the chat template file in the Tokenizer saved files list

  #34974 merged Jan 7, 2025

• [Whisper] fix docstrings typo

  #35338 merged Jan 7, 2025

• [Qwen2Audio] handle input ids expansion during processing

  #35534 merged Jan 7, 2025

• Release GPU memory after Optuna trial

  #35440 merged Jan 7, 2025

• Check whether rescale is requested before checking is_scaled_image

  #35439 merged Jan 7, 2025

• Fix bug when requesting input normalization with EnCodec

  #34756 merged Jan 7, 2025

• Add diffllama

  #34083 merged Jan 7, 2025

• NPU support SDPA

  #35165 merged Jan 7, 2025

• Replace tokenizer to processing_class in Seq2SeqTrainer

  #35452 merged Jan 7, 2025

• ci: mark model_parallel tests as cuda specific

  #35269 merged Jan 7, 2025

• Zamba new attention standard

  #35375 merged Jan 7, 2025

• [Dinov2 with Registers] Some fixes

  #35411 merged Jan 6, 2025

• added logic for deleting adapters once loaded

  #34650 merged Jan 6, 2025

• Fixed typo in Llama configuration docstring

  #35520 merged Jan 6, 2025

• 🌐 [i18n-KO] Remove duplicates in toctree

  #35496 merged Jan 6, 2025

• [GGUF] Refactor and decouple gguf checkpoint loading logic

  #34385 merged Jan 6, 2025

• Bump jinja2 from 3.1.4 to 3.1.5 in /examples/research_projects/decision_transformer

  #35408 merged Jan 6, 2025

• Update llm_optims docs for sdpa_kernel

  #35481 merged Jan 6, 2025

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

  #34863 merged Jan 6, 2025

• Add check for if num_items_in_batch is not None

  #35102 merged Jan 6, 2025

• Add position_ids in XLMRobertaXLForCausalLM.prepare_inputs_for_generation

  #35044 merged Jan 6, 2025

• Add French translation of task_summary and tasks_explained

  #33407 merged Jan 6, 2025

• Idefics: fix docstring

  #35079 merged Jan 6, 2025

• Fix Llava conversion for models that use safetensors to store weights

  #35406 merged Jan 6, 2025

33 Pull requests opened by 27 people

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

  #35517 opened Jan 5, 2025

• Remove batch size argument warning when unjustified

  #35519 opened Jan 6, 2025

• Validate the num imgs and vids tokens

  #35521 opened Jan 6, 2025

• Add proper jinja2 error

  #35533 opened Jan 6, 2025

• change README FILE

  #35535 opened Jan 6, 2025

• Security fix for `self-comment-ci.yml`

  #35548 opened Jan 7, 2025

• Add AIMv2 to Transformers

  #35550 opened Jan 7, 2025

• Add support for nested images to LLava and VipLLava

  #35558 opened Jan 7, 2025

• BLIPs clean-up

  #35560 opened Jan 8, 2025

• Support QuestionAnswering Module for ModernBert based models.

  #35566 opened Jan 8, 2025

• Trainer Refactor: Part 1

  #35567 opened Jan 8, 2025

• add qwen2.5vl

  #35569 opened Jan 8, 2025

• Exploit symmetry of covariance matrices for faster & more stable diag…

  #35571 opened Jan 8, 2025

• Multimodal Granite Support

  #35579 opened Jan 9, 2025

• Get latest complete checkpoint directory when auto-resume from checkpoint

  #35580 opened Jan 9, 2025

• apply_chat_template: consistent behaviour for return_assistant_tokens_mask=True return_tensors=True

  #35582 opened Jan 9, 2025

• A new Traditional Chinese version of the README.md for run_on_remote.py

  #35585 opened Jan 9, 2025

• [generation] Support cache-cropping methods

  #35591 opened Jan 9, 2025

• Fix the config class comparison for remote code models

  #35592 opened Jan 9, 2025

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

  #35604 opened Jan 10, 2025

• [tests] make cuda-only tests device-agnostic

  #35607 opened Jan 10, 2025

• Fix device in rope module when using dynamic updates

  #35608 opened Jan 10, 2025

• Still more model refactors!

