Skip to content

Refresh OSS export from refactor_2604: fused CUDA kernels, vanilla we…#1

Open
StarrickLiu wants to merge 1 commit into
mainfrom
oss-export-2604-refresh
Open

Refresh OSS export from refactor_2604: fused CUDA kernels, vanilla we…#1
StarrickLiu wants to merge 1 commit into
mainfrom
oss-export-2604-refresh

Conversation

@StarrickLiu

Copy link
Copy Markdown
Owner

…bsocket serving, DMuon LIBERO configs

@StarrickLiu
StarrickLiu force-pushed the oss-export-2604-refresh branch from 10490a1 to 262cb8e Compare July 22, 2026 09:45
…rving, restructured-harrix vendoring, DMuon LIBERO configs

- Bake fused CUDA kernels (two_expert MoE / rmsnorm / silu / rope) into
  csrc/ with install-time CUDAExtension build (internal-parity MFU ~53%).
- Vendor the restructured harrix (e2e_infer / model_executors / serving.replica
  + handlers/monitoring) for the LIBERO eval + vanilla websocket serving; the
  vanilla policy path is validated bit-identical to the eval predict_batch.
- run_serving.sh exposes the full vanilla CLI and rejects full-stack-only flags
  with guidance; returns raw action chunks.
- DMuon LIBERO/lerobot configs use recommended LRs (muon_lr=1e-4,
  adamw_lr=5e-5) + ZeRO-2; add libero_dmuon.yml.
- Preserve published hotfixes (X-Square-Robot#110 arXiv links, X-Square-Robot#112 normalizer delta==0 guard).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@StarrickLiu
StarrickLiu force-pushed the oss-export-2604-refresh branch from 262cb8e to d8ce5ff Compare July 23, 2026 07:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant