8000 Qualify Rollout Runtime v2 for v0.11.0 by DaoyuanLi2816 · Pull Request #98 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content

Qualify Rollout Runtime v2 for v0.11.0 - #98

Merged
DaoyuanLi2816 merged 6 commits into
mainfrom
v0.11.0-hf-cached-speed-gate
Aug 29, 2026
Merged

DaoyuanLi2816 merged 6 commits into
mainfrom
v0.11.0-hf-cached-speed-gate

Conversation

@DaoyuanLi2816
Copy link
Copy Markdown
Owner

Summary

  • add a compiled BF16 rollout mirror for NF4 actors with exact live-LoRA synchronization and StaticCache compatibility
  • enable the qualified managed vLLM CUDA Graph path for direct GKD while keeping PG-k1 fail-closed on hf_cached
  • publish a new exact-wheel 24-cell RTX 4080 result without changing the preregistration, baseline, first failed candidate, or frozen scientific artifacts

Measured outcome

  • hf_cached / hf_reference, response 256: 2.375-4.305x in every paired cell
  • hf_cached / hf_reference, response 512: 2.541-4.228x in every paired cell
  • vLLM / hf_cached, required cells: 3.076-5.966x
  • peak total GPU memory: 6,504 MiB (hf_cached), 11,931 MiB (vLLM)
  • eight unique policy refreshes and clean teardown passed for both backends
  • local NF4/mirror conformance passed; vLLM PG-k1 log-probability conformance remains failed closed

Closes #97.

Validation

  • ruff check .
  • ruff format --check .
  • mypy src/miniverl
  • actionlint 1.7.12
  • pytest -q -m "not gpu and not network" --cov=miniverl: 2486 passed, 10 skipped, 24 deselected, 82.8% coverage
  • pytest -q -m gpu: 10 passed
  • pytest -q -m network: 16 passed
  • generated-artifact, privacy, text-integrity and Markdown-link checks
  • mkdocs build --strict and 64 Playwright screenshots across four viewports
  • package build and twine check

The aggregate result is 0fb2066a8b567adbb011113141b2b4e3a4fa281cf3f3d13f6eecd1d7b0499a5d. The frozen calculator remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc.

@DaoyuanLi2816
DaoyuanLi2816 merged commit fb9bbd5 into main Aug 29, 2026
16 checks passed
@DaoyuanLi2816
DaoyuanLi2816 deleted the v0.11.0-hf-cached-speed-gate branch August 29, 2026 22:58
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.

Optimize hf_cached to satisfy the frozen Rollout Runtime v2 speed gate

1 participant

0