[WS2][PR7][Attention] Add FlashInfer RoPE-fused paged attention scaffold

[inaniloquentee]

Scope

PR7 is the CUDA paged-KV adapter and strict production Attention entry point for #235. FlashInfer remains the cache-layout adapter; strict Attention arithmetic now reuses the qualified FA4 CuTe core instead of the RL-Kernel reference core.

Boundary	Production path	Fixed strict control
Paged KV	FlashInfer layout adapter	Logical K/V and position order materialized before arithmetic
Attention core	flash_attn.cute.interface.flash_attn_func	Same FA4 core on training and rollout
Split-K	FA4 num_splits=1	Contract-level Split-KV disabled; no auto split
Backward	FA4 deterministic=true	Dropout 0; deterministic backward required
GQA / LSE	pack_gqa fixed from head layout; return_lse=true	FP32 attention-domain LSE with exact shape checks
Batch execution	One logical batch row per core call	Stable batch-composition behavior
RoPE	RL-Kernel SM90 deterministic RoPE	Same post-RoPE boundary on both sides
CP communication	Self-owned cuda_ag_rs	Q/K/V/positions AG, FA4 core, Out/LSE RS
Reference	RL-Kernel deterministic Attention core	Reference/debug oracle only; rejected as production provenance

The loader inspects the installed FA4 CuTe signature and fails closed when num_splits, deterministic backward, GQA packing, or LSE export controls are absent. There is no silent fallback to another Attention kernel.

Q/K RMSNorm and projection policy are supplied by #230: TE Q/K RMSNorm is reused after an exact probe; QKV/O projection stays on RL-Kernel deterministic GEMM because expanded vendor GEMM probes drifted.

Validation

Current head: 044f5b1.

Gate	Result
Attention regression without optional Triton file	172 passed, 741 skipped
Focused PR7 + correctness suite	58 passed, 129 skipped
FA4 API and provenance tests	Fixed num_splits=1, deterministic backward, FP32 LSE, production/reference identity checks
H100 current-head FA4 matrix	Pending: the H100 host is currently unreachable

Previous H100 results for the RL-Kernel deterministic core remain reference evidence only. They are not presented as FA4 current-head acceptance. Production bitwise status requires rerunning prefill/decode, B=1/B>1, representative sequence lengths, and 2/4/8-rank CUDA AG+RS on this exact head.

Reproduce

python -m pytest -q   tests/test_attention_correctness.py   tests/test_flashinfer_pr7_attention.py

python scripts/ws2_pr7_flashinfer_attention_check.py   --strict --device cuda --no-dry-run   --mode decode --batch-size 2 --query-len 1   --kv-seq-len 128 --page-size 16   --output results/cuda-attention/strict-decode-b2-kv128.json

torchrun --standalone --nproc-per-node=4   scripts/ws2_p2p_nccl_attention_reference_check.py   --transport cuda_ag_rs --strict-shared-core   --output results/cuda-attention/cuda-ag-rs-4r.json

Related to #235 and #294.

[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: 96419c29-bf9c-4862-9261-7e5cd75dc8a7

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