2 Releases published by 1 person

• v4.52.4 Patch release: v4.52.4

  published May 30, 2025

• v4.52.4-ColQwen2-preview ColQwen2 (based on v4.52.4)

  published Jun 2, 2025

61 Pull requests merged by 39 people

• Updated deprecated typing imports with equivalents for Python 3.9+

  #38546 merged Jun 4, 2025

• New gpt neo model card

  #38505 merged Jun 4, 2025

• tests/roformer: fix couple roformer tests on gpus

  #38570 merged Jun 4, 2025

• [Dinov2] Enable device_map="auto" support

  #38487 merged Jun 4, 2025

• Add Expectations for three AMD tests

  #38581 merged Jun 4, 2025

• Pin Scipy version to >=1.12.0

  #38469 merged Jun 4, 2025

• Janus seamless transfer

  #38580 merged Jun 4, 2025

• feat: add repository field to benchmarks table

  #38582 merged Jun 4, 2025

• Docs: fix code formatting in torchao docs

  #38504 merged Jun 4, 2025

• allow custom head_dim for qwen2_moe

  #37188 merged Jun 4, 2025

• fix(attention_visualizer): add default value for image_seq_length

  #38577 merged Jun 4, 2025

• [FlexAttn] Fix models with unique characteristics

  #38433 merged Jun 4, 2025

• Expectation changes and more AMD expectations

  #38529 merged Jun 4, 2025

• Fix deepseekv3

  #38562 merged Jun 4, 2025

• update utils/notification_service.py for AMD vs Nvidia

  #38563 merged Jun 4, 2025

• Fix chameleon tests

  #38565 merged Jun 4, 2025

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

  #35831 merged Jun 4, 2025

• Relaxed the output_attention condition for ValueError

  #38560 merged Jun 4, 2025

• [janus] Fix failing tests on mi3XX

  #38426 merged Jun 4, 2025

• Added guards against device mismatch errors in some LogitsProcessor classes

  #38558 merged Jun 4, 2025

• Fixed a multiple-devices issue in SmolVLMModel

  #38557 merged Jun 4, 2025

• [docs] Format fix

  #38414 merged Jun 3, 2025

• Fix hqq issue

  #38551 merged Jun 3, 2025

• Name change AOPermod -> ModuleFqn

  #38456 merged Jun 3, 2025

• Fix utils/notification_service.py

  #38556 merged Jun 3, 2025

• Explicitly setting encoding in tokenization_utils_base.py

  #38553 merged Jun 3, 2025

• [TP] Change command in tests to python3

  #38555 merged Jun 3, 2025

• [bugfix] fix apply_rotary_emb error on Ascend NPU

  #38491 merged Jun 3, 2025

• Update docker image to use av==10.0.0

  #38548 merged Jun 3, 2025

• update emu3 test

  #38543 merged Jun 3, 2025

• Don't use default attn if pre-set in sub-config

  #38526 merged Jun 3, 2025

• [tests] expand flex-attn test for vision models

  #38434 merged Jun 3, 2025

• Fix blip2 tests

  #38510 merged Jun 2, 2025

• Fix Gemma2IntegrationTest

  #38492 merged Jun 2, 2025

• Remove type annotation in Siglip Attention Module

  #38503 merged Jun 2, 2025

• Num parameters in model.safetensors.index.json

  #38531 merged Jun 2, 2025

• [flax/mistral] support sliding_window: null in config

  #37402 merged Jun 2, 2025

• Fix amp deprecation issue

  #38100 merged Jun 2, 2025

• remove unhandled parameter

  #38145 merged Jun 2, 2025

• Add ColQwen2 to 🤗 transformers

  #35778 merged Jun 2, 2025

• [generate] move SinkCache to a custom_generate repo

  #38399 merged Jun 2, 2025

• [generate] add soft deprecations on custom generation methods

  #38406 merged Jun 2, 2025

• Update Loss Functions to Accept Tensor num_items_in_batch

  #38029 merged Jun 2, 2025