  #35610 opened Jan 10, 2025

• Uniformize LlavaNextVideoProcessor kwargs

  #35613 opened Jan 10, 2025

• 🚨🚨🚨 An attempt to fix #29554. Include 'LayerNorm.' in gamma/beta rename scope, optimize string search.

  #35615 opened Jan 10, 2025

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

  #35616 opened Jan 10, 2025

• xpu: fix benchmarking scripts for xpu devices

  #35620 opened Jan 11, 2025

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

  #35627 opened Jan 11, 2025

• Guard against unset resolved_archive_file

  #35628 opened Jan 11, 2025

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

  #35633 opened Jan 12, 2025

• Removed some duplicated code

  #35637 opened Jan 12, 2025

• [ViTPose] Convert more checkpoints

  #35638 opened Jan 12, 2025

• modular_model_converter bugfix on assignments

  #35642 opened Jan 12, 2025

58 Issues closed by 19 people

• Add keypoint-detection task

  #24044 closed Jan 12, 2025

• tokenizer.json modified after tokenizer.save_pretrained of OLMO models

  #34744 closed Jan 12, 2025

• Bug of eval loss when enabling average_tokens_across_devices

  #35078 closed Jan 12, 2025

• size mismatch for lm_head.weight: copying a param with shape torch.Size([32, 768]) from checkpoint, the shape in current model is torch.Size([59744, 768]).

  #35084 closed Jan 12, 2025

• Documentation for SWAG contradicts itself when constructing the first sentence.

  #35095 closed Jan 12, 2025

• `dataloader_persistent_workers=True` causes fork-bomb due to repeated creation of `eval_dataloader`

  #28469 closed Jan 11, 2025

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

  #33357 closed Jan 11, 2025

• long-standing Bug in Adafactor optimizer if beta1 > 0

  #34506 closed Jan 11, 2025

• VLlama3ForCausalLM in SmolVLM

  #35039 closed Jan 11, 2025

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

  #35051 closed Jan 11, 2025

• Get "NotImplementedError: Cannot copy out of meta tensor; no data!" error while deploying model

  #35057 closed Jan 11, 2025

• issues when i change the lm_head to a 32 node layer

  #35071 closed Jan 11, 2025

• Multiple training runs not working with deepspeed

  #35073 closed Jan 11, 2025

• OverflowError: out of range integral type conversion attempted

  #35540 closed Jan 10, 2025

• `T5ForSequenceClassification`

  #14097 closed Jan 10, 2025

• Incorrect average calculation in `Perplexity of fixed-length models`

  #34138 closed Jan 10, 2025

• `Nan` logits when performing inference using ModernBERT

  #35574 closed Jan 10, 2025

• PaliGemma2 Processor returns wrong labels array when <image> token is present in `text`

  #35200 closed Jan 10, 2025

• Issues counting passing rates on tests which use subTest()

  #34755 closed Jan 10, 2025

• FA2 broken for Cohere2 if Optional `Mask` is not passed in `forward`

  #35547 closed Jan 9, 2025

• Training Evaluation Display on VSCode

  #22694 closed Jan 9, 2025

• `train_new_from_iterator()` does not work when pre_tokenizer is null

  #35315 closed Jan 9, 2025

• torch.compile DataCollatorWithFlattening flash_attention_2.7 causes crash when training

  #35590 closed Jan 9, 2025

• <spam>

  #35577 closed Jan 9, 2025

• qwen2 rope device matching bug

  #35505 closed Jan 9, 2025

• is_causal arg appears twice in FAttention call from GPT2Attention.forward()

  #35380 closed Jan 9, 2025

• VisualBert: Why isn't the pooler_output used to calculate the logits in VisualBertForQuestionAnswering?

