2626 Commits
Author SHA1 Message Date
d7f1a42109 feat(fsdp): load module-level compute kernels from the Hub (#2829)
Co-authored-by: Zhichen Zeng <zhichen.zeng@radixark.ai>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-30 19:24:47 -07:00
Yikai Zhang 077fb59228 [AMD][LoRA] Add a gpt-oss-20b MoE LoRA RL launcher (#3157) 2026-09-30 21:17:08 -04:00
Jiajun Li 55ec3d870f test(e2e): rework e2e CI coverage on top of the H200 migration (#3699) 2026-09-30 15:54:37 -07:00
Zhichen Zeng 8f409c7c49 fix(docker): patch TileLang's fp16 256-bit store on Blackwell (#3647) 2026-09-30 12:11:13 -07:00
Xinyu JiangandZhiyao Jiang 874b3e2756 [AMD] docker: build the ROCm images on the sglang v0.5.20 release images (#3873)
Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
2026-09-30 00:10:03 -07:00
Jiajun Li 79ef601d6d fix(ray): pin each trainer to the NUMA node of the GPU it actually uses (#3869) 2026-09-29 21:07:58 -07:00
Jiajun Li f602ca207c docs: add Claude rule for launch and request args (#3867) 2026-09-29 20:36:38 -07:00
Zhichen Zeng bb35ecc490 tests: fix two fast tests related with megatron rpc(#3872) 2026-09-29 19:26:08 -07:00
Xinyu JiangandZhiyao Jiang 5ebc06f8d2 [AMD] ci: expand ROCm nightly coverage and refresh runtime estimates (#2541)
Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
2026-09-29 17:28:58 -07:00
HEJIAN SANGandClaude Opus 5.5 3439ec7513 perf(dsv4): balance the CSA indexer across contiguous CP ranks (#3689)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-29 09:05:40 -07:00
Jiajun Li 9e4260de04 Use the base image's Mooncake instead of reinstalling a dev wheel (#3738) 2026-09-27 21:55:14 -07:00
Jiajun LiandClaude Opus 5.5 f84b1496e1 test(e2e): move 8x H100 e2e tests to H200 stages and widen parallelism coverage (#3660)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-27 20:41:43 -07:00
Jiajun LiandShi-Dong 833d5bf78b fix: exec command worker launch commands so actor death frees their ports (#3718)
Co-authored-by: Shi-Dong <shi.dong@radixark.ai>
2026-09-27 20:13:25 -07:00
fzyzcjy 23d41d711f Drop the empty-batch timeout only multi-LoRA v1 ever raised (#3723) 2026-09-26 20:50:59 +08:00
fzyzcjy 447438eefa Keep the live-rollout fault comparison bitwise up to the fault (#3357) 2026-09-26 20:50:30 +08:00
fzyzcjy 9fecf0ff24 Bind the sglang router on the ip a placed hostname resolves to (#3356) 2026-09-26 20:49:59 +08:00
fzyzcjy a85437d470 Move the wandb utils tests under tests/ (#3122) 2026-09-26 20:49:27 +08:00
fzyzcjy 207d892ca6 Build model script test arguments from the launcher environment (#3121) 2026-09-26 20:48:57 +08:00
fzyzcjy a0d2ac6242 Shut the worker manager down on every driver exit path (#3120) 2026-09-26 20:48:28 +08:00
fzyzcjy 5d8dda0b1a Extract the mooncake store config into one table (#3119) 2026-09-26 20:47:58 +08:00
fzyzcjy 47170b94dc Restore the event history before the first process writes events (#3118) 2026-09-26 20:47:28 +08:00
fzyzcjy f3f1717336 Let a CPU-only controller render the inference engine command (#3078) 2026-09-26 20:46:57 +08:00
fzyzcjy f580bef3bd Declare trainer concurrency groups once (#3077) 2026-09-26 20:46:26 +08:00
fzyzcjy fdb973993f Enable the long solver-verifier test on calibrated eval gates (#3076) 2026-09-26 20:45:55 +08:00
fzyzcjy b6c2692beb Pause generation with retract in the solver-verifier recipe (#3075) 2026-09-26 20:45:20 +08:00
fzyzcjy 126cbc0064 Save every sample of a multi-policy run, keyed by policy (#3074) 2026-09-26 20:44:51 +08:00
fzyzcjy fbb1450cab Report the raw reward of every generated group, filter or no filter (#3073) 2026-09-26 20:44:18 +08:00
fzyzcjy 6419b7258a Dispatch shared-engine eval from the multi-policy training loop (#3072) 2026-09-26 20:43:50 +08:00
fzyzcjy 2b42fdb24c Keep realistic hot restart WandB identity stable (#3070) 2026-09-26 20:43:22 +08:00
fzyzcjy 87cbb8ac80 Schedule hot restarts through virtual cells (#3069) 2026-09-26 20:42:51 +08:00
fzyzcjy e771de58ed Keep hot restart tests outside automatic FT recovery (#3068) 2026-09-26 20:42:22 +08:00
fzyzcjy f050555e05 Keep event analysis off the trainer event loop (#3067) 2026-09-26 20:41:51 +08:00
fzyzcjy 2098042990 Gate Ray fault injection between weight updates (#3066) 2026-09-26 20:41:21 +08:00
fzyzcjy 36ce06f8a3 Report multi-policy verifier reward outcomes (#3065) 2026-09-26 20:40:53 +08:00
fzyzcjy b96ab1262f Inject faults at quiescent points and witness recovery one-sidedly (#3064) 2026-09-26 20:40:24 +08:00
fzyzcjy f96336534c Start a run over with an empty log when a take-over finds no checkpoint (#3063) 2026-09-26 20:39:53 +08:00
fzyzcjy 12f7f8c648 Let disposing the rollout release a wedged producer (#3062) 2026-09-26 20:39:24 +08:00
fzyzcjy 8d15269131 Settle the hot restart comparison lifecycle (#3061) 2026-09-26 20:38:55 +08:00
fzyzcjy dfcb8e0597 Keep the first multi-policy failure and stop followers by flag (#3060) 2026-09-26 20:38:24 +08:00
fzyzcjy d1286304c8 Stamp the startup weight sync with the restored rollout id (#3059) 2026-09-26 20:37:56 +08:00
fzyzcjy f2e0a9ab59 Add a debug switch that computes the stored log probabilities under grad (#3058) 2026-09-26 20:37:25 +08:00
fzyzcjy 1860b1363a Reclaim the Ray worker manager and its gpus when a run ends (#3057) 2026-09-26 20:36:56 +08:00
fzyzcjy 3d3fb97bac Give every worker that watches pods the platform read access it needs (#3056) 2026-09-26 20:36:27 +08:00
fzyzcjy 4ce14ed227 Diagnose a failed run across every release it installed (#3055) 2026-09-26 20:35:58 +08:00
fzyzcjy 220bc2bb0d Refuse a multi-node command that a node pin cannot schedule (#3054) 2026-09-26 20:35:29 +08:00
fzyzcjy f4f49f1d17 Let the environment override the model and data directories (#3053) 2026-09-26 20:35:00 +08:00
fzyzcjy fb1d3861c8 Forbid a backend from overriding the exec methods (#3052) 2026-09-26 20:34:31 +08:00
fzyzcjy 4daa1026cc Print every stack when a test hangs (#3051) 2026-09-26 20:34:01 +08:00
fzyzcjy fd68758b2b Give the e2e harness its own command runner (#3050) 2026-09-26 20:33:30 +08:00
fzyzcjy adc558d10e Sort the CI pytest arguments by directory (#3049) 2026-09-26 20:33:01 +08:00