• [seamless_m4t] Skip some tests when speech is not available

  #38430 merged Jun 2, 2025

• Fix setting FLASH_ATTENTION_DETERMINISTIC after importing

  #37185 merged Jun 2, 2025

• Remove deprecated use_flash_attention_2 parameter

  #37131 merged Jun 2, 2025

• [docs] add xpu environment variable for gpu selection

  #38194 merged May 30, 2025

• protect dtensor import

  #38496 merged May 30, 2025

• Align TP check

  #38328 merged May 30, 2025

• [Tests] Reduced model size for albert-test model

  #38480 merged May 30, 2025

• Bump torch from 2.2.0 to 2.6.0 in /examples/flax/vision

  #37618 merged May 30, 2025

• Fix incorrect bbox_embed initialization when decoder_bbox_embed_share=False in GroundingDINO

  #38238 merged May 30, 2025

• Fix convert_internvl_weights_to_hf.py to support local paths

  #38264 merged May 30, 2025

• Make patch helper more helpful

  #38409 merged May 30, 2025

• fix: handle no scheduler passed by user

  #38407 merged May 30, 2025

• [Qwen2.5-Omni] Fix dtype of cos,sin when used with flash attention

  #38453 merged May 29, 2025

• Fix Gemma3IntegrationTest

  #38471 merged May 29, 2025

• Cleanup BatchFeature and BatchEncoding

  #38459 merged May 29, 2025

• Fix TypeError in save_pretrained error handling (fixes #38422)

  #38449 merged May 29, 2025

• 🔴 [VLM] modeling updates

  #38317 merged May 29, 2025

• [Tests] Clean up test cases for few models

  #38315 merged May 29, 2025

53 Pull requests opened by 40 people

• Add glpn fast processor

  #38461 opened May 29, 2025

• fix torch_dtype on awq

  #38463 opened May 29, 2025

• Fix trainer.py not showing signature columns

  #38465 opened May 29, 2025

• Fix HQQ model param device transfer issue

  #38466 opened May 29, 2025

• [VLMs] support passing embeds along with pixels

  #38467 opened May 29, 2025

• Add detailed ConvBERT model card with usage, architecture, and refere…

  #38470 opened May 29, 2025

• Updated Aria model card

  #38472 opened May 29, 2025

• docs: Add Turkish translation for README

  #38473 opened May 29, 2025

• Avoid overwrite existing local implementation when loading remote custom model

  #38474 opened May 29, 2025

• Refactor DBRX tests to use CausalLMModelTest base classes

  #38475 opened May 29, 2025

• Fix meta tensor copy error

  #38478 opened May 29, 2025

• [static cache] fix device map per layer in VLMs

  #38488 opened May 30, 2025

• lazy cache init

  #38495 opened May 30, 2025

• Add ZoeDepthImageProcessorFast: PyTorch-native Fast Image Preprocessing for ZoeDepth

  #38497 opened May 30, 2025

• Add fast imageprocessor vitpose

  #38502 opened May 31, 2025

• Fixed markdown for BertTokenizer's '[CLS]' token.

  #38506 opened May 31, 2025

• Fix initialization of a pretrained backbone

  #38512 opened Jun 1, 2025

• Update blip model card

  #38513 opened Jun 1, 2025

• added fast image processor for ZoeDepth and expanded tests accordingly

  #38515 opened Jun 1, 2025

• Fix `return_dict=False` giving errors in a few VLM models

  #38519 opened Jun 1, 2025

• Add QuasarV4 model

  #38520 opened Jun 1, 2025

• [qwen-omni] fix sliding window

  #38525 opened Jun 2, 2025

• Logging message for ``` is_bitsandbytes_available() ```

  #38528 opened Jun 2, 2025

• Fix to make vllm happy

  #38530 opened Jun 2, 2025

• On branch fix-void-segment-mask-input [WIP]