  #35025 closed Jan 9, 2025

• [Idefics 3] Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1

  #35031 closed Jan 9, 2025

• LlamaTokenizer being recognized as a bool

  #35037 closed Jan 9, 2025

• ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation

  #24915 closed Jan 8, 2025

• iframe

  #35559 closed Jan 8, 2025

• Qwen2-VL used to work with `inputs_embeds` instead of `input_ids`, but no more

  #35463 closed Jan 8, 2025

• ModernBERT does not have `inputs_embeds` input

  #35555 closed Jan 8, 2025

• Transformers cannot load ModernBERT for sequence classification

  #35362 closed Jan 8, 2025

• Any plans for a Test-Time Compute Scaling Package or Module?

  #35561 closed Jan 8, 2025

• Qwen2VLProcessor cannot handle odd number of video frames

  #35412 closed Jan 8, 2025

• LLaVa 1.5 and 1.6 not working with text-only inputs

  #35424 closed Jan 8, 2025

• LlavaForConditionalGeneration._merge_input_ids_with_image_features throws error

  #35169 closed Jan 8, 2025

• Assistant decoding w. Llava-Next does not work

  #35450 closed Jan 8, 2025

• Flash attention 2 broke when batch inference

  #34824 closed Jan 8, 2025

• Data collator class type integrity is not intact throughout the runtime

  #34830 closed Jan 8, 2025

• Data prefetching does not occur for iterable datasets

  #34867 closed Jan 8, 2025

• Deepspeed integration: support batch sizes that are less than the number of gpus/ranks

  #27299 closed Jan 8, 2025

• Remove sortish_sampler from Seq2SeqTrainingArgument and the docs

  #34986 closed Jan 7, 2025

• Potentially incorrect calculation of `total_updates` on >=4.46.0 since #34198 affecting multi gpu training

  #35387 closed Jan 7, 2025

• Add support for Allegro

  #34347 closed Jan 7, 2025

• How to run the model on another machine and send the answer to another machine.

  #35485 closed Jan 7, 2025

• xpu: parallelize() not supported for PyTorch XPU backend

  #35252 closed Jan 7, 2025

• Inference with FSDP during training affects checkpoints

  #34530 closed Jan 7, 2025

• SAMProcessor padding for rectangular aspect input images isn't symmetric

  #35017 closed Jan 7, 2025

• Unable to export GLM models to ONNX

  #35021 closed Jan 7, 2025

• Enabling Access to Currently Training Model in Callback Handler

  #35542 closed Jan 7, 2025

• tokenizers.apply_chat_template with `continue_final_message=True` with trailing spaces in input

  #35433 closed Jan 6, 2025

• Bug of self.accelerator.gather(num_items_in_batch) with enabling average_tokens_across_devices

  #35076 closed Jan 6, 2025

• IdeficsImageProcessor raises unexpected ValueError

  #35391 closed Jan 6, 2025

• Redundant Operations.

  #34958 closed Jan 6, 2025

• Memory Leak When Using padding="max_length" in T5 Text Encoder On CPU

  #34988 closed Jan 6, 2025

• OpenBLAS Warning : Detect OpenMP Loop and this application may hang. Please rebuild the library with USE_OPENMP=1 option.

  #35002 closed Jan 6, 2025

45 Issues opened by 42 people

• tokenizer.decode() and tokenizer.convert_ids_to_tokens() return different results

  #35641 opened Jan 12, 2025

• Expected `tensors` and `new_tensors` to have the same type but found <class ‘tuple’> and <class ‘torch.Tensor’>

  #35640 opened Jan 12, 2025

• Breaking change in v4.48.0 and Python 3.9

  #35639 opened Jan 12, 2025

• FSDP OOM error

  #35636 opened Jan 12, 2025

• set_initialized_submodules too slow when loading big model like DeepSeekV3

  #35635 opened Jan 12, 2025

• ValueError: MllamaForConditionalGeneration does not support Flash Attention 2.0 yet

  #35634 opened Jan 12, 2025

• Trying To Convert Paligemma model in npz to hf model format

  #35632 opened Jan 12, 2025

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

  #35630 opened Jan 11, 2025

• Dose `num_logits_to_keep` in `model.generate()` really work?

