Skip to content

[Runtime][TP2 CP2] Publish strict bitwise runtime and user modes - #344

Open
inaniloquentee wants to merge 2 commits into
testfrom
codex/qwen3-strict-runtime
Open

[Runtime][TP2 CP2] Publish strict bitwise runtime and user modes#344
inaniloquentee wants to merge 2 commits into
testfrom
codex/qwen3-strict-runtime

Conversation

@inaniloquentee

Copy link
Copy Markdown
Collaborator

Summary

This publishes the complete RL-Kernel side of the Qwen3-8B TP=2, CP=2 strict
train/rollout consistency work on top of the current test branch.

  • restores the strict Megatron and vLLM Attention, FFN, collective, and
    linear_logp runtime routes used by the 8xH100 validation;
  • keeps every strict Attention, GEMM, FFN, and logprob route fail-closed;
  • defaults the aligned Hopper runtime to deterministic cuBLASLt no-split-K
    GEMM while retaining the explicit SM90 backend;
  • adds RL_KERNEL_MODE=strict|audit|auto|off as the user-facing integration
    contract;
  • reports interface, operator, actual, and fallback once per module on
    rank 0 by default;
  • exposes the canonical provider as
    rl_engine.integrations.vime.linear_logp.provider, while keeping the old
    provider path compatible;
  • keeps all framework changes in RL-Kernel-owned adapters. No Vime, Megatron,
    or vLLM native source is modified.

The branch was rebuilt from origin/test before publication, so the earlier
local WIP snapshot is not present in the PR history.

8xH100 evidence

Qwen3-8B, TP=2, CP=2, strict aligned R/R:

mismatch_count=0
max_abs_diff=0
torch_equal=true
steady rollout=1.0173s
steady actor=0.8536s
steady step=22.8104s

Artifact directory on the validation host:

/home/ellm/ljj/vime-debug-main/outputs/rlkernel/fair-rr-aligned-cublaslt-20260826T152921Z

Tests

66 passed, 44 skipped

The skipped cases require CUDA. bash -n and git diff --check also pass.

@coderabbitai

coderabbitai Bot commented Aug 26, 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: be66767f-cbf4-4e8f-9d66-294057cdc1b4

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

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

@inaniloquentee inaniloquentee changed the title [Vime][TP2 CP2] Publish strict bitwise runtime and user modes [Runtime][TP2 CP2] Publish strict bitwise runtime and user modes Aug 26, 2026
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