Commit Graph
258 Commits
Author SHA1 Message Date
Yueming Yuan 2799fe3863 Add --colocate-memory-peak-device: overlap the trainer/rollout handoff on the GPU (#2781) 2026-08-31 20:17:22 -07:00
Yueming Yuan 6d0cd93f10 DeepSeek-V4 support-v2: native megatron implementation as option (#2706) 2026-08-31 13:21:05 -07:00
Yueming YuanandJiajun Li 2c78e2abd5 feat: resolve --hardware from the node in every launcher (#2738)
Co-authored-by: Jiajun Li <guapisolo@gmail.com>
2026-08-25 13:14:03 -07:00
580040b725 Bump Megatron-LM to miles-main-20260819 (latest NVIDIA dev) (#2673)
Co-authored-by: Jiajun Li <guapisolo@gmail.com>
Co-authored-by: Zhichenzzz <zczeng@uw.edu>
2026-08-24 12:37:44 -07:00
Yueming Yuan 851ddba2a7 add DeepSeek-V4-Flash-0731 support and mxfp4->fp8 converter (#2717) 2026-08-23 23:42:36 -07:00
Yuzhen ZhouandJiajun Li cc92260df1 fix: align MLA RoPE types with model configs (#2716)
Co-authored-by: Jiajun Li <guapisolo@gmail.com>
2026-08-23 04:14:30 -07:00
Zhiyao Jiang 02098f1e59 [AMD] Point the ROCm image at the v0.5.16 wheels release with TE 2.17.0 (#2660) 2026-08-19 19:52:21 -07:00
Zhichen Zeng bb6f7dbd31 feat: add the Qwen3.8-27B recipe (#2575) 2026-08-16 14:50:54 -07:00
afd088cb6b Make the class-based rollout the default and convert legacy path to env var gated (#2522)
Co-authored-by: Jiajun Li <jiajun.li@radixark.ai>
Co-authored-by: guapisolo <guapisolo@gmail.com>
2026-08-13 22:05:39 -07:00
943b8acd52 [AMD] Enable amd pr ci (#2347)
Co-authored-by: Xinyu Jiang <xinyuj2@andrew.cmu.edu>
Co-authored-by: Yuankai Chen <206619631+yuankaichen-amd@users.noreply.github.com>
Co-authored-by: Xinyu Kang <93342727+Xinyu-Kang@users.noreply.github.com>
Co-authored-by: Shekhar <38083203+indianspeedster@users.noreply.github.com>
Co-authored-by: Sree Rohith Pulipaka <119916571+sreerohi@users.noreply.github.com>
Co-authored-by: lizamd <161388580+lizamd@users.noreply.github.com>
Co-authored-by: Jiajun Li <guapisolo@gmail.com>
2026-08-13 20:14:11 -07:00
Tao LinandClaude Fable 5 946ade07be docs: label-only sidebar groups, explicit Overview pages, scoped link color (#2489)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-13 14:20:23 -07:00
Tao LinandClaude Fable 5 b4f45e2e5d docs: mirror examples/ READMEs into the Examples tab (#2481)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-12 15:26:43 -07:00
Yueming YuanandZhichen Zeng 2b2ef6f8b9 Replace all the .sh launch scripts with .py launch script (#2356)
Co-authored-by: Zhichen Zeng <zhichen.zeng@radixark.ai>
2026-08-11 21:51:40 -07:00
Yueming Yuan 6a6843f1db [fix] fix the bugs/outdated commands in .sh scripts and the corresponding snapshots (#2355) 2026-08-11 19:12:18 -07:00
Yueming Yuan 221dbc7eaa Delete the glm4-9B, mimo-7B, moonlight-16B and deepseek-r1 launch scripts (#2354) 2026-08-11 18:22:26 -07:00
Shi-Dong dc3bf8f58b scripts: download the DAPO dataset the NPU recipe trains on (#2402) 2026-08-11 17:00:25 -07:00
Jiajun LiandYuzhen Zhou 26c82b599c fix(kimi): align YaRN parameters with checkpoints (#2377)
Co-authored-by: Yuzhen Zhou <zyzshishui@gmail.com>
2026-08-11 14:03:51 -07:00
Shi-Dong 355cac452d scripts: enable the Miles dashboard in the quick-start launcher (#2300) 2026-08-10 11:27:34 -07:00
fzyzcjy fe30593dd6 Replace the model config shell scripts with python (#1910) 2026-08-09 18:51:21 +08:00
fzyzcjy 063af76f32 Remove non-reproducible file arguments by supporting inline base64 payloads (#1905) 2026-08-09 18:46:33 +08:00
fzyzcjy e06329bdad Move the shell exec helpers next to their only consumers (#1904) 2026-08-09 18:45:45 +08:00
fzyzcjy 49f9aef18a Rename exec_command by the resource its command needs (#1903) 2026-08-09 18:43:51 +08:00
fzyzcjy 1e51cb838d Derive the miles checkout location instead of hardcoding it in launch scripts (#1898) 2026-08-09 18:24:08 +08:00
