[WS2][Cross-config][Attention] Integrate the Qwen3-8B TP=2 CP=2 BF16 Attention setting to cross-config alignment framework

[zhangj1an]

Summary

Implements issue #235 PR4 for the Qwen3-8B BF16 TP=2/CP=2 Megatron + vLLM attention cross-configuration target.

This PR now depends on the current heads of:

• [WS2] Add CP-aware Attention contract and dispatch metadata  #236: first-class attention, sharding, reduction, RoPE, KV-cache, and Split-KV contracts
• feat(attention): add deterministic CP reference #238: deterministic CP reference and complete Split-KV runtime-plan provenance
• feat(alignment): add cross-configuration framework #230: cross-configuration planning/runtime framework

What is enforced

Area	Strict rule
Semantic identity	Checkpoint/model/weight/tokenizer/token/mask/position/padding/pre-update identity must match
TP/CP topology	TP/CP ranks, world sizes, Qwen3 GQA head ownership, sequence ownership, and global block layout must match; topology is not a recorded free difference
Reduction	Online-softmax LSE merge, FP32 accumulation, global block order, final-write downcast, exported attention-domain LSE
Split-KV request	AttentionContract.split_kv is first-class and shared by training and rollout
Split-KV execution	Both runtimes must provide the complete batch x TP x CP x KV-owner SplitKVRuntimePlanSet
Split-KV fail-closed cases	Missing plans, topology/KV-length/request/boundary/order/dtype/downcast mismatch, unknown plan, or fallback all fail
Runtime evidence	Constructed/configured contracts are UNOBSERVABLE; only post-execution runtime readback can be APPLIED
CP fallback	Requested CP=2 becoming effective CP=1 cannot pass strict TP=2/CP=2 acceptance

bind_attention_runtime_readbacks(...) is the strict handoff for real Megatron/vLLM launchers after they reconstruct the actual contract, actual knob values, frozen-scope verification, and all-rank Split-KV plans.

The scenario now uses TP=2/CP=2 on both sides. The vLLM flash_attn_max_num_splits_for_cuda_graph value remains diagnostic only; it is not treated as proof of the executed logical Split-KV chunk boundaries.

Scope boundary

This PR defines and validates the binding/readback boundary. It does not launch Megatron or vLLM, initialize process groups, or claim GPU execution. A real engine integration must inject AttentionRuntimeReadback after execution. Decode logical KV replay remains in #260.

Validation

Current head: 48a4130

WSL Ubuntu, Python 3.11, PyTorch 2.4.1+cpu:

• full rl_engine/ MyPy passed (123 source files)
• updated adapters/tests: Black, isort, and flake8 passed
• attention binding/runtime/contract/CP focused tests: 164 passed, 1 skipped
• cross-config runner tests: 8 passed
• CI-equivalent Attention tests: 24 passed, 2 deselected
• CI-equivalent KV-cache tests: 15 passed; Attention contract tests: 42 passed
• unsupported reduction requests invalidate every dependent knob application instead of allowing a missing contract to reach runtime readback
• the CUDA Attention package exports only modules present in this PR; PR7-only cp_comm and FlashInfer exports are not falsely exposed

GitHub linting, unit-tests, and docs checks all pass on this head. GPU/Megatron/vLLM execution is intentionally not claimed by these CPU contract tests.

Diff audit

The PR currently shows 63 changed files against test. The original PR4 head already changed 62 of them. The only added path is rl_engine/kernels/gtest/operator_specs.py from the PR1 dependency, which registers the cp_attention test operator. The broader cross-config framework files are pre-existing PR4 scope, not newly added by the latest review fixes.

[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: ef21cddb-6596-42f9-850f-248ffcc709db

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.}