  #35629 opened Jan 11, 2025

• static cache with mixtral will cause CUDA error: device-side assert triggered

  #35626 opened Jan 11, 2025

• The Phi model does not have lm_head bias after upgraded to v4.48.0

  #35625 opened Jan 11, 2025

• Segmentation fault: address not mapped to object at address 0x100000007

  #35624 opened Jan 11, 2025

• Unsupported: hasattr SkipFunctionVariable when i compile the mixtral model with muti-gpus

  #35623 opened Jan 11, 2025

• running utills.fx.symbolic_trace on gp2 raised an error: torch.fx.proxy.TraceError: Proxy object cannot be iterated, which does not occur in the previous version

  #35622 opened Jan 11, 2025

• The argument "dim" is gone from LlamaRotaryEmbedding initializer. Intentional?

  #35621 opened Jan 11, 2025

• from_pretrained fails to save weights.py and layers.py into cache, therefore fails to find them in cache

  #35619 opened Jan 11, 2025

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

  #35618 opened Jan 10, 2025

• Better handeling of hardcoded component in PretrainedModel.from_pretrained.

  #35617 opened Jan 10, 2025

• Trainer: TensorBoardCallback not working for "on_save" and "on_save_end" events

  #35612 opened Jan 10, 2025

• Trainer sets `state.best_model_checkpoint` even when it doesn't save there; leads to training crash

  #35609 opened Jan 10, 2025

• Prompt_ids feature causing repetitions and hallucinations

  #35603 opened Jan 10, 2025

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

  #35602 opened Jan 10, 2025

• weird criterion to decide if needed to adjust the padding size

  #35599 opened Jan 9, 2025

• Inconsistent saving of tokenizer with custom code from HF hub vs. local directory

  #35597 opened Jan 9, 2025

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

  #35588 opened Jan 9, 2025

• Malformed config when saving & loading locally custom models

  #35584 opened Jan 9, 2025

• Unused kwargs: ['bnb_8bit_quant_type', 'bnb_8bit_use_double_quant', 'bnb_8bit_compute_dtype'] when using bnb quantization? These kwargs are not used in <class 'transformers.utils.quantization_config.BitsAndBytesConfig'>

  #35581 opened Jan 9, 2025

• Tokenizer outputs same offsets for different tokens.

  #35575 opened Jan 9, 2025

• Error occurs when using model.generate with Gemma2 in ZeRO3 environment

  #35572 opened Jan 9, 2025

• Transformers can create unconventional python module names when loading certain repositories

  #35570 opened Jan 8, 2025

• Any plans to integrate GTE model natively into transformers

  #35568 opened Jan 8, 2025

• Add cosmos from Nvidia

  #35565 opened Jan 8, 2025

• 4.47.1 Hugging Face Trainer loss accumulated by sum instead of mean

  #35556 opened Jan 7, 2025

• ModernBERT export to onnx error

  #35545 opened Jan 7, 2025

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

  #35543 opened Jan 7, 2025

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

  #35538 opened Jan 6, 2025

• Mask2FormerImageProcessor support overlapping features

  #35536 opened Jan 6, 2025

• RagTokenizer Missing patch_token_id, patch_token, and encode Functionality

  #35532 opened Jan 6, 2025

• SAM mask-generation - crops_n_layers

  #35530 opened Jan 6, 2025

• Trainer: update `state.num_input_tokens_seen` to use `num_items_in_batch`

  #35529 opened Jan 6, 2025

• Trainer: Use second last checkpoint if last checkpoint loading fails

  #35525 opened Jan 6, 2025

• Warning 'The attention mask is not set'

  #35524 opened Jan 6, 2025

• How about adding a combined step and epoch feature to save_strategy?

