2,217 commits

Author	SHA1	Message	Date
jiaqiw09	8669a22e9c	[fix] fix liger kernel patch for npu (#10583)	2026-06-16 18:21:52 +08:00
souljoy	cabe59a343	[model] add MiniCPM5-1B-Chat (#10558)	2026-06-10 16:18:27 +08:00
Co-Cl2	9ca4026efe	[model] handle unsloth model loading fallback during checkpoint resume (#7156) (#10551)	2026-06-09 01:01:01 +08:00
Ximing Xing	0b7aaf8f6a	[fix] correctly place new token embeddings when embedding is padded (#10547)	2026-06-05 10:47:51 +08:00
codingma	8a4f6a3da5	[model] add gemma-4-12B-it (#10549)	2026-06-04 23:43:20 +08:00
A1waysBeenHere	409e8a477f	[model] Patch GDN for NPU (#10504) ... Co-authored-by: jiaqiw09 <jiaqiw960714@gmail.com>	2026-06-04 16:39:02 +08:00
Cui-yshoho	053d43c0ac	[feat] support HyperParallel PT training and activation optimization (#10370)	2026-06-02 22:39:32 +08:00
Yaowei Zheng	8ef7335b6a	[misc] set dev version (#10533)	2026-05-31 00:16:07 +08:00
Yaowei Zheng	7af909522a	[version] release v0.9.5 (#10532)	2026-05-30 23:57:09 +08:00
xvxuopop	e016d2480e	[fix] Fix NPU FusedMoE and RMSNorm (#10512)	2026-05-30 21:42:54 +08:00
jiaqiw09	7d719182c9	[model] fix non-packing batch (bsz>1) for Qwen3.5 with flash attention (#10529)	2026-05-30 21:41:41 +08:00
jiaqiw09	01398eb18d	[v1] fix padding free with sp (#10513)	2026-05-26 23:49:21 +08:00
cxy	8e68764b65	[v1] Implement dynamic padding-free stretrgy for batching (#10507) ... Co-authored-by: cxy-thinkbook <xuanyuchen@seu.edu.cn>	2026-05-25 20:40:21 +08:00
Copilot	16ff5a23cb	[fix] use getattr for profiler attrs to support MCA TrainingArguments (#10506) ... Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: hiyouga <16256802+hiyouga@users.noreply.github.com>	2026-05-21 17:26:29 +08:00
jiaqiw09	bdcb92d035	[v1] Add FlashAttention selection and implement normal / padding-free / dynamic batching (#10469)	2026-05-21 17:14:19 +08:00
sunyi0505	7e20db5735	[v1] support liger_kernel (#10493)	2026-05-21 11:44:56 +08:00
浮梦	2322bf1cc2	[v1] add cuda fused moe kernel, implementing with triton (#10481)	2026-05-20 20:49:42 +08:00
浮梦	368c48968f	[callback] add torch profiler callback (#10463)	2026-05-20 20:47:52 +08:00
浮梦	8b5ea65770	[v1] support reward training stage (#10431) ... Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2026-05-20 20:46:52 +08:00
Dennis Huang	40e786d016	[data] add missing return statement in MiniCPM V Plugin (#10500) ... Co-authored-by: Cursor <cursoragent@cursor.com>	2026-05-20 01:50:00 +08:00
马境远	ca50f22c38	[fix] Fix MiniCPM-V-4.6 image preprocessing behavior (#10478)	2026-05-12 11:35:23 +08:00
马境远	53e77a9bfa	[model] support MiniCPM-V-4.6 (#10472)	2026-05-08 18:14:34 +08:00
Tai An	7e09152275	fix(data/converter): handle None tool_calls in OpenAI-style messages (#10455)	2026-05-07 17:44:41 +08:00
luca-888	8752280dd7	[data] Optimize QwenVL video dataset preprocessing (#10404) ... Co-authored-by: Kingsley <kingsleydodonow@gmail.com>	2026-05-03 18:36:56 +08:00
Kingsley	468723c5d9	[packing] fix GDN crash when meeting dummy image (#10453)	2026-05-01 12:10:13 +08:00
