fix(integrations): restore strict Qwen3 bitwise parity

[inaniloquentee]

Summary

• add the strict FA4 BSHD entrypoint plus bundled-vLLM FA4/rotary compatibility
• align Megatron and vLLM Attention projections on DetGemmOp and match Megatron's RMSNorm arithmetic
• replace every duplicate sampled-token logprob column before vLLM response conversion

Validation

• python -m pytest -q tests/test_framework_runtime_adapters.py tests/test_flashinfer_pr7_attention.py: 77 passed, 1 skipped
• 8x H100, Qwen3-8B Dense, TP=2, CP=2, 128 response tokens, debug environment variables unset:
  • train_rollout_logprob_abs_diff: 0.0
  • mismatch_count: 0 / 128
  • max_abs_diff: 0.0

The validation job subsequently hits Vime/Megatron's existing distributed optimizer checkpoint assertion (torch.Size([]), 0, 311427072), after the strict consistency metric has passed.

Fixes #335

[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: 97a29115-1913-4926-8dac-0185368841b8

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.}

[Flink-ddd]

LGTM, Thank you for update.