  #38532 opened Jun 2, 2025

• another way to use shift_labels

  #38533 opened Jun 2, 2025

• fixed a bug which was causing only partial files to be imported

  #38534 opened Jun 2, 2025

• Update data collator to support sequence_length

  #38536 opened Jun 2, 2025

• Allow `mlm_probability` to be set to `None` when `mlm=False` in DataCollatorForLanguageModeling (#38522)

  #38537 opened Jun 2, 2025

• [docs] transformers-cli command

  #38539 opened Jun 2, 2025

• Image processor compile fix

  #38540 opened Jun 3, 2025

• [WIP chat template] return assistant mask in processors

  #38545 opened Jun 3, 2025

• Fix CTRL model DataParallel compatibility

  #38547 opened Jun 3, 2025

• Improve GPTNeoX model card following standardization guidelines

  #38550 opened Jun 3, 2025

• Better CI

  #38552 opened Jun 3, 2025

• [masking utils] check `None` instead of try/except

  #38561 opened Jun 3, 2025

• Fix ModernBERT tokenizer issue with is_split_into_words flag

  #38564 opened Jun 3, 2025

• Fix `FalconMambaIntegrationTests`

  #38566 opened Jun 3, 2025

• Update Wav2Vec2 documentation to create model cards

  #38568 opened Jun 3, 2025

• Add Bagel

  #38569 opened Jun 3, 2025

• enable more test cases on xpu

  #38572 opened Jun 4, 2025

• Fix `MiniMax` (docs and integration tests checkpoint)

  #38575 opened Jun 4, 2025

• Disable custom MRA kernels for ROCm

  #38578 opened Jun 4, 2025

• blt wip

  #38579 opened Jun 4, 2025

• update `ColQwen2ModelIntegrationTest`

  #38583 opened Jun 4, 2025

• Fix zero rotary dim

  #38584 opened Jun 4, 2025

• [don't merge yet] Fix RAG

  #38585 opened Jun 4, 2025

• docs: fix dark mode logo display.

  #38586 opened Jun 4, 2025

• Refactor Bamba tests to inherit from CausalLMModelTester base classes

  #38587 opened Jun 4, 2025

• Remove custom pytest and pluggy

  #38589 opened Jun 4, 2025

• Fix: Correctly handle integer device_map for NPU devices in _load_sta…

  #38591 opened Jun 4, 2025

• [docs] Add int4wo + 2:4 sparsity example to TorchAO README

  #38592 opened Jun 4, 2025

• Remove IPEX requirement for bitsandbytes on CPU

  #38594 opened Jun 4, 2025

50 Issues closed by 22 people

• Qwen2.5-VL using ascend NPU with flash-attention-2 raises error

  #38189 closed Jun 4, 2025

• Community contribution: enabling `device_map="auto"` support for more vision and multimodal models

  #29786 closed Jun 4, 2025

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

  #34855 closed Jun 4, 2025

• allow custom head_dim for qwen2_moe

  #37187 closed Jun 4, 2025

• A local variable 'image_seq_length' leading to UnboundLocalError: cannot access local variable 'image_seq_length' where it is not associated with a value

  #38576 closed Jun 4, 2025

• Support for excel files

  #38567 closed Jun 4, 2025

• 404 Client Error when accessing https://router.huggingface.co/nebius/v1/chat/completions endpoint

  #38524 closed Jun 4, 2025

• TapasTokenizer Produces All Zero token_type_ids Even with Tutorial Data

  #37183 closed Jun 4, 2025

• Whisper chunking algorithm increases WER

  #37789 closed Jun 4, 2025

• AutomaticMaskGeneration does not work with batch_size greater than 1

  #37805 closed Jun 4, 2025

• ValueError: size must contain 'shortest_edge' and 'longest_edge' keys.

  #37811 closed Jun 4, 2025

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

  #35710 closed Jun 4, 2025

• 1

  #38574 closed Jun 4, 2025

• register_quantizer or register_quantization_config does not add new method to QuantizationMethod

  #38462 closed Jun 4, 2025

• Hang in quantized_phi::ModelWeights::forward() with Phi-2 GGUF on CPU (Candle main branch)

  #38516 closed Jun 3, 2025

• Why do you remove sample_indices_fn for processor.apply_chat_template?

