Issues

• [BUG] Gemma 4 assistant models fail to build since KV sharing was formalized

  bug
  huggingface/transformers#50316 · Filter by author harshaljanjani opened on Nov 3, 2026
  1

• ZeroShotObjectDetectionPipeline silently drops batched images and misroutes candidate labels

  bug
  huggingface/transformers#50315 · Filter by author BlazeSR10hawk opened on Nov 3, 2026

• Trainer: resume from a checkpoint for models sharded at load time (DistributedConfig)

  Feature request
  huggingface/transformers#50306 · Filter by author qgallouedec opened on Nov 3, 2026

• Quantizing gpt-oss on load with Mxfp4Config() crashes its forward and saves noise

  huggingface/transformers#50300 · Filter by author qgallouedec opened on Nov 3, 2026

• Docs: hidden_states[-1] of DINOv3 / ViT changed from post-norm to pre-norm in 4.57.2 (not in release notes or migration guide)

  huggingface/transformers#50297 · Filter by author ravediamond opened on Nov 3, 2026
  3

• [serge] integration failure triage - 2026-11-02

  huggingface/transformers#50280 · Filter by author github-actions opened on Nov 2, 2026
  1

  

• QuantizedCache.crop() raises IndexError across FP8, Quanto and HQQ, breaking speculative decoding

  bug
  huggingface/transformers#50273 · Filter by author BlazeSR10hawk opened on Nov 2, 2026

• Saving a gpt-oss loaded with Mxfp4Config(dequantize=True) loses the experts

  huggingface/transformers#50270 · Filter by author qgallouedec opened on Nov 2, 2026

• PaddleOCR-VL: position embedding interpolation changed from align_corners=False to True

  bugShould Fix
  huggingface/transformers#50241 · Filter by author scyyh11 opened on Nov 2, 2026
  4

• Add a calibration test for WatermarkDetector

  huggingface/transformers#50232 · Filter by author aparnaittekot opened on Nov 1, 2026
  1

• [Doge] DogeCDMoE router .view(2, bsz * seq_len, -1) mixes tokens: with is_moe=True, logits depend on future tokens and on other sequences in the batch

  huggingface/transformers#50230 · Filter by author Andreas-Hum opened on Nov 1, 2026
  4

• TopHLogitsWarper does not implement the documented Top-H rule (it bounds the unnormalized partial entropy, and takes H(p) from the top 100 tokens only)

  huggingface/transformers#50228 · Filter by author Andreas-Hum opened on Nov 1, 2026
  9

• [CLAP] ClapModel(return_loss=True) computes the text-to-audio loss twice and never the audio-to-text loss

  huggingface/transformers#50227 · Filter by author Andreas-Hum opened on Nov 1, 2026
  4

• CI: vllm-ci failing: uv 'No solution found' torch==2.14.0+cpu unavailable for vllm cpu build; step uses sh with [[

  huggingface/transformers#50223 · Filter by author fitzyracing1 opened on Nov 1, 2026
  2

• CI: Torch pipeline CI failing: 39 pipeline tests fail with RuntimeError ignore_mismatched_sizes=False

  huggingface/transformers#50222 · Filter by author fitzyracing1 opened on Nov 1, 2026
  1

• CI: check_tiny_models failing: from datasets import load_dataset ImportError — pyarrow needs NumPy>=2.0 (found 1.26.4)

  huggingface/transformers#50221 · Filter by author fitzyracing1 opened on Nov 1, 2026
  4

• CI: build-docker-images Doc builder failing: pyarrow requires NumPy>=2.0 but image has 1.26.4

  huggingface/transformers#50220 · Filter by author fitzyracing1 opened on Nov 1, 2026
  1

• ModernBERT mean pooling overflows in float16 on long inputs and returns NaN logits

  huggingface/transformers#50216 · Filter by author clemlesne opened on Nov 1, 2026
  4

• SDPA launch failures on ROCm Windows surface as misleading async errors with no fallback — propose error translation or startup probe

  huggingface/transformers#50211 · Filter by author AIwork4me opened on Nov 1, 2026
  2

• [Bug]: resume_from_checkpoint silently skips the root-level (trained) adapter when the checkpoint contains adapter subdirectories — DPO loss stuck at ln(2) ≈ 0.6931, no error or warning

  bug
  huggingface/transformers#50204 · Filter by author TheSeaByTheMorning opened on Oct 31, 2026
  4

• DeBERTa: from_pretrained(dtype=torch.bfloat16) leaves q_bias/v_bias in float32, forward fails (expected m1 and m2 to have the same dtype)

  huggingface/transformers#50200 · Filter by author tarekziade opened on Oct 31, 2026
  

• GGUF reader: llama.cpp Mixtral loads as dense Llama with random MLPs; Qwen3 head_dim and Mistral detection not mapped

  huggingface/transformers#50197 · Filter by author JohnnyTeutonic opened on Oct 31, 2026
  1

• GGUF: Gemma 3 loses global-layer RoPE linear scaling and query_pre_attn_scalar when loaded from a llama.cpp GGUF

  huggingface/transformers#50196 · Filter by author JohnnyTeutonic opened on Oct 31, 2026
  4

• Mamba (Mamba-1): forward of several tokens over a primed cache restarts the selective scan from zero and returns wrong logits

  huggingface/transformers#50195 · Filter by author JohnnyTeutonic opened on Oct 31, 2026
  4

• Multimodal pipelines silently discard generated candidates during postprocessing

  bug
  huggingface/transformers#50194 · Filter by author BlazeSR10hawk opened on Oct 31, 2026