github-featured/modelscope--ms-swift
2
0
Fork 0
Code Issues Pull Requests Deployments Models Agent Notes Packages Projects Releases Wiki Activity

Commit Graph

  • afd33b1129 lint main 雨泓 2026-08-24 15:43:44 +08:00
  • 93c1651d31 fix: honor sampled fps in Qwen-Omni video_second_per_grid (#9955) Light 2026-08-24 15:41:58 +08:00
  • 481d7dbb9e perf(rlhf): vectorize padding-free sequence reductions (#9958) Ruihan11 2026-08-24 15:39:07 +08:00
  • c1972888dd Fix Qwen3.5 A_log reinitialization (#9957) cherry77-cloud 2026-08-24 15:32:43 +08:00
  • 8e3f164121 fix(dataset): honor explicit hub prefixes (#9971) YZJF,YCDG,DJLY,ZZZB 2026-08-24 15:04:53 +08:00
  • d6644babab fix(metrics): count empty NLG predictions in score averages (#9962) Excelius 2026-08-24 14:59:22 +08:00
  • ec8d1b010e fix(agent): serialize Qwen tool argument JSON literals (#9968) Zhijun-Xu 2026-08-24 13:45:39 +08:00
  • 27ac25c8d1 Fix MOSS-VL align test default model id for ModelScope hub (#9974) SSSSuperC 2026-08-24 13:36:50 +08:00
  • a8e8631dbc fix(utils): scope download cache by URL (#9969) YZJF,YCDG,DJLY,ZZZB 2026-08-24 11:50:20 +08:00
  • b1b3f97b95 wip dev-swift-v5 z0o0ey 2026-08-23 17:48:16 +08:00
  • 3602360a56 sync missing version bump on main (#9967) jinghanhu 2026-08-22 23:51:23 +08:00
  • 86e55a5d9a [NPU] Adapt vLLM-Ascend 0.23 runtime (#9959) addsubmuldiv 2026-08-21 17:31:26 +08:00
  • 5f50af04fc Support MOSS-VL (#9944) SSSSuperC 2026-08-21 15:39:50 +08:00
  • 7d354df70d fix:Compatibility with Python 3.14 (#9961) z0o0ey 2026-08-21 15:23:47 +08:00
  • abfe0c4476 fix：(megatron 0.16.1) support PackedSeqParams without total_tokens (#9960) Hazeldxq 2026-08-21 14:43:54 +08:00
  • b43edd2270 wip z0o0ey 2026-08-21 10:09:01 +08:00
  • a54a4ae5c8 Fix save_missing_weights arg (#9956) tastelikefeet 2026-08-21 00:10:55 +08:00
  • 86d4403fab The ViT gradient checkpoint was incorrectly turned off during multimodal model training. (#9945) z0o0ey 2026-08-20 09:46:23 +08:00
  • e78a63caa5 Merge branch 'dev-swift-v5' of github.com:modelscope/ms-swift into dev-swift-v5 z0o0ey 2026-08-19 21:29:13 +08:00
  • 868385000f wip z0o0ey 2026-08-19 21:28:49 +08:00
  • 4a49a70b83 docs(grpo): fix reward model spelling (#9940) YZ 2026-08-19 15:03:52 +08:00
  • 302b074dd4 test(ci): drop legacy structbert tests and dependency pins (#9938) jinghanhu 2026-08-18 19:45:03 +08:00
  • 3a76ee2c67 fix:change lint chekc to ruff (#9939) z0o0ey 2026-08-18 15:41:37 +08:00
  • 1240f48238 doc:Frequently asked questions update(#9937) z0o0ey 2026-08-18 13:40:35 +08:00
  • 191313a4f5 feat(npu): Add FSDP2 LoRA example for Qwen3.8-27B for Ascend NPU (#9920) Rain 2026-08-18 10:44:13 +08:00
  • 1a1ba3ee86 Support UEmbed-2B (#9930) z0o0ey 2026-08-17 14:36:42 +08:00
  • 11b23c2615 Keep SGLang and LMDeploy streaming response IDs stable (#9926) cherry77-cloud 2026-08-17 10:20:47 +08:00
  • e7b50c398a [Megatron] Fix duplicated GRPO KL gathering (#9921) cherry77-cloud 2026-08-17 01:10:25 +08:00
  • cf6316ca2b fix: handle tool messages in last round loss scaling (#9923) Bakuma-sea 2026-08-16 21:53:23 +08:00
  • bc231c8ed7 Fix single token embedding (#9924) tastelikefeet 2026-08-16 21:33:54 +08:00
  • 197d31f303 fix: carry rounded seconds in format_time so it never prints '60s' (#9919) Coro 2026-08-16 02:23:30 -06:00