  #38527 closed Jun 3, 2025

• torch.compile fails for gemma-3-1b-it

  #38501 closed Jun 2, 2025

• num_items_in_batch should be moved to logits.device in ForCausalLMLoss too

  #37886 closed Jun 2, 2025

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

  #38095 closed Jun 2, 2025

• Release Tag Changed, Breaking Checksums, and AUR Package Building

  #37090 closed Jun 2, 2025

• Loading and Saving Pretrained model to the same directory raises SafeTensorError: IOError

  #37713 closed Jun 2, 2025

• Failed to load model with transformers 4.51.3 when WORLD_SIZE set to 1 on nvidia gpu

  #37737 closed Jun 2, 2025

• [Trainer] As gradient_accumulation_steps increases, the loss also increases

  #37766 closed Jun 2, 2025

• error: subprocess-exited-with-error when install transformerspython

  #37775 closed Jun 2, 2025

• DTensor Import Path Changes in PyTorch 2.5 Causing Compatibility Issues

  #38251 closed Jun 2, 2025

• TypeError: CustomTrainer.compute_loss() got an unexpected keyword argument 'num_items_in_batch'

  #36331 closed Jun 1, 2025

• Can't perform inference with images on Gemma-3-12b-it-qat-int4.0

  #37710 closed Jun 1, 2025

• Very slow model instantiation

  #37712 closed Jun 1, 2025

• Facing "NotImplementedError: ``config.do_stable_layer_norm is False`` is currently not supported" in FlaxWav2vec2Models

  #37731 closed Jun 1, 2025

• Model Request: SLaM (Sparse Latent Mixer) – Multimodal Flamingo Alternative

  #38508 closed May 31, 2025

• bug in new prefill_chunk_size implementation

  #38028 closed May 31, 2025

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

  #36041 closed May 31, 2025

• Object detection tutorial uses buggy dataset, may lead to crash during training

  #36516 closed May 31, 2025

• Tokenizing with `apply_chat_template` behaves differently from regular tokenizing

  #37686 closed May 31, 2025

• num_items_in_batch larger than the actual useful token when computing loss

  #38448 closed May 31, 2025

• ImportError: cannot import name 'DTensor' from 'torch.distributed.tensor'

  #38494 closed May 30, 2025

• [Tests] Testing for ALBERT is quite slow

  #38344 closed May 30, 2025

• Unexpected Zero Probabilities with siglip2-base-patch16-224 Model

  #38175 closed May 30, 2025

• VLM reverse mapping logic in modeling_utils.py save_pretrained not doing anything?

  #38489 closed May 30, 2025

• TypeError in Llama-4-Maverick-17B-128E-Instruct-FP8 Resolved with Workaround

  #38283 closed May 30, 2025

• A shallow copy in groundingdino

  #37333 closed May 30, 2025

• convert_internvl_weights_to_hf.py does not support model in a local path

  #38200 closed May 30, 2025

• [Bug - Qwen2.5-Omni] FlashAttention 2 BF16 dtype mismatch persists in `apply_rotary_pos_emb_flashatt`

  #38451 closed May 29, 2025

• Bug in error handling routine in save_pretrained

  #38422 closed May 29, 2025

• add Flash Attention Support for Helsinki-NLP/opus models

  #36169 closed May 29, 2025

• Gemma3 (and Paligemma) position_ids 1-indexed?