fzyzcjy 8da9c60e17 Fix various launch scripts errors about missing line concatenations or paths (#1897) 2026-08-09 18:22:52 +08:00
fzyzcjy 5ed3fd170b Fix typo environment variable and unbuffer python outputs (#1895) 2026-08-09 18:19:56 +08:00
Yueming Yuan 52ac174d97 Remove --disable-weights-backuper; default eligible colocate launchers to rematerialize (#2223) 2026-08-07 12:24:04 -07:00
Zhichen Zengandyueming-yuan 98eec31762 Nemotron RL support (#1284)
Co-authored-by: yueming-yuan <yym022502@gmail.com>
2026-08-04 22:49:18 -07:00
Zhiyao Jiang 3507ce1543 [AMD] retune AMD 4-node dsv4 config (#2139) 2026-08-04 11:04:58 -07:00
Yueming Yuan 38ab6d46ee GLM-5.2 kernel fix and GB300 training config (#1571) 2026-08-03 17:57:52 -07:00
Zhichen Zeng 92ccb87de8 [tml] native LoRA support for inkling model (#2122) 2026-08-03 14:15:51 -07:00
Zhichen Zeng 5c517599fb [tml] Inkling model support (#1683) 2026-08-03 13:59:01 -07:00
Shi-Dong 7401900847 scripts, examples: stop forcing the deprecated Miles router in launchers (#2015) 2026-08-03 12:57:29 -07:00
Jiajun Li 66c46bca9b fix glm47-flash: use paged MLA prefill on B200 (#2047) 2026-08-02 19:38:11 -07:00
Xinyu JiangandZhiyao Jiang d9e3c3b394 [AMD] DeepSeek-V4: use the ROCm precision-parity norm path (#2040)
Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
2026-07-31 15:09:50 -07:00
6bbc317fc1 [AMD] Drop inert DSv4 rollout knobs and add an MTP recipe (#1733)
Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-07-29 13:14:58 -07:00
Shi-DongandZhichen Zeng 0916d04f2f refactor(examples): group examples into infra_features/ and experimental/ (#1953)
Co-authored-by: Zhichen Zeng <zhichen.zeng@radixark.ai>
2026-07-29 12:10:19 -07:00
Zhichen ZengandJiajun Li 5b809d9d74 [fsdp] fsdp backend support fix (#1469)
Co-authored-by: Jiajun Li <guapisolo@gmail.com>
2026-07-27 17:02:43 -07:00
Jiajun Li 978bddcada refactor(tito): remove --tito-allowed-append-roles and derive append surfaces from fixed templates (#1818) 2026-07-27 13:45:04 -07:00
1fca8f8b6f NVFP4 RL (#1261)
Signed-off-by: Ziang Li <ziangli@umich.edu>
Co-authored-by: Yueming Yuan <yueming@Yuemings-MacBook-Air.local>
Co-authored-by: yueming-yuan <yym022502@gmail.com>
2026-07-24 15:43:28 -07:00
Zhihao Wang 8046e3e3a1 [feat] Deepseek V4 Mxfp8 (#1340)
Signed-off-by: zhihaow6 <zhihaow6@illinois.edu>
2026-07-22 00:20:52 -07:00
Zhihao Wang 7e7926b131 fix: decide NVTE_FP8_BLOCK_SCALING_FP32_SCALES by hardware instead of hardcoding it (#1584) 2026-07-21 13:01:57 -07:00
Yueming Yuan eb17279732 Bump sglang to v0.5.15 (#1655) 2026-07-16 11:55:52 -07:00
Zhiyao JiangandXinyu Jiang a60c23ce89 [AMD] Add AMD MI350X/MI355X (gfx950) blockwise FP8 support for run_qwen3_30b_a3b (#1465)
Co-authored-by: Xinyu Jiang <xinyuj2@andrew.cmu.edu>
2026-07-15 16:52:14 -07:00
Ethan (Yusheng) Su d4adb93983 [WIP] feat: lora GLM-5.2 FP8 rollout support (#1678)
Signed-off-by: Yusheng Su <yushengsu.thu@gmail.com>
2026-07-15 16:22:36 -07:00
Ethan (Yusheng) Su bf536263c4 [Lora] Add GDN (Qwen3.5/Qwen3-Next) LoRA target-module mapping (#1597)
Signed-off-by: Yusheng Su <yushengsu.thu@gmail.com>
2026-07-15 14:15:35 -07:00
Jiajun Li ced0b9e7b6 fix(ci): align model-script jobs with GPU allocation (#1671) 2026-07-14 17:35:50 -07:00
Xinyu Kang d4f5a80189 [AMD][DeepSeek V4][2/N] Enable decode CUDA graphs on MI355X (#1664) 2026-07-13 22:58:31 -07:00
Xinyu JiangandZhiyao Jiang 352b4aafcb [AMD] Enable DeepSeek-V4-Flash FP8 RL training on MI355X (#1607)
Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
2026-07-09 19:49:03 -07:00
5bdd942329 Add GLM-5/5.1/5.2 (744B MoE) LoRA RL support (#1559)
Signed-off-by: Yusheng Su <yushengsu.thu@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: yueming-yuan <yym022502@gmail.com>
2026-07-07 15:46:00 -07:00
Yueming Yuan 00a698ea7d Bump sglang to v0.5.14 (#1587) 2026-07-06 23:40:52 -07:00