  • ff5128777d bump version v4.5.2 release/4.5 雨泓 2026-08-16 10:23:55 +08:00
  • cbaf2de9e4 remove duplicate model-ids 雨泓 2026-08-16 10:22:34 +08:00
  • 4065c1d061 remove duplicate model-ids 雨泓 2026-08-16 10:22:34 +08:00
  • 7aebb021ea bump version v4.5.1 雨泓 2026-08-16 10:00:59 +08:00
  • 8e07ec0ebb fix wrong model-ids 雨泓 2026-08-16 09:58:21 +08:00
  • a45f1d4f73 fix wrong model-ids 雨泓 2026-08-16 09:58:21 +08:00
  • 074cacc1b8 bump mcore-bridge v4.5.0 hjh0119 2026-08-15 00:34:25 +08:00
  • ce8b083a2b bump version hjh0119 2026-08-15 00:30:32 +08:00
  • cc25166a11 fix(README): restore broken star history chart (#9913) PingouinFerreux 2026-08-14 18:25:20 +02:00
  • 5f894d3f54 rename qwen3.5 doc (#9918) jinghanhu 2026-08-14 19:53:54 +08:00
  • ab726e9d44 [model] qwen3.8 (#9914) jinghanhu 2026-08-14 19:43:14 +08:00
  • 598053e186 remove auto_config patch for vllm (#9911) jinghanhu 2026-08-14 18:03:30 +08:00
  • 00fc223be0 fix nemotron grpo (#9905) jinghanhu 2026-08-14 18:02:54 +08:00
  • c08a1ca7b4 fix megatron common pt save (#9908) jinghanhu 2026-08-14 14:38:53 +08:00
  • f27bf45c2a fix z0o0ey 2026-08-13 20:36:52 +08:00
  • 1a7977b7f3 fix(megatron/grpo): clear MegatronGRPOTrainer._metrics per rollout HSYZhang 2026-08-13 16:56:39 +08:00
  • 1e2a17d769 Fix alpha UMI Next loss scale (#9900) cherry77-cloud 2026-08-13 14:08:03 +08:00
  • b88a5117f4 fix: add set_epoch() to DataLoaderDispatcher for streaming datasets (#9902) diaby24 2026-08-13 14:07:20 +08:00
  • 93525b1774 Synchronize MuonClip max logits across ranks (#9901) cherry77-cloud 2026-08-13 11:35:25 +08:00
  • 5d02ae9722 Fix sp with transformers5.15 (#9899) tastelikefeet 2026-08-12 23:15:53 +08:00
  • e061ed9cf1 wip z0o0ey 2026-08-12 20:18:17 +08:00
  • 7f97926e43 wip z0o0ey 2026-08-12 20:17:39 +08:00
  • d17f031895 fix:MiMo audio input (#9898) z0o0ey 2026-08-12 18:02:43 +08:00
  • 73f081bfc1 Fix dcp_validation bug in megatron distributed saving tastelikefeet 2026-08-12 15:29:20 +08:00
  • d211b633e3 Fix MathORM false positives on parse failures (#9889) Ruihan11 2026-08-12 14:30:19 +08:00
  • ca937fbaf8 [model] support nemotron (#9894) jinghanhu 2026-08-11 23:41:06 +08:00
  • 0321ff2cd8 fix(rlhf): reduce sequence-parallel log-prob memory (#9882) yongyaoduan 2026-08-11 20:43:52 +08:00
  • fa351dcb75 Support muse-glimmer (#9887) tastelikefeet 2026-08-11 16:55:55 +08:00
  • 1dbd1bf64a fix(template): keep MiMo dependencies optional (#9886) Ziyang Guo 2026-08-11 16:08:41 +08:00
  • c1da149fff support Xiaomi-MiMo-V2.5 inference( sglang / vllm ) (#9880) z0o0ey 2026-08-11 10:24:01 +08:00
  • 6aa7e4f65e [Infer] Fix transformers streaming logprobs (#9866) cherry77-cloud 2026-08-10 19:11:58 +08:00
  • 3738096873 feat: add loss scale for thinking prefixes (#9875) 杨鹏晖(Yang Penghui) 2026-08-09 12:06:38 +08:00
  • 4d978f6a8a Support empty content openai message format and standalone tool message (#9861) tastelikefeet 2026-08-08 14:56:49 +08:00
  • c625b8981f update 4.4.3 image (#9869) jinghanhu 2026-08-07 17:23:19 +08:00
  • cad6e5cf4e wip z0o0ey 2026-08-06 23:07:10 +08:00
  • dc40c652fa fix(qwen3.5): add Torch fallback for Qwen3.5 linear attention SP (#9865) meichangsu1 2026-08-06 18:33:38 +08:00
  • 0a8541f8c8 fix(utils): preserve streamed download bytes (#9857) Ziyang Guo 2026-08-06 10:46:46 +08:00
