Issues

• SmolVLM: built-in video chat template drops non-empty string message content

  huggingface/transformers#50613 · Filter by author arjhinety opened on Nov 11, 2026

• TimesFM 2.5: training loss feeds the point-forecast channel to the quantile loss

  bug
  huggingface/transformers#50612 · Filter by author m1o0 opened on Nov 11, 2026
  

• finegrained_fp8 models are silently mis-dequantized when loaded on CPU if a quantized weight has partial blocks

  bug
  huggingface/transformers#50603 · Filter by author YancyLii opened on Nov 11, 2026
  1

• Default configs for 11 models can't build or run their own model (CohereCompass, DiffusionGemma, DeepseekOcr2, Granite4Vision, HunYuanVL, Moshi, PPOCRV6MediumDet, Qwen4Exp, Qwen3ASR, Molmo2, Step3p7)

  huggingface/transformers#50601 · Filter by author hameerabbasi opened on Nov 11, 2026
  2

• Inkling: BF16 router rounds distinct scores into ties before Top-K

  Triaged
  huggingface/transformers#50599 · Filter by author taking-lying-flat opened on Nov 11, 2026
  1

• Universal speculative decoding with a pruned assistant LM head can map logits to the wrong target tokens

  huggingface/transformers#50598 · Filter by author Incheonkirin opened on Nov 11, 2026

• [serge] integration failure triage - 2026-11-10

  huggingface/transformers#50597 · Filter by author github-actions opened on Nov 10, 2026
  

• FineGrainedFP8 dequantize with dtype=float32 leaves non-quantized weights in bfloat16

  huggingface/transformers#50594 · Filter by author qgallouedec opened on Nov 10, 2026

• center_crop pads the odd leftover pixel on opposite sides in the torchvision vs numpy/PIL backends

  huggingface/transformers#50584 · Filter by author SID-6921 opened on Nov 10, 2026
  4

• v5: full fine-tuning a bf16 checkpoint with Trainer silently runs pure bf16 AdamW, and most updates round away

  huggingface/transformers#50581 · Filter by author Andreas-Hum opened on Nov 10, 2026
  2

• Docs: dead links in model_doc pages (12 URLs return 404)

  huggingface/transformers#50568 · Filter by author Hisernberg opened on Nov 10, 2026
  1

• apply_chat_template crashes with DeepSeek-V3/R1 templates on OpenAI-style tool-call messages since #45422

  bug
  huggingface/transformers#50551 · Filter by author albertvillanova opened on Nov 10, 2026
  1

  

• FSDP checkpoint resumption deadlocks during collective optimizer state sync with PEFT LoRA in Trainer

  bug
  huggingface/transformers#50549 · Filter by author utkarshxdev opened on Nov 10, 2026

• MoE load-balancing loss adds no gradient under reentrant gradient checkpointing

  huggingface/transformers#50545 · Filter by author hivaze opened on Nov 10, 2026
  2

• quantizer_torchao imports torchao.prototype at module load, which initializes CUDA when torchao is installed

  huggingface/transformers#50540 · Filter by author Thangnguyenvn98 opened on Nov 9, 2026
  1

• Wav2Vec2-family feature encoder sets requires_grad on an intermediate every training step, breaking torch.compile(fullgraph=True)

  bug
  huggingface/transformers#50527 · Filter by author kamboj-sajan opened on Nov 9, 2026
  4

• Wav2Vec2CTCTokenizer drops the word delimiter | around <unk> (and other special tokens) when encoding

  bug
  huggingface/transformers#50519 · Filter by author jcsilva opened on Nov 9, 2026
  1

• save_pretrained deletes shards of an existing variant when saving default weights to the same folder

  bug
  huggingface/transformers#50491 · Filter by author Dhruv-Sharma29 opened on Nov 9, 2026

• New model addition: Iwin Transformer (Hierarchical Vision Transformer using Interleaved Windows)

  New model
  huggingface/transformers#50489 · Filter by author Cominder opened on Nov 9, 2026
  2

• KimiLinear: padding in cached continuation advances convolution and recurrent states

  huggingface/transformers#50484 · Filter by author taking-lying-flat opened on Nov 9, 2026
  2

• EfficientNet/ALIGN/MobileNet pass a Keras-style BatchNorm momentum to nn.BatchNorm2d, so running stats are the last training batch

  huggingface/transformers#50473 · Filter by author Andreas-Hum opened on Nov 8, 2026
  2

• DynamicCache.batch_select_indices leaves Falcon-H1 convolution and recurrent states unchanged

  bug
  huggingface/transformers#50470 · Filter by author adl5423 opened on Nov 7, 2026
  1

• Weight-preserving patch embedding initialization for vision models with non-RGB channels (num_channels != 3)

  Feature request
  huggingface/transformers#50460 · Filter by author utkarshxdev opened on Nov 7, 2026
  1

• Qwen3-ASR: masked mel frames affect valid CNN outputs

  huggingface/transformers#50453 · Filter by author taking-lying-flat opened on Nov 7, 2026

• LayoutXLM / LayoutLMv2 / MarkupLM tokenizers label several sub-tokens per word with only_label_first_subword=True (offset-based first-token check)

  huggingface/transformers#50452 · Filter by author 14H034160212 opened on Nov 7, 2026
  2