  #35523 opened Jan 6, 2025

• Very slow to load deep seekv3 int4 model and device_map="auto" "sequential" bug

  #35522 opened Jan 6, 2025

• Batch size deprecation warning issued even when it is not used

  #35518 opened Jan 6, 2025

136 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 Prompt Depth Anything Model

  #35401 commented on Jan 9, 2025 • 77 new comments

• Samhq model addition

  #35147 commented on Jan 10, 2025 • 48 new comments

• Add Relation DETR

  #34900 commented on Jan 10, 2025 • 32 new comments

• Enhanced Installation Section in README.md

  #35094 commented on Jan 11, 2025 • 26 new comments

• Add TimesFM Time Series Forecasting Model

  #34082 commented on Jan 10, 2025 • 25 new comments

• Implement SuperGlue model

  #29886 commented on Jan 8, 2025 • 24 new comments

• Add GOT-OCR 2.0 to Transformers

  #34721 commented on Jan 7, 2025 • 24 new comments

• Add Molmo (7B-D, 7B-O, 70B)

  #33962 commented on Jan 10, 2025 • 20 new comments

• Support constant lr with cooldown

  #35453 commented on Jan 11, 2025 • 17 new comments

• Add common test for `torch.export` and fix some vision models

  #35124 commented on Jan 10, 2025 • 10 new comments

• Enable gptqmodel

  #35012 commented on Jan 10, 2025 • 7 new comments

• Use AMD CI workflow defined in hf-workflows

  #35058 commented on Jan 9, 2025 • 7 new comments

• Add autoquant support for torchao quantizer

  #35503 commented on Jan 8, 2025 • 7 new comments

• Improve support for image generation with Chameleon & Anole

  #32013 commented on Jan 7, 2025 • 6 new comments

• docs: Clarify descriptions for mask_labels in Mask2Former

  #35514 commented on Jan 11, 2025 • 5 new comments

• Grounding DINO Processor standardization

  #34853 commented on Jan 7, 2025 • 4 new comments

• support telechat2

  #35415 commented on Jan 12, 2025 • 4 new comments

• [WIP] Possible bug - adding option to save/reload scaler

  #34932 commented on Jan 8, 2025 • 3 new comments

• ModernBERT FlexAttention

  #35423 commented on Jan 12, 2025 • 3 new comments

• Universal Speculative Decoding `CandidateGenerator`

  #35029 commented on Jan 12, 2025 • 2 new comments

• Add dithering to the `Speech2TextFeatureExtractor` API.

  #34638 commented on Jan 9, 2025 • 2 new comments

• OmDet Turbo processor standardization

  #34937 commented on Jan 7, 2025 • 2 new comments

• uniformize kwargs for OneFormer

  #34547 commented on Jan 6, 2025 • 2 new comments

• enable StaticCache for assisted generation

  #34797 commented on Jan 7, 2025 • 2 new comments

• Fix Qwen2RotaryEmbedding Device Matching Bug

  #35506 commented on Jan 7, 2025 • 1 new comment

• Efficient Inference Kernel for SpQR

  #34976 commented on Jan 9, 2025 • 1 new comment

• Bart: new cache format

  #35314 commented on Jan 10, 2025 • 1 new comment

• OwlViT/Owlv2 post processing standardization

  #34929 commented on Jan 7, 2025 • 1 new comment

• Integrate xlstm cleanly.

  #35377 commented on Jan 11, 2025 • 1 new comment

• Pass callbacks kwarg to study.optimize() in run_hp_search_optuna()