  #36856 closed May 29, 2025

• bitnet

  #37632 closed May 29, 2025

• Error message is misleading for missing protobuf

  #37641 closed May 29, 2025

• phi-4-mm HF format

  #38120 closed May 29, 2025

• _register_pytree_node error in torch2.1.0 and bf16 assertion error for XPU and NPU

  #37838 closed May 29, 2025

32 Issues opened by 32 people

• Consider Deprecating Sigopt from Hyperparameter Search

  #38593 opened Jun 4, 2025

• "facebook/opt-125m" gives wrong results

  #38590 opened Jun 4, 2025

• Transformers fail to load deepseek-ai/DeepSeek-V3 with vllm

  #38588 opened Jun 4, 2025

• Request to add the small-doge model

  #38573 opened Jun 4, 2025

• Oneke not utilizing much from GPU(Nvidia L20)

  #38571 opened Jun 4, 2025

• Possible Typo in "Mask2FormerLoss"

  #38559 opened Jun 3, 2025

• hidden_states, self_attn_weights = self.self_attn( ValueError: too many values to unpack (expected 2)

  #38554 opened Jun 3, 2025

• Clarification on default top_k sampling parameter

  #38549 opened Jun 3, 2025

• Paligemma model card needs update

  #38544 opened Jun 3, 2025

• enable GraniteMoeHybridIntegrationTest in UT

  #38542 opened Jun 3, 2025

• `eager_attention_forward` and `repeat_kv` code duplication

  #38541 opened Jun 3, 2025

• Hidden states are different for model() and model.generate()

  #38538 opened Jun 2, 2025

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

  #38535 opened Jun 2, 2025

• "Size mismatch" error when trying to download pretrained ChatGPT-4 using transformers

  #38523 opened Jun 2, 2025

• Allow `mlm_probability` to be set to None when `mlm`=False in `DataCollatorForLanguageModeling`

  #38522 opened Jun 2, 2025

• Error for `return_assistant_tokens_mask` in MLLM processor

  #38521 opened Jun 2, 2025

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

  #38518 opened Jun 1, 2025

• model_type = self._reverse_config_mapping[key.__name__] KeyError: 'Qwen2RMConfig'

  #38517 opened Jun 1, 2025

• Can not reproduce Blip2ForImageTextRetrieval example from docs, getting different results

  #38514 opened Jun 1, 2025

• SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference

  #38509 opened May 31, 2025

• id2label assignment problem in run_glue.py

  #38507 opened May 31, 2025

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

  #38500 opened May 30, 2025

• ModernBERT for MLM outputs incorrect hidden state shape.

  #38499 opened May 30, 2025

• [Florence-2] SyntaxWarning: invalid escape sequence '\d' in processing_florence2.py

  #38498 opened May 30, 2025

• Clarification on per_device_train_batch_size in Trainer

  #38484 opened May 30, 2025

• Transformers 4.41.0 does not recognize 'gemma2' model type for google/gemma-2-2b

  #38482 opened May 29, 2025

• Token shape issue in LLaVA-onevision fine-tuning

  #38481 opened May 29, 2025

• ImportError: DLL load failed while importing _safetensors_rust: The specified module could not be found

  #38479 opened May 29, 2025

• Pickle error when downloading DeepSeek model

  #38476 opened May 29, 2025

• AssertionError: Torch not compiled with CUDA enabled when using device_map="auto" in Ascend NPU

