transformers · Issues· 2421 open

Support 2D bool mask for logits_to_keep, expose last_hidden_state

VibeVoice-ASR: float32 ceil in get_audio_features miscounts audio tokens for long audio, crashing generate() with a mask shape mismatch

HfArgumentParser cannot parse dict-typed dataclass fields from the CLI (registers them with type=dict)

Non-persistent buffers materialized uninitialized by from_pretrained (_move_missing_keys_from_meta_to_device)

Qwen3.5-9B fine-tuning with run_clm.py on 8x B200 ran the gated DeltaNet layers on the fp32 torch chunk loop because flash-linear-attention was not installed

Please remove dependency on the unmantained module pyctcdecode: it wasn't updated for years and is still at numpy-1.x

Four MoE models expose `aux_loss` in their output type and never set it

gemma 4 can not load audio file and produce Audio features and audio tokens do not match, tokens: 31, features: 468480

Vectorized (batched) logits processors for assisted decoding

[Qwen3.8-flash-next] QSAIndexer can not train

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

GenerationConfig should validate that pad_token_id is not in eos_token_id list

Llama 4 declares `output_router_logits` and `router_aux_loss_coef` but never reads them

[i18n-zh] Translating docs to Chinese (Simplified)

Rewrite audio feature processors to accept torch tensors