Repository navigation

All pull requests

[train] Left-pad generation during scheduled SFT evaluation
#10892YueshenZ opened Oct 14, 20262 tasks done
[data] support non-string enum values in the default tool format
#10891cyx5445 opened Oct 14, 2026
[train] Fix hinge and IPO objectives with TRL 0.20+
#10889YueshenZ opened Oct 13, 2026
[data] composite transparent images on a white background
#10888drakeo338 opened Oct 12, 20262 tasks done
[data] fix gemma4 tool_extractor reading colons inside string arguments as keys
#10885hpdkhoa opened Oct 10, 20262 tasks done
[data] pad position_ids before unpadding on the neat_packing FA2 path
#10884hpdkhoa opened Oct 10, 20262 tasks done
[train] check TRL version for PPO when parsing arguments
#10882stanleys12 opened Oct 9, 20262 tasks done
docs: install contributor tools from requirements/dev.txt
#10878Styleofm3 opened Oct 8, 20261 of 2 tasks
feat: adopt HyperParallel unified model setup
#10877Cui-yshoho opened Oct 8, 2026Contributor2 tasks done
[xpu] support profiler, module profiler and env report on Intel GPU
#10876chinyixiang opened Oct 8, 20262 tasks done
[export] keep exported tokenizer_config loadable with transformers<5
#10875er-s-an opened Oct 7, 20262 tasks done
feat(data): precompute length column for efficient group_by_length sa…
#10874VovyH opened Oct 7, 2026
[v1] reduce CUDA fused MoE activation memory
#10873CharlesXu-HQ opened Oct 7, 20262 tasks done
[v1] add batched Muon orthogonalization
#10872CharlesXu-HQ opened Oct 7, 20262 tasks done
[deps] allow accelerate 1.15
#10871qingshungLI opened Oct 6, 20262 tasks done
[export] sync export dtype to sub-configs of composite models
#10870aron-intframe opened Oct 2, 2026
[deps] upgrade transformers to 5.17.0
#10869chinyixiang opened Oct 2, 20261 of 2 tasks
enable fp16 in dtype inference on Intel XPU
#10868chinyixiang opened Oct 2, 20261 of 2 tasks
[fix] preserve v0 torchrun argument boundaries
#10867NaCr05 opened Oct 2, 20262 tasks done
[chat] judge finish_reason against the full stop token set
#10866longzhenren opened Oct 1, 2026
[test] check attn implementation instead of attention class names in test_attention
#10865singhalshubham03 opened Oct 1, 2026Contributor
[v1] fix fake MoE model in test_fsdp2_weight_convert for transformers v5
#10864singhalshubham03 opened Oct 1, 2026Contributor
[data] add tiny_aya and gemma4n_nothink templates
#10863adzcai opened Oct 1, 20262 tasks done
fix: respect temporal_patch_size in Qwen video grid metadata
#10862aniketkrs opened Oct 1, 2026
1
[webui] add an optional modern layout for LLaMA Board (LLAMABOARD_UI=modern)
#10861Li2O0577 opened Oct 1, 20262 tasks done

Pull requests