  #38468 opened May 29, 2025

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

  #38464 opened May 29, 2025

• Incorrect API call

  #38457 opened May 28, 2025

116 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 support for Florence-2

  #38188 commented on Jun 4, 2025 • 73 new comments

• GLM-4-0414 Change

  #38431 commented on Jun 4, 2025 • 39 new comments

• Split `transformers chat` and `transformers serve`

  #38443 commented on Jun 3, 2025 • 31 new comments

• Add EoMT Model

  #37610 commented on Jun 4, 2025 • 27 new comments

• support MiniCPM-o2.6

  #37917 commented on Jun 3, 2025 • 14 new comments

• Add LightGlue model

  #31718 commented on Jun 4, 2025 • 11 new comments

• Add Fast Image Processor for mobileViT

  #37143 commented on Jun 2, 2025 • 7 new comments

• Add X-Codec model

  #38248 commented on May 30, 2025 • 7 new comments

• Encoder-Decoder Gemma

  #38332 commented on May 30, 2025 • 6 new comments

• fix total batch size calculation in trainer

  #38286 commented on Jun 4, 2025 • 5 new comments

• support overlapping masks in mask2former image processor

  #37357 commented on Jun 3, 2025 • 3 new comments

• Skip non-selected experts for qwen3_moe

  #38133 commented on Jun 2, 2025 • 3 new comments

• [WIP] Perception lm

  #37878 commented on May 31, 2025 • 2 new comments

• Add Ovis2 model and processor implementation

  #37088 commented on Jun 4, 2025 • 2 new comments

• Add kernelize to transformers

  #38205 commented on Jun 3, 2025 • 2 new comments

• Enable tracing for Moshi

  #36894 commented on Jun 2, 2025 • 2 new comments

• Fix Whisper inference regression with backward-compatible logprob calculation

  #38388 commented on Jun 3, 2025 • 2 new comments

• Add Dia model

  #38405 commented on Jun 4, 2025 • 2 new comments

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

  #38086 commented on Jun 2, 2025 • 1 new comment

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

  #37220 commented on Jun 2, 2025 • 1 new comment

• [trainer] ensure special tokens in model configs are aligned with tokenizer at train time

  #38441 commented on Jun 3, 2025 • 1 new comment

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

  #36534 commented on May 30, 2025 • 1 new comment

• Update tokenization_utils_base.py

  #37512 commented on Jun 2, 2025 • 0 new comments

• internalize build_inputs_with_special_tokens and prepare_for_model

  #37522 commented on Jun 2, 2025 • 0 new comments

• Fix interpolation of convnext image processor

  #37460 commented on Jun 4, 2025 • 0 new comments

• Fast tokenizer encoding doesn't handle empty string input

  #37537 commented on Jun 2, 2025 • 0 new comments

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

  #37394 commented on Jun 2, 2025 • 0 new comments

• Add configurable normalization schemes to SigLIP image processors

  #38444 commented on May 29, 2025 • 0 new comments

• handle training summary when creating modelcard but offline mode is set

  #37095 commented on Jun 2, 2025 • 0 new comments

• [draft] random tests order

  #37082 commented on Jun 2, 2025 • 0 new comments

• Adding a stub for MiniCPM-o to the models

  #37049 commented on Jun 3, 2025 • 0 new comments

• Add Fast Segformer Processor

  #37024 commented on Jun 4, 2025 • 0 new comments

• Add Fast SamImageProcessor

  #36999 commented on Jun 1, 2025 • 0 new comments

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

  #36852 commented on Jun 2, 2025 • 0 new comments

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

  #36808 commented on Jun 4, 2025 • 0 new comments

• Add Aimv2 model

  #36625 commented on May 29, 2025 • 0 new comments

• Fix edge case for tokenize (#36277)

  #36555 commented on May 29, 2025 • 0 new comments

• [Qwen2.5-VL] Fix empty string input crash in processor

  #38421 commented on May 29, 2025 • 0 new comments

• Lag kv cache

  #38364 commented on Jun 3, 2025 • 0 new comments

• align xpu's autocast behavior w/ cuda by using device agnostic torch APIs

  #38284 commented on Jun 4, 2025 • 0 new comments

• Add zero dim tensor check when using flash_attention

  #38280 commented on May 30, 2025 • 0 new comments

• [docs] Tensor parallelism

  #38241 commented on Jun 2, 2025 • 0 new comments

• Add SVE implementation for Mamba Sequential Scan Algorithm

  #38185 commented on Jun 3, 2025 • 0 new comments

• [WIP] new BLT

  #38173 commented on May 30, 2025 • 0 new comments

• Fix FSDP + llava-next/llava-onevision

  #38141 commented on Jun 4, 2025 • 0 new comments

• Cache System Refactor: Layered Architecture

  #38077 commented on Jun 4, 2025 • 0 new comments

• update loss computation in modeling code

  #37993 commented on Jun 2, 2025 • 0 new comments

• Add dia

  #37941 commented on Jun 3, 2025 • 0 new comments

• [WIP] Add MM Grounding DINO

  #37925 commented on May 30, 2025 • 0 new comments

• Add DEIM object detection model

  #37875 commented on Jun 1, 2025 • 0 new comments

• [WiP] Add xcodec2 model

  #37868 commented on Jun 4, 2025 • 0 new comments

• Fixed a bug calculating cross entropy loss in `JetMoeForCausalLM`

  #37830 commented on May 30, 2025 • 0 new comments

• qwen null pointer check.

