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.

• Port ESMC and ESMFold2 to Transformers

  #46419 commented on Jul 30, 2026 • new comments

• Add Unlimited OCR

  #46836 commented on Jul 30, 2026 • new comments

• Support Granite Speech NAR (NLE)

  #46031 commented on Jul 30, 2026 • new comments

• wip step 3 7 flash

  #46658 commented on Jul 29, 2026 • new comments

• Add support for Nemotron Omni

  #46509 commented on Jul 25, 2026 • new comments

• [`FA`] Native torch integration

  #45153 commented on Jul 30, 2026 • new comments

• Executorch exporter fixes

  #47243 commented on Jul 29, 2026 • new comments

• Add MOSS-TTS-v1.5 (delay) and MOSS-Audio-Tokenizer

  #46447 commented on Jul 30, 2026 • new comments

• Exportable kimi

  #47096 commented on Jul 30, 2026 • new comments

• DO NOT MERGE - model creation skill

  #45213 commented on Jul 29, 2026 • new comments

• 🚨 Enable SDPA (and other attention backends) for T5 and propagate to the T5 family

  #47014 commented on Jul 30, 2026 • new comments

• Add Fun-ASR-Nano model

  #46180 commented on Jul 30, 2026 • new comments

• [CB] [Major] Add multimodal support to continous batching

  #46376 commented on Jul 23, 2026 • new comments

• Add Granite-swa and Granitemoe-swa model support

  #47179 commented on Jul 30, 2026 • new comments

• Add pose estimation keypoint preprocessing to Sapiens2ImageProcessor

  #47199 commented on Jul 27, 2026 • new comments

• Add V-JEPA 2.1 inference support

  #45497 commented on Jul 26, 2026 • new comments

• model: Add NVIDIA Canary-1B-v2 to Transformers

  #46825 commented on Jul 30, 2026 • new comments

• add MTP for GLM OCR and other MROPEs

  #47462 commented on Jul 29, 2026 • new comments

• [Improvement] Formalize KV sharing

  #47290 commented on Jul 27, 2026 • new comments

• Fix batched Qwen2.5/3-Omni audio generation

  #47186 commented on Jul 28, 2026 • new comments

• Implement VibeVoice

  #40546 commented on Jul 27, 2026 • new comments

• 🚨[wav2vec2] Support attn_implementation=sdpa dispatch

  #46196 commented on Jul 29, 2026 • new comments

• adding amd quark config class changes

  #47322 commented on Jul 30, 2026 • new comments

• Add heterogeneous model support (per-layer config and modeling)

  #45332 commented on Jul 30, 2026 • new comments

• Add new model `EfficientVit-SAM` to transformers

  #47275 commented on Jul 28, 2026 • new comments

• Improve Trainer DataLoader Controls for Streaming and Multiprocessing

  #47164 commented on Jul 29, 2026 • new comments

• fix bugs for clvp model

  #47127 commented on Jul 27, 2026 • new comments

• Enhance `get_balanced_memory` to ensure adequate GPU allocation for small models with large layers

  #47426 commented on Jul 30, 2026 • new comments

• Fix _is_packed_sequence for 1D and 3D position_ids

  #47342 commented on Jul 29, 2026 • new comments

• Batch Rebalance Data Sampler

  #47340 commented on Jul 28, 2026 • new comments

• Route byte-level llama tokenizers to TokenizersBackend

  #47017 commented on Jul 26, 2026 • new comments

• OpenVINO HF Exporter

  #47003 commented on Jul 27, 2026 • new comments

• TP dtensor

  #47194 commented on Jul 28, 2026 • new comments

• Also size the device_map buffer from the largest leaf module (follow-up to #47203)

  #47211 commented on Jul 29, 2026 • new comments

• pytorch 2.13 is released , updating transformer to handle the fix tf32 issue

  #47222 commented on Jul 25, 2026 • new comments

• Respect min_tokens_to_keep in TopHLogitsWarper

  #47262 commented on Jul 23, 2026 • new comments

• Pipeline parallel naive inference

  #47289 commented on Jul 28, 2026 • new comments

• [DeepSpeed] Composite models are not initialized correctly under ZeRO-3

  #47297 commented on Jul 23, 2026 • new comments

• Experiments with Neuron

  #47306 commented on Jul 25, 2026 • new comments

• Add chat input sanitization

  #47386 commented on Jul 29, 2026 • new comments

• fix(trainer): fix resume_from_checkpoint logic bug (issue #47375)

