Skip to content

feat(integrations): add H100 Megatron/vLLM runtime adapters - #338

Merged
Flink-ddd merged 1 commit into
testfrom
codex/h100-framework-runtime-adapters
Aug 25, 2026
Merged

feat(integrations): add H100 Megatron/vLLM runtime adapters#338
Flink-ddd merged 1 commit into
testfrom
codex/h100-framework-runtime-adapters

Conversation

@inaniloquentee

Copy link
Copy Markdown
Collaborator

Summary

  • install one shared R/R operator plan in Megatron and vLLM workers without editing either framework
  • route structural hidden-state and LM-head inputs through the separate strict linear_logp boundary
  • bind the merged Attention/FFN semantic operators and persist CUDA-only runtime provenance/readbacks
  • register the opt-in vLLM general plugin and add an exact train/rollout artifact validator
  • add focused tests for runtime hooks, strict readback policy, and validation artifacts

Context

This completes the RL-Kernel-side runtime bridge exercised by #335 and the companion RL-Align/vime#42.

Depends on #336 for the separate strict SM90 sm90_deterministic_linear_logp entry points. The high-performance linear_logp route remains unchanged.

This PR targets RL-Align/RL-Kernel:test. It does not modify or target vllm-project/vime.

Validation

  • adapter/integration suite: 97 passed
  • with feat(linear-logp): add strict bitwise SM90 contract #336 temporarily applied: 59 passed
  • wider Attention/CP/FlashInfer suite: 224 passed, 83 skipped
  • Ruff, Flake8, Black, isort, compileall, and git diff --check: passed
  • package egg_info contains vllm.general_plugins -> rl_engine.integrations.vllm_runtime:register_vllm_plugin

Two existing H100 preprocessing tests remain environment-blocked because the host Transformer Engine extension has a PyTorch ABI mismatch (undefined c10::Error symbol); the adapter-related tests pass.

@coderabbitai

coderabbitai Bot commented Aug 25, 2026

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 81fc1a8d-0e00-48ba-bc37-ca62247def0f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@Flink-ddd Flink-ddd left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@maxiaosong1124
maxiaosong1124 self-requested a review August 25, 2026 03:39
@Flink-ddd
Flink-ddd merged commit edde365 into test Aug 25, 2026
4 of 6 checks passed
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.

3 participants