  #37810 commented on Jun 2, 2025 • 0 new comments

• Update ruff to 0.11.7 and some fixes

  #37809 commented on Jun 4, 2025 • 0 new comments

• fix qwen2.5-omini cant be loaded from AutoModel

  #37795 commented on Jun 2, 2025 • 0 new comments

• Adding features like Tokenizer evaluation/benchmarking

  #37792 commented on Jun 2, 2025 • 0 new comments

• Updated Albert model Card

  #37753 commented on Jun 4, 2025 • 0 new comments

• refactor create_token_type_ids_from_sequences

  #37681 commented on Jun 2, 2025 • 0 new comments

• Non model inits

  #37653 commented on Jun 2, 2025 • 0 new comments

• Add DeepSeek V2 Model into Transformers

  #36400 commented on Jun 3, 2025 • 0 new comments

• 💡 Proposal: Add temporal-grounding pipeline for video-language tasks

  #38450 commented on Jun 1, 2025 • 0 new comments

• Convnext image preprocessor raises an AssertionError when comparing logits

  #37461 commented on Jun 1, 2025 • 0 new comments

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

  #38061 commented on Jun 1, 2025 • 0 new comments

• Please support GGUF format for UMT5EncoderModel

  #36774 commented on May 31, 2025 • 0 new comments

• Any plans on adding Flash Attention 3?

  #33373 commented on May 31, 2025 • 0 new comments

• 401 Unauthorized Error: "Invalid credentials" on POST requests to Inference API from multiple services

  #38289 commented on May 31, 2025 • 0 new comments

• [BUG] Batch inference DDP + zero stage 3 = inference code hangs

  #36638 commented on May 31, 2025 • 0 new comments

• ModernBert Tokenizer flag `is_split_into_words` not working

  #37883 commented on May 31, 2025 • 0 new comments

• Error in input expansion for `generate` with `num_return_sequences` > 1 for multi-image inputs to `AutoModelForImageTextToText`

  #37900 commented on May 31, 2025 • 0 new comments

• Object detection training/fine-tuning for Owl-vit/Owlv2

  #33664 commented on May 31, 2025 • 0 new comments

• OWL-ViT training / fine-tuning code

  #20091 commented on May 31, 2025 • 0 new comments

• Gibberish generations with FSDP2 and MixedPrecisionPolicy

  #38190 commented on May 30, 2025 • 0 new comments

• A type error in the Template writing document

  #37524 commented on May 30, 2025 • 0 new comments

• ImageInput doesn't include JAX ndarray and TensorFlow tensor

  #37857 commented on May 30, 2025 • 0 new comments

• BUG: ModernBERT flash-attention2 incompatible on Ascend NPU

  #37859 commented on May 30, 2025 • 0 new comments

• Llama2 can output scores normally, but Llama3 outputs full inf

  #37862 commented on May 30, 2025 • 0 new comments

• WhisperForCTC

  #26242 commented on May 30, 2025 • 0 new comments

• Potential mix-up with IMAGENET_STANDARD and IMAGENET_DEFAULT values

  #38318 commented on May 30, 2025 • 0 new comments

• Version 4.52.3 leads to error after bundling with pyinstaller

  #38402 commented on May 29, 2025 • 0 new comments

• Memory saving by upcasting logits for only non-ignored positions

  #38452 commented on May 29, 2025 • 0 new comments

• accelerate + device_map auto = error

  #38408 commented on May 29, 2025 • 0 new comments

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

  #36560 commented on May 29, 2025 • 0 new comments

• The same situation as #31377 occurred when using Qwen/Qwen2-VL-7B-Instruct

  #33399 commented on May 29, 2025 • 0 new comments

• Gemma3: Cuda error: misaligned address

  #36961 commented on May 29, 2025 • 0 new comments

• Decoder Attention Mask is not passed to the VisionEncoderDecoderModel during training!!