  #47391 commented on Jul 23, 2026 • new comments

• Add LoMa keypoint matching model

  #47401 commented on Jul 28, 2026 • new comments

• Make _added_tokens_decoder the single source of truth for added-token state

  #47440 commented on Jul 28, 2026 • new comments

• Fix ESMFold error in fp16 precision

  #47472 commented on Jul 23, 2026 • new comments

• Fix CLVP `c_proj` residual scaled-init silently never applied

  #47478 commented on Jul 29, 2026 • new comments

• Chat template inconsistencies in tool-calling support

  #45419 commented on Jul 23, 2026 • new comments

• ESMFold throws error when explicitly cast to fp16

  #47470 commented on Jul 23, 2026 • new comments

• generation error when same token in "forced_eos_token_id" and "supress_token" parameter.

  #24099 commented on Jul 24, 2026 • new comments

• MetalConfig quantization: batched generate silently corrupts all rows after the first — affine_qmm_t computes only batch element 0 of 3D inputs (verified, one-line fix)

  #47331 commented on Jul 24, 2026 • new comments

• Return updated attention mask from Wav2Vec 2.0

  #25307 commented on Jul 24, 2026 • new comments

• Added-token encoder cache goes stale when tokenizers mutate _added_tokens_decoder directly, giving ids that do not round-trip

  #47439 commented on Jul 25, 2026 • new comments

• YOLO model failing on multi GPU device map auto

  #46869 commented on Jul 25, 2026 • new comments

• Unable to resume model training with a locally downloaded model and trust remote code

  #46165 commented on Jul 25, 2026 • new comments

• Is transformers 5.12.1 requires tokenizers<=0.23.0,>=0.22.0 really right?

  #46785 commented on Jul 26, 2026 • new comments

• [BUG] FP8 model cannot inference on B200 (SM>=100)

  #46209 commented on Jul 27, 2026 • new comments

• tokenizers<=0.23.0 upper bound blocks the only existing 0.23.x release

  #47429 commented on Jul 28, 2026 • new comments

• Batch size schedulers

  #31222 commented on Jul 28, 2026 • new comments

• ZeRO-3 zero.Init does not partition composite minimax_m3_vl language submodule -> OOM on multi-GPU load

  #46822 commented on Jul 29, 2026 • new comments

• Allow to give the dataset multiprocessing_context

  #34793 commented on Jul 29, 2026 • new comments

• Please update `tokenizers` version check

  #45736 commented on Jul 29, 2026 • new comments

• Checkpoint conversion silently loads transposed fused-expert weights when the swapped dims are equal (Transpose(check_dims=True), qwen3_vl_moe)

  #47031 commented on Jul 29, 2026 • new comments

• FineGrainedFP8 DeepGEMM path returns silently wrong results on SM100 (UE8M0 scale rounding without requantization)

  #47030 commented on Jul 29, 2026 • new comments

• Remove device to host sync triggered in _flash_attention_forward

  #39213 commented on Jul 29, 2026 • new comments

• Mamba Models - Missing MambaForSequenceClassification

  #30431 commented on Jul 29, 2026 • new comments

• SentencePieceBackend should decode byte-fallback tokens

  #47473 commented on Jul 29, 2026 • new comments

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

  #39921 commented on Jul 29, 2026 • new comments

• Add Molmo2

  #43451 commented on Jul 30, 2026 • new comments

• add HyperClovaX Vision

  #44314 commented on Jul 29, 2026 • new comments

• 🚨🚧 FeatureExtractor → AudioProcessor

  #44394 commented on Jul 28, 2026 • new comments

• Add Audio-Visual Flamingo model

  #45586 commented on Jul 28, 2026 • new comments

• Enable MTP support for Qwen3.5

  #45638 commented on Jul 28, 2026 • new comments

• [generation] Encode multimodal data only once

  #45783 commented on Jul 29, 2026 • new comments

• LocateAnything 3B

  #46300 commented on Jul 25, 2026 • new comments

• 🚨 [DiffusionGemma] Return router logits and load balancing loss

  #46642 commented on Jul 27, 2026 • new comments

• Fix CPM-Ant LM head tying (vocabulary-slice tie)

  #46835 commented on Jul 27, 2026 • new comments