huggingface/transformers

Pull Requests Commits

Point Metal kernel repo to ArthurZ/gguf-dequant until transfer to kernels-community

ArthurZucker committed 58 days ago

b4fb5d71

serve: auto-select metal-flash-sdpa attention on MPS

ArthurZucker committed 58 days ago

9fc9687b

GGUF: optional Metal dequant fast path via kernels-community/gguf-dequant

ArthurZucker committed 58 days ago

be73645e

Fix Gemma2/3 GGUF rename rules + drop GGUF runtime-aux tensors

ArthurZucker committed 58 days ago

0cd7bce4

Free per-source tensors in fuse ops + drain MPS pool between stages

ArthurZucker committed 58 days ago

ed9c07e7

Restore gpt_oss rope_scaling reconstruction + minimax_m2 scoring_func mapping; minor GGUFQuantizedTensor perf

ArthurZucker committed 59 days ago

ec2f3198

Update src/transformers/modeling_gguf_pytorch_utils.py

ArthurZucker committed 59 days ago

Verified b8832746

Update src/transformers/modeling_gguf_pytorch_utils.py

ArthurZucker committed 59 days ago

Verified 0cb151ce

Guard heavy imports behind is_torch_available() — fixes tokenizer-only installs

ArthurZucker committed 59 days ago

67de05d9

Fix MPS dequant + Qwen biases; further consolidate T5 FFN

ArthurZucker committed 59 days ago

0b23297f

Fix type check: load_config.hf_quantizer can be None in is_hqq_or_quark guard

ArthurZucker committed 59 days ago

2203bc30

docs: add 'Adding GGUF support for a new architecture' contributor guide

ArthurZucker committed 59 days ago

33c7f99d

Simplify GGUF renaming with regex alternations; minimise diff vs main

ArthurZucker committed 59 days ago

f7378ff5

Inline GGUF spawn into GGUFQuantizer; drop spawn_gguf_materialize

ArthurZucker committed 59 days ago

cf271730

Dequant on the target device; drop gguf-py fallback

ArthurZucker committed 59 days ago

06589a69

Replace gguf-py NumPy dequant with pure-PyTorch port (3-8x faster)

ArthurZucker committed 59 days ago

d6d63593

Drop gguf_rename factory; use plain WeightRenaming/WeightConverter

ArthurZucker committed 59 days ago

1b33217a

Refactor GGUF loading: new converter chain & GGUFQuantizer

ArthurZucker committed 59 days ago

b027f8ae

Require input_ids for repetition penalty (#45389)

ruben-aghayan committed 59 days ago

Verified eee7039b

Fix undefined 'input' variable (#45895)

fullyz committed 59 days ago

Verified e51a963d

Deepseek v4 csa mask collapse (#45928)

ArthurZucker committed 59 days ago

Verified 2ad5a9b8

[fix] Add `fatal_error` to `ContinuousBatchingManager` so the serving layer can detect a dead CB worker (#45898)

qgallouedec committed 59 days ago

Verified 9c8a1b83

Fix fsdp2 for gemma models with shared kv states (#45912)

Cyrilvallez committed 59 days ago

Verified 0d476f2c

[docs] update model cards (#45612)

stevhliu committed 59 days ago

Verified 7ee56fc2

torch.backends.fp32_precision cascade conv/rnn so removing the temp fix (#45777)

khushali9 committed 60 days ago

Verified 4968b639

Fix slow Trainer path with 4D attention mask (#45852)

qflen committed 60 days ago

Verified 909ece88

:rotating_light: [`Attn`] Remove all old mask APIs from modeling (#43924)

vasqu committed 60 days ago

Verified 5206626c

fix(text-generation): use token-level slicing for return_full_text=Fa… (#45860)

Abineshabee committed 60 days ago

Verified 91ba4c9a

[CI] AMD docker: bump to ROCm 7.2.2 / PyTorch 2.10 + prebuilt FA wheel (#45913)

Abdennacer-Badaoui committed 60 days ago

Verified 0d34e1ab

Automatically inherit properties from children in composite models (#45909)

Cyrilvallez committed 60 days ago

Verified a25b8efa

Older