  #37823 commented on May 29, 2025 • 0 new comments

• AttentionMaskVisualizer hard-code sliding_window to 5 in transformers code.

  #37851 commented on May 29, 2025 • 0 new comments

• Will Trainer.predict() return data in the same order as the original dataset during multi-machine and multi-gpus inference?

  #33728 commented on May 29, 2025 • 0 new comments

• Add support for BAGEL from ByteDance

  #38267 commented on May 29, 2025 • 0 new comments

• Add evolla rebase main

  #36232 commented on Jun 3, 2025 • 0 new comments

• Add Doge model

  #35891 commented on Jun 3, 2025 • 0 new comments

• Integrate xlstm cleanly.

  #35377 commented on Jun 4, 2025 • 0 new comments

• Correctly support resuming from checkpoint with a dataset without length

  #33544 commented on Jun 3, 2025 • 0 new comments

• Add Segment Anything 2 (SAM2)

  #32317 commented on Jun 4, 2025 • 0 new comments

• [Community contributions] Model cards

  #36979 commented on Jun 4, 2025 • 0 new comments

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

  #34446 commented on Jun 4, 2025 • 0 new comments

• Processor multiprocessing error when load custom processor

  #37637 commented on Jun 4, 2025 • 0 new comments

• MedGemma worked find prior to 4.52.3 release but now errors

  #38333 commented on Jun 4, 2025 • 0 new comments

• LagKV for key-value compression

  #38312 commented on Jun 4, 2025 • 0 new comments

• `ConditionalDetrImageProcessor` still accepts the deprecated parameter `max_size`

  #37939 commented on Jun 4, 2025 • 0 new comments

• Errors using TinyLlama-1.1B-Chat-v1.0 and DirectML

  #38340 commented on Jun 4, 2025 • 0 new comments

• Add RoMa keypoint matcher

  #36718 commented on Jun 3, 2025 • 0 new comments

• Maybe the vocab_size can be duplicated to the mainconfig for PEFT to pick up

  #38017 commented on Jun 3, 2025 • 0 new comments

• Shape Error in Llama4VisionMLP2

  #37321 commented on Jun 3, 2025 • 0 new comments

• [Bug] Gemma3Processor.apply_chat_template returns Tensor instead of dict with long multimodal few-shot inputs

  #37943 commented on Jun 3, 2025 • 0 new comments

• Alternative to trainer.hyperparameter_search for models used with custom optimizer / lrscheduler etc.

  #37945 commented on Jun 3, 2025 • 0 new comments

• Add examples that showcase the use of Hyperparameter search with Transformers

  #37947 commented on Jun 3, 2025 • 0 new comments

• [Contributions Welcome] Add Fast Image Processors

  #36978 commented on Jun 2, 2025 • 0 new comments

• Model implmenetation using Liger Kernel layers

  #38416 commented on Jun 2, 2025 • 0 new comments

• quantizer_hqq should not require a gpu/cuda device to run

  #38439 commented on Jun 2, 2025 • 0 new comments

• Add Gemma 3 For Sequence Classification

  #36755 commented on Jun 2, 2025 • 0 new comments

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

  #34928 commented on Jun 2, 2025 • 0 new comments

• request the support for training support for QuantizationMethod.FP8

  #37927 commented on Jun 2, 2025 • 0 new comments

• Updates in type-checking specifications have broken transformers' types

  #37928 commented on Jun 2, 2025 • 0 new comments

• Is Llama4TextL2Norm meant to be RMS norm?

  #37934 commented on Jun 2, 2025 • 0 new comments

• [i18n-TR] Translating docs to Turkish

  #27088 commented on Jun 1, 2025 • 0 new comments

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

  #38071 commented on Jun 1, 2025 • 0 new comments