Search results

Repetition penalty is gauge-dependent. Proposing opt-in normalized variant.

Feature requestRequest for a new featureRequest for a new feature

Status: Open.

#47595 In huggingface/transformers;

· captainpete opened on Jul 28, 2026

NemotronH: no longer respected after #47452 (unconditionally fetches Hub kernel)

Status: Open.

#47577 In huggingface/transformers;

· albertvillanova opened on Jul 27, 2026

[serge] integration failure triage - 2026-07-27

Status: Open.

#47567 In huggingface/transformers;

· github-actions opened on Jul 27, 2026

create_bidirectional_sliding_window_mask builds the mask on encoder_hidden_states' device — same model-parallel bug as #46221, which only fixed create_bidirectional_mask

Status: Open.

#47559 In huggingface/transformers;

· abcgco opened on Jul 26, 2026

Wav2Vec2 tokenizer rebuilds special-token lists per token in decode paths; phoneme variant carries a filter that never matches

Status: Open.

#47556 In huggingface/transformers;

· ishan-1010 opened on Jul 26, 2026

"ValueError: Tokenizer class TokenizersBackend does not exist or is not currently imported"

bug

Status: Open.

#47553 In huggingface/transformers;

· DXXS opened on Jul 25, 2026

Mamba2-family cuda_kernels_forward decode step is broken since #47452: missing seq-dim squeeze

Status: Open.

#47532 In huggingface/transformers;

· qgallouedec opened on Jul 25, 2026

Phi3/Phimoe: prepare_inputs_for_generation forwards logits_to_keep=None into forward() -> 4-dim logits -> torch.multinomial crash

bug

Status: Open.

#47530 In huggingface/transformers;

· pwasiewi opened on Jul 24, 2026

Transformers 5.13 CPU memory growth loading Qwen3-235B-A22B with DeepSpeed ZeRO-3; 4.51.0 remains bounded

bug

Status: Open.

#47514 In huggingface/transformers;

· narankhuut opened on Jul 24, 2026

MistralCommonBackend API documentation appears to render the dummy implementation instead of the actual class

Status: Open.

#47504 In huggingface/transformers;

· samyuktahegde opened on Jul 23, 2026

DataCollatorWithFlattening uses 0-based position_ids, silently degrading RoBERTa-family models

bug

Status: Open.

#47496 In huggingface/transformers;

· tomaarsen opened on Jul 23, 2026

[BUG] CANINE: dynamic ONNX export fixes sequence length to example input

bug

Status: Open.

#47491 In huggingface/transformers;

· sgpublic opened on Jul 23, 2026