Go
Home
FAQ
Install
Home
FAQ
Install
Login
via GitHub
huggingface/transformers
Pull Requests
Commits
Open
Closed
Fix Unigram tokenizers built from a bare `spiece.model` splitting text at user-defined pieces
#51084 opened 2026-11-22 05:38 by
XiaoyuanWang111
[Trainer] Preserve MoE router auxiliary loss with label smoothing
#51083 opened 2026-11-22 05:12 by
taking-lying-flat
Add Qwen2-VL and Qwen2.5-VL to the GGUF loading path
#51081 opened 2026-11-22 04:56 by
taeyang0505
Trainer: upcast fp16 models to fp32 when fp16 AMP is enabled
#51080 opened 2026-11-22 04:54 by
Kayvan-Zahiri
Keep left padding when AutoTokenizer loads TokenizersBackend in place of LlamaTokenizer
#51078 opened 2026-11-22 02:49 by
vedant-1807
Fix Gemma3n sparse MLP synchronization in eager CUDA inference
#51077 opened 2026-11-22 00:52 by
hlin99
fix: pass position_ids to create_causal_mask in GptOss, AfMoE, and GraniteMoeHybrid (#51072)
#51074 opened 2026-11-21 20:20 by
shotsan
Fix batch-dependent text output in image-text-to-text and any-to-any pipelines
#51073 opened 2026-11-21 20:04 by
Enricoftw
[docs] Fix the output recorder registration in the MoE routing replay recipe
#51071 opened 2026-11-21 18:51 by
AlexxRussell
[Chat Parsing] Preserve explicit string arguments in typed json and python-call-list tool calls
#51069 opened 2026-11-21 17:20 by
xgbtxy
Fix axk1 / dots3_note integration tests after fp32 router gate change
#51067 opened 2026-11-21 15:11 by
ydshieh
[distributed] experts offload: read misses with safetensors' Reader
#51066 opened 2026-11-21 15:00 by
McPatate
[CB] GLM-5.3 faster than vLLM on 8x B300 (draft, to be split)
#51065 opened 2026-11-21 13:26 by
qgallouedec
Fix T5Tokenizer dropping newlines when vocab contains whitespace tokens
#51064 opened 2026-11-21 12:57 by
AyhamJo7
Fix universal speculative decoding crash when the assistant shares the target tokenizer
#51063 opened 2026-11-21 12:33 by
samichellia90-bit
Inkling2 Mini: native NVFP4 / MXFP8 experts on MPS
#51062 opened 2026-11-21 11:14 by
ArthurZucker
fix(mluke): keep the fairseq token ids when loading a SentencePiece checkpoint
#51059 opened 2026-11-21 03:57 by
houyunsun
[Whisper] Avoid encoder self-attention maps for token timestamps
#51058 opened 2026-11-21 03:25 by
luanweslley77
Fix dead links to onnxruntime issues in exporters doc
#51057 opened 2026-11-20 23:16 by
Dani-Google
Saving: copy the gathered files to CPU on all ranks at once
#51055 opened 2026-11-20 22:29 by
michaelbenayoun
Saving: checkpoint consolidation
#51053 opened 2026-11-20 20:57 by
michaelbenayoun
Ignore masked padding in `encoder_repetition_penalty`
#51052 opened 2026-11-20 20:28 by
sohailshk
[docs] Replace deprecated torch_dtype with dtype in doc examples
#51051 opened 2026-11-20 20:20 by
abhitorch81
Cast the Inkling vision tower input to its weight dtype
#51048 opened 2026-11-20 18:45 by
albertvillanova
[docs] Fix the AWQ install step and ExLlamaV2 example
#51046 opened 2026-11-20 18:36 by
phatstraws
perf(data_collator): make _whole_word_mask O(batch * length)
#51045 opened 2026-11-20 17:55 by
DrSkyle
regression add_bos_token, add_eos_token preserve tokenizer.json behav…
#51044 opened 2026-11-20 17:45 by
itazap
[cb] sample per-request settings without sorting the vocabulary
#51043 opened 2026-11-20 16:53 by
SunMarc
Fix strided checkpoint reads on PyTorch 2.9
#51042 opened 2026-11-20 16:50 by
satsuki2486441738
Fix NumPy output for DataCollatorForMultipleChoice
#51041 opened 2026-11-20 16:40 by
nazek52
Older