  • 4ffff52927 fix(utils): close daemon event loops on shutdown (#9858) Ziyang Guo 2026-08-06 10:34:30 +08:00
  • 9298fb8a97 Fix missing LoRARequest injection in GRPOVllmEngine.infer_async (#9856) Zuozhuo 2026-08-05 10:42:05 +08:00
  • 552e00aa87 Support ds v4 flash 0731 (#9834) tastelikefeet 2026-08-04 23:11:44 +08:00
  • 15c97e06a8 docs(readme): fix v3.0 release note links (#9853) Ziyang Guo 2026-08-04 20:40:33 +08:00
  • 2390f43bd5 fix(infer): reject prompts without generation space (#9851) Ziyang Guo 2026-08-04 20:39:17 +08:00
  • 2ce97ce4f4 Fix GRPO offload compatibility with FSDP2 CPU offload (#9844) addsubmuldiv 2026-08-04 20:35:02 +08:00
  • d20d208d48 fix: handle BF16 MindSpeed grouped linear activation offload (#9846) addsubmuldiv 2026-08-04 19:09:27 +08:00
  • 3886b1658e fix(infer): propagate streaming errors (#9849) Ziyang Guo 2026-08-04 17:47:01 +08:00
  • f4fd7089d0 Fix GKD Liger teacher routing (#9841) cherry77-cloud 2026-08-04 14:17:41 +08:00
  • d39405fd06 remove qoder-review action (#9842) jinghanhu 2026-08-04 11:35:00 +08:00
  • 49efcbfe59 Fix Gemma 3 mixed-batch token type IDs (#9786) Hyungwook Choi 2026-08-03 23:19:05 +09:00
  • f97ab73a9b fix Qwen3.5 non-packing causal conv routing (#9812) cherry77-cloud 2026-08-03 22:18:43 +08:00
  • 0b9b00c92c Support OpenAI and Anthropic message processors (#9809) Lumin 2026-08-03 22:17:55 +08:00
  • 8d1f7a9443 Fix export merge_lora with quantization not working (#9830) Yongtao Huang 2026-08-03 22:13:41 +08:00
  • 05b5685a78 fix: preserve configured PEFT LoRA dtype (#9832) buduoqiu 2026-08-03 22:12:07 +08:00
  • e390c9ce0d Merge tag 'v4.4.3' into release/4.4 release/4.4 hjh0119 2026-08-03 21:00:44 +08:00
  • e1287928be bump transformers and peft (#9837) v4.4.3 jinghanhu 2026-08-03 20:35:44 +08:00
  • ccd0104800 [Bugfix (rlhf)] skip model reload during colocate vLLM init on FSDP2 (#9793) ys2025-AI 2026-08-03 16:52:27 +08:00
  • 850fe6baf9 docs: fix broken Reward Model example link in README (#9831) Recoordinate 2026-08-01 14:46:26 +12:00
  • a5229fda7e fix(npu): optimize MoE weight synchronization for NPU colocate training with vLLM-Ascend. (#9828) addsubmuldiv 2026-07-31 17:46:17 +08:00
  • 61cc11fdb9 align sft dev-v5 hjh0119 2026-07-31 16:04:38 +08:00
  • ab7a19a040 feat(rollout): support min_p sampling in GRPO/DAPO rollout (#9824) (#9826) Tai An 2026-07-30 19:08:25 -07:00
  • abbfc58674 [bugfix] report per-device maximum memory (#9825) kaining-never-stop 2026-07-30 19:44:13 +08:00
  • 2d0b823aae Merge branch 'main' into dev-v5 hjh0119 2026-07-30 15:14:45 +08:00
  • 69cf0cbfa7 init commit hjh0119 2026-07-30 15:13:08 +08:00
  • 5afcce57a5 [megatron] Fix FA4 backward recompilation: pass max_seqlen as int in PackedSeqParams (#9798) 徐旺 2026-07-30 13:48:54 +08:00
  • 9c8a0135a0 perf(megatron): vectorize channel loss aggregation for padding_free (#9795) gakkii 2026-07-30 13:48:20 +08:00
  • 9ca151490a Add kimi-k3 template (#9811) tastelikefeet 2026-07-30 13:28:00 +08:00
  • 8d9936fed5 fix(sample): don't crash on torch_dtype in engine_kwargs (#9816) Yufeng He 2026-07-30 13:27:44 +08:00
  • ff9baeb535 update wechat (#9819) tastelikefeet 2026-07-30 13:22:42 +08:00
  • 4805d7ff68 fix gkd use_vllm false (#9818) jinghanhu 2026-07-30 11:25:02 +08:00
  • fb60b11117 fix(npu): expose GDN helpers for context parallelism (#9814) addsubmuldiv 2026-07-29 14:18:33 +08:00