Default Branch

d7f1a42109 · feat(fsdp): load module-level compute kernels from the Hub (#2829) · Updated 2026-10-01 10:24:47 +08:00

Branches

a846a0968a · build: bump CUDA 13 images to SGLang v0.5.21 · Updated 2026-10-01 17:22:43 +08:00

0
1

b9db6d6777 · Disable the overlap scheduler in the shared deterministic rollout recipe · Updated 2026-10-01 15:43:37 +08:00

359
770

ad84212202 · Preserve backend arguments and checkpoint origin in HF conversion · Updated 2026-10-01 15:43:36 +08:00

359
754

acb267b22f · Use private mutable arguments in the multi-agent example · Updated 2026-10-01 15:43:36 +08:00

359
755

ffd9bb40de · Copy evaluation engine GPU counts under the controller lock · Updated 2026-10-01 15:43:36 +08:00

359
756

5f31717312 · Read random-async settings from their typed owners · Updated 2026-10-01 15:43:36 +08:00

359
757

57c6f1071d · Keep plain HF weights distinct from resumed training iterations · Updated 2026-10-01 15:43:36 +08:00

359
758

fd3a11fb49 · Compress runtime snapshot bases with reversible stage deltas · Updated 2026-10-01 15:43:36 +08:00

359
759

9c29266d52 · Normalize generated values using registered launch provenance · Updated 2026-10-01 15:43:36 +08:00

359
760

eb9e76737b · Record runtime baselines from validated E2E captures · Updated 2026-10-01 15:43:36 +08:00

359
762

6501f53c71 · Migrate unit tests and preserve Verifiers contract coverage · Updated 2026-10-01 15:43:36 +08:00

359
763

c5422087b4 · Enforce explicit dynamic attribute access in examples · Updated 2026-10-01 15:43:36 +08:00

359
764

de6766031e · Add e2e scenarios that scale the engine pool and the trainer's DP cells up and down under a live run · Updated 2026-10-01 15:43:35 +08:00

359
747

a17c8b7e7b · Resize a pool as a soak action · Updated 2026-10-01 15:43:35 +08:00

359
745

31a3479884 · Assert a hot restart's changed argument reached only the restarted components · Updated 2026-10-01 15:43:35 +08:00

359
741

03c84e3a66 · Share the kubectl runner, the cell counts of a mode and the soak forms type · Updated 2026-10-01 15:43:35 +08:00

359
742

8d3e6263a5 · Share starting and following launchers across deploy soak sessions · Updated 2026-10-01 15:43:35 +08:00

359
743

c87b257c83 · Observe a pool of a live run as a soak target · Updated 2026-10-01 15:43:35 +08:00

359
744

bffd93a3ca · Add the runner the scaling scenarios share · Updated 2026-10-01 15:43:35 +08:00

359
746

e2a4d09dc7 · Make LoRA Hugging Face target ordering deterministic · Updated 2026-10-01 15:43:35 +08:00

359
751