[WS2][PR5][Attention] Add distributed CP drift benchmark artifacts

[inaniloquentee]

Summary

• Add a rank-aware WS2 CP attention drift benchmark for Qwen3-8B TP=1/2 and CP=1/2.
• Record CP merge drift, BF16-vs-FP32 drift, RoPE provenance, per-logical-CP-rank metrics, and optional PR8 backward drift.
• Reuse Transformer Engine only as an optional merge oracle via the context-parallel correction helpers; RL-Kernel remains the reference merge contract.

Testing

Windows / local CI checks:

• pre-commit run --all-files
• py -3.13 -m pytest tests/test_ws2_cp_attention_drift_benchmark.py tests/test_cp_attention.py tests/test_operator_inputs.py -q (45 passed)
• py -3.13 -m pytest tests/test_attention.py tests/test_kv_cache_attention.py tests/test_attention_correctness.py -q (85 passed, 83 skipped)
• py -3.13 -m mypy --ignore-missing-imports rl_engine/
• py -3.13 -m mkdocs build --strict -f mkdocs.yaml
• py -3.13 scripts/check_operator.py --op cp_attention --check-grad
• py -3.13 benchmarks/benchmark_ws2_cp_attention_drift.py --smoke --json
• py -3.13 benchmarks/benchmark_ws2_cp_attention_drift.py --smoke --tp-world-sizes 2 --cp-world-sizes 2 --kv-chunk-sizes none,1 --include-backward --device cpu --output $env:TEMP\ws2-cp-attention-drift-pr5.json

WSL validation:

• python3 -m pytest tests/test_ws2_cp_attention_drift_benchmark.py tests/test_cp_attention.py tests/test_operator_inputs.py -q (45 passed)
• python3 -m pytest tests/test_attention.py tests/test_kv_cache_attention.py tests/test_attention_correctness.py -q (39 passed, 129 skipped)
• python3 benchmarks/benchmark_ws2_cp_attention_drift.py --smoke --tp-world-sizes 2 --cp-world-sizes 2 --kv-chunk-sizes none,1 --include-backward --device cpu --json
• python3 -m torch.distributed.run --nproc-per-node=2 benchmarks/benchmark_ws2_cp_attention_drift.py --smoke --tp-world-sizes 2 --cp-world-sizes 2 --kv-chunk-sizes none --device cpu --json

Notes:

• WSL tests used CPU execution. The WSL PyTorch install warned that the local NVIDIA driver is older than the CUDA wheel; no CUDA path is claimed by this PR.
• Windows torchrun was blocked by the host PyTorch libuv rendezvous build before script launch; WSL torch.distributed.run passed for the torchrun-style rank-aware path.

Notes

• Stacked on PR feat(attention): add CP backward drift validation #284 / codex/ws2-pr8-cp-attention-backward.
• Related issue: [WS2] CP-aware deterministic Attention for cross-config alignment (Qwen3-8B TP=2 CP=2 BF16) #235.

[coderabbitai]

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: 78c077c6-8010-4c2f-9b90-39bbc449c024

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

• X
• Mastodon
• Reddit
• LinkedIn

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