  #34732 commented on Jan 9, 2025 • 0 new comments

• Add helper for torch dynamo

  #35478 commented on Jan 9, 2025 • 0 new comments

• fix: Handle BLIP-2 model output format

  #34705 commented on Jan 6, 2025 • 0 new comments

• Update config validation

  #34726 commented on Jan 8, 2025 • 0 new comments

• Clean-up composite configs

  #34603 commented on Jan 6, 2025 • 0 new comments

• Add support for Apple's Depth-Pro

  #34583 commented on Jan 7, 2025 • 0 new comments

• Add Zamba2

  #34517 commented on Jan 7, 2025 • 0 new comments

• Added `segmentation maps` support for DPT image processor

  #34345 commented on Jan 10, 2025 • 0 new comments

• LLaVA-NeXT: add new model checkpoints

  #34195 commented on Jan 6, 2025 • 0 new comments

• Added resource class configuration option for `check_circleci_user` job

  #32866 commented on Jan 12, 2025 • 0 new comments

• fix: Updated BridgeTower Image processor

  #32384 commented on Jan 12, 2025 • 0 new comments

• [GroundingDino] Fix grounding dino loss 🚨

  #31828 commented on Jan 10, 2025 • 0 new comments

• [docs] Redesign

  #31757 commented on Jan 11, 2025 • 0 new comments

• Add LightGlue model

  #31718 commented on Jan 8, 2025 • 0 new comments

• Support Kosmos-2.5

  #31711 commented on Jan 10, 2025 • 0 new comments

• [WIP] Add implementation of `_extract_fbank_features_batch`

  #31579 commented on Jan 7, 2025 • 0 new comments

• Option to Disable Model Caching When Using "pipeline"

  #35337 commented on Jan 6, 2025 • 0 new comments

• Add FAST

  #35476 commented on Jan 11, 2025 • 0 new comments

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

  #35455 commented on Jan 8, 2025 • 0 new comments

• T5 static cache

  #35445 commented on Jan 7, 2025 • 0 new comments

• Add D-FINE Model into Transformers

  #35400 commented on Jan 6, 2025 • 0 new comments

• Add support for H2O cache eviction with LLaMA

  #35381 commented on Jan 6, 2025 • 0 new comments

• Several fixes related to rotary position embeddings

  #35376 commented on Jan 11, 2025 • 0 new comments

• Add support for post-processing kwargs in image-text-to-text pipeline

  #35374 commented on Jan 10, 2025 • 0 new comments

• FineTuning AutoModelForSequenceClassification.from_pretrained(meta-llama/Llama-3.2-1B) Bug:RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:1 and cuda:0! (when checking argument for argument target in method wrapper_CUDA_nll_loss_forward) and awq importing

  #35365 commented on Jan 7, 2025 • 0 new comments

• Add JinaBERT model

  #35320 commented on Jan 10, 2025 • 0 new comments

• Add support for DeepSpeed sequence parallelism (Ulysses)

  #35301 commented on Jan 7, 2025 • 0 new comments

• speed up PrefixConstrainedLogitsProcessor

  #35275 commented on Jan 9, 2025 • 0 new comments

• Enhance DataCollatorForLanguageModeling with Configurable Token Replacement Probabilities

  #35251 commented on Jan 9, 2025 • 0 new comments

• Fix Gemma2 synced multi-GPU generation

  #35232 commented on Jan 7, 2025 • 0 new comments

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

  #35193 commented on Jan 12, 2025 • 0 new comments

• Pixtral: vectorize patch embeddings and enable tests

  #35122 commented on Jan 10, 2025 • 0 new comments

• Output dicts support in text generation pipeline

  #35092 commented on Jan 10, 2025 • 0 new comments

• Refactoring of ImageProcessorFast

  #35069 commented on Jan 6, 2025 • 0 new comments

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

  #35010 commented on Jan 7, 2025 • 0 new comments

• Fix max size deprecated warning

  #34998 commented on Jan 12, 2025 • 0 new comments

• Remove _supports_static_cache = True for some model classes

  #34975 commented on Jan 9, 2025 • 0 new comments

• Enable different torch dtype in sub models

  #34873 commented on Jan 9, 2025 • 0 new comments

• add a new flax example for Bert model inference

  #34794 commented on Jan 10, 2025 • 0 new comments

• Adding RTDETRv2

  #34773 commented on Jan 11, 2025 • 0 new comments

• Request for a Vision Transformer Model for Digital Image Segmentation

  #35477 commented on Jan 7, 2025 • 0 new comments

• Instructions to raise PR for addition of shared library files(.so) and .cpp files