Peilin Li	887ee2b121	[refactor] Add KTransformers AMX MoE SFT support via Accelerate (#10430) ... Co-authored-by: mrhaoxx <mr.haoxx@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-05-01 01:47:58 +08:00
Kingsley	6b08b948c9	[misc] bump transformers version upperbound (#10446)	2026-05-01 01:30:11 +08:00
Hertz	f7f3bfcbd7	[model] support Hy3-Preview (#10432)	2026-04-29 23:21:13 +08:00
Kingsley	3475198d1e	[fa2] fix IMA when train qwen3_5 (#10448)	2026-04-29 20:20:55 +08:00
sunyi0505	50945ef850	[v1] fix device_mesh and sp for fsdp2 (#10429)	2026-04-28 11:20:11 +08:00
Octopus	2f0bef207a	[export] handle NotImplementedError in export_model for transformers>=5.0 (fixes #10410) (#10438) ... Co-authored-by: octo-patch <octo-patch@github.com>	2026-04-27 23:36:23 +08:00
curnane-lab	2092abc217	[npu] add Qwen3.5 support with Partial RoPE and Hybrid Attention (#10421) ... Co-authored-by: Curnane <mingliangfu@users.noreply.github.com>	2026-04-27 23:36:07 +08:00
Kingsley	99464b3d03	[misc] code lint (#10439) ... Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2026-04-27 14:07:31 +08:00
jiaqiw09	9a0cfdccfa	[v1] fix init on meta in transformers v5 (#10414)	2026-04-27 00:37:09 +08:00
Kingsley	c8890c32db	[data] support discard history cot for multiturn (#10435)	2026-04-27 00:32:44 +08:00
Kingsley	79c8332e4c	[train] add qwen35 patch for neat_packing (#10436)	2026-04-27 00:31:49 +08:00
jiaqiw09	e0bc3c1971	[v1] fix epoch and steps (#10422)	2026-04-23 17:29:06 +08:00
浮梦	ecca167eb4	[model] support qwen3.6 models (#10415) ... Co-authored-by: frozenleaves <frozen@Mac.local>	2026-04-22 19:44:01 +08:00
jiaqiw09	28a6ea1cdc	[v1] add deepspeed zero3 trigger for low memory usage weight loading (#10300)	2026-04-21 14:09:52 +08:00
sunyi0505	f5d739b132	[v1] fix device mesh and clip_grad_norm for ulysses cp (#10366)	2026-04-21 10:54:54 +08:00
浮梦	c4bbac49b2	[v1] support resume training from checkpoint (#10280) ... Co-authored-by: frozenleaves <frozen@Mac.local> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2026-04-20 20:28:08 +08:00
Cocoon-Break	c5aecaf31d	[data] fix SeedToolUtils.tool_extractor returns content when no tool calls found (#10408) ... Signed-off-by: Cocoon-Break <54054995+kuishou68@users.noreply.github.com>	2026-04-20 12:22:55 +08:00
Kingsley	436d26bc28	fix: projector lookup for gemma4 modules (#10382) ... Co-authored-by: yiluoAK_47 <yiluoAK_47@163.com>	2026-04-12 08:32:14 +08:00
Kingsley	c109c061e5	[model] set mm_projectors for omni models (#10378)	2026-04-10 18:12:57 +08:00
Kingsley	fa09c01c36	fix: gemma4 mm_token_type_ids padding (#10359)	2026-04-06 13:14:45 +08:00
Kingsley	eae6f0b541	[model] gemma4 (#10346)	2026-04-05 12:10:28 +08:00
Kingsley	acac63ef35	[data] fix qwen3vl timestamp (#10338)	2026-04-01 22:40:12 +08:00
浮梦	e5e8546493	[misc] fix moe (#10334) ... Co-authored-by: frozenleaves <frozen@Mac.local>	2026-03-31 23:04:45 +08:00
Cui-yshoho	97433c53b6	[feat] support LlamaFactory SFT training by HyperParallel FSDP2 backend (#10289)	2026-03-30 10:47:20 +08:00
sunyi0505	b5afabe3d2	[v1] support ulysses cp for fsdp2 (#10262)	2026-03-27 16:22:48 +08:00