transformers
> 数据库🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference a
About
🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference a
Issues· 2355 open
- #49252
AutoTokenizer with tokenizer_class=LlamaTokenizerFast silently builds a sentencepiece pipeline for a byte-level BPE tokenizer.json, dropping all CJK input
bugUpdated Oct 2, 2026 - #49233
Encoder repetition penalty uses incorrect source rows after beam or sampled-return expansion
bugUpdated Oct 2, 2026 - #49238
ExponentialDecayLengthPenalty creates NaN EOS scores when combined with minimum-length constraints
bugUpdated Oct 2, 2026 - #49190
[eager MOE] Selected expert IDs should probably not be treated as 0-d GPU tensors
bugUpdated Oct 2, 2026 - #48428
RoPE corruption due to TF32 on B200 (possibly others) when using `torch.set_float32_matmul_precision("high")`
bugUpdated Oct 2, 2026 - #49166
Rename `gradient_checkpointing` to `activation_checkpointing` in the next major
Updated Oct 2, 2026 - #49237
Trainer counts fp16 steps skipped by GradScaler as completed update steps (`global_step`, `max_steps`, `on_optimizer_step`)
Updated Oct 2, 2026 - #49241
[serge] integration failure triage - 2026-10-01
Updated Oct 2, 2026 - #48256
Right-padded prefill produces a corrupted cache in Mamba-family models (zero conv state, decayed SSM state)
Updated Oct 1, 2026 - #48308
Native TP does not shard qwen3_vl_moe experts (weights fully replicated per rank)
Updated Oct 1, 2026