  #35492 commented on Jan 7, 2025 • 0 new comments

• RuntimeError: shape '[1, 3098, 6, 5, 128]' is invalid for input of size 12689408

  #35146 commented on Jan 7, 2025 • 0 new comments

• PeftModel is not an instance of PreTrainedModel. `No liger kernels will be applied.`

  #34016 commented on Jan 7, 2025 • 0 new comments

• Stop requiring CacheConfig in GenerationConfig with StaticCache

  #35026 commented on Jan 7, 2025 • 0 new comments

• `transformers.image_transforms.resize` doesnot work for negative values

  #34920 commented on Jan 7, 2025 • 0 new comments

• Deprecation Warning for `max_size` in `DetrImageProcessor.preprocess`

  #34977 commented on Jan 7, 2025 • 0 new comments

• Deepseek v2

  #35317 commented on Jan 7, 2025 • 0 new comments

• Add AudioQuestionAnswering pipeline

  #33782 commented on Jan 7, 2025 • 0 new comments

• Any plans to add AIMv2 in the model?

  #35351 commented on Jan 7, 2025 • 0 new comments

• Impossible to change attention implementation

  #35153 commented on Jan 8, 2025 • 0 new comments

• Qwen2vl float16 inference bug in naive attention

  #35151 commented on Jan 8, 2025 • 0 new comments

• Cuda OOM

  #35150 commented on Jan 8, 2025 • 0 new comments

• Discrepancy in Training Loss Behavior with Gradient Accumulation using DeepSpeed

  #34694 commented on Jan 8, 2025 • 0 new comments

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

  #35447 commented on Jan 8, 2025 • 0 new comments

• Multi-GPU training crashes with IterableDataset and different length input (e.g. Next token prediction)

  #35308 commented on Jan 8, 2025 • 0 new comments

• Custom 4D tensor caused shape mismatch error

  #35290 commented on Jan 5, 2025 • 0 new comments

• Special token ids are not longer typed properly in 4.47.0

  #35126 commented on Jan 6, 2025 • 0 new comments

• The dot in the model name when using auto_map will cause a path parsing error.

  #35082 commented on Jan 6, 2025 • 0 new comments

• CPU processing is extremely slow for models loaded with `torch_dtype = torch.float16`

  #34692 commented on Jan 6, 2025 • 0 new comments

• FlaxWhisperForConditionalGeneration Out Of Memory Error

  #34668 commented on Jan 6, 2025 • 0 new comments

• SinkCache (StreamLLM) implemented over Post-RoPE Key cache might result in confused position for inference

  #35350 commented on Jan 6, 2025 • 0 new comments

• 'do_sample' model default cannot be overridden

  #35372 commented on Jan 6, 2025 • 0 new comments

• rework `test_multi_gpu_data_parallel_forward`

  #31087 commented on Jan 6, 2025 • 0 new comments

• Set output_attentions=True for model.geneate

  #35393 commented on Jan 6, 2025 • 0 new comments

• Allow static cache to be larger than sequence length / batch size for encoder-decoder models

  #35444 commented on Jan 6, 2025 • 0 new comments

• Training config that worked with transformers v4.4.6.3 results in OOM error with v4.47.0 (using SFTTrainer)

  #35108 commented on Jan 6, 2025 • 0 new comments

• DeepSeek V3 Support

  #35425 commented on Jan 6, 2025 • 0 new comments

• `trainer.evaluate` always creates a new MLFlow run, separate from the one used during `train()`

  #35074 commented on Jan 6, 2025 • 0 new comments

• HfArgumentParser error when using LoraConfig dataclass

  #34834 commented on Jan 6, 2025 • 0 new comments

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

  #34747 commented on Jan 6, 2025 • 0 new comments

• `GPT2Attention()` class with `_attn()` method when `add_cross_attention=True` and therefore `is_cross_attention=True`.

  #35430 commented on Jan 6, 2025 • 0 new comments

• Training issues latest version

  #35407 commented on Jan 6, 2025 • 0 new comments

• Log multiple losses used along with the combined losses when a model returns a dictionary of losses.

  #31081 commented on Jan 10, 2025 • 0 new comments

• Uniform kwargs for processors

  #31911 commented on Jan 10, 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 10, 2025 • 0 new comments

• run_mlm_flax on tpu v5-pods

  #35205 commented on Jan 11, 2025 • 0 new comments

• logged loss is not correct with gradient accumulation

  #35204 commented on Jan 11, 2025 • 0 new comments

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

  #35203 commented on Jan 11, 2025 • 0 new comments

• Accelerate x Trainer issue tracker:

  #33345 commented on Jan 11, 2025 • 0 new comments

• Initializing via AutoImageProcessor before AutoProcessor is imported causes `AttributeError`

  #34307 commented on Jan 11, 2025 • 0 new comments

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

  #35244 commented on Jan 12, 2025 • 0 new comments

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

  #34855 commented on Jan 12, 2025 • 0 new comments

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

  #34427 commented on Jan 12, 2025 • 0 new comments

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

  #33552 commented on Jan 12, 2025 • 0 new comments

• Wav2Vec2BertForSequenceClassification. return_attention_mask work wrong

  #35495 commented on Jan 12, 2025 • 0 new comments

• How can I disable legacy processing in llava-next

  #35457 commented on Jan 12, 2025 • 0 new comments

• apply class transformers.SequenceBiasLogitsProcessor on Qwen model

  #35432 commented on Jan 12, 2025 • 0 new comments

• Beit image classification have different results compared from versions prior to 4.43.0

  #34446 commented on Jan 12, 2025 • 0 new comments

• Potential fix for #30819. Check for best metric only on gpu 0

  #31268 commented on Jan 6, 2025 • 0 new comments

• Mimi model gives different outputs when using batch encode vs single encode

  #35166 commented on Jan 9, 2025 • 0 new comments

• [tests] run one test but got 2 test results

  #35159 commented on Jan 9, 2025 • 0 new comments

• When extending embeddings, multivariate distribution isn't correctly estimated even when the calculated sigma matrix is symmetric and positive definite

  #35075 commented on Jan 9, 2025 • 0 new comments

• Plain-DETR

  #27496 commented on Jan 9, 2025 • 0 new comments

• Calling Trainer.create_model_card() with an empty dataset list causes an IndexError

  #35163 commented on Jan 9, 2025 • 0 new comments

• modernbert logits do not have gradient

  #35386 commented on Jan 9, 2025 • 0 new comments

• tracker: `generate` compatibility with `torch.compile`

  #28981 commented on Jan 9, 2025 • 0 new comments

• VLMs Processors are not fully consistent in the inputs formats they accept

  #34545 commented on Jan 9, 2025 • 0 new comments

• Unclear what happens when using torchrun, multi-gpu and trainer arguments.

  #35311 commented on Jan 9, 2025 • 0 new comments

• DynamicCache does not support variable lengths, except for FA2

  #35168 commented on Jan 9, 2025 • 0 new comments

• Export to ExecuTorch

  #32253 commented on Jan 9, 2025 • 0 new comments

• Can't load models with a gamma or beta parameter

  #29554 commented on Jan 10, 2025 • 0 new comments

• Qwen2vl support for GGUF

  #35282 commented on Jan 10, 2025 • 0 new comments

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

  #35186 commented on Jan 10, 2025 • 0 new comments

• QuantizedCache first token processing is counterintuitive / worse than in papers

  #35185 commented on Jan 10, 2025 • 0 new comments

• Saving model with shared tensors fails on cpu but succeeds on gpu

  #33688 commented on Jan 10, 2025 • 0 new comments

• Unknown quantization type, got fp8

  #35471 commented on Jan 10, 2025 • 0 new comments