[MRV2] Pass FULL replay context to synchronous KV loads

[whx-sjtu]

Overview

Pass the actual FULL graph mode to MRV2's synchronous KV-load pre-forward hook, so a connector can wait for its loads before graph replay bypasses Python layer hooks. This is the standalone FULL-context fix extracted from #58968.

Rebased onto main bfa7e5b7afe40ebe48288801a2baf5c1135cd881. This remains an independent child of closed #58968.

Claims

• Run kv_connector.pre_forward() under a forward context whose runtime mode is CUDAGraphMode.FULL on the FULL replay branch.
• Preserve eager/piecewise paths and avoid an additional DP batch-coordination operation.

FULL context -> pre_forward / synchronous load wait -> graph replay

Validation

Validated on 61398a91a907fafb364cb0ffd7b2f902844abdb3 (2026-10-12): 17 passed. Changed-file pre-commit passed, including mypy for Python 3.11 and test/CI tethering checks.

VLLM_TARGET_DEVICE=cpu .venv/bin/python -m pytest -q -ra \
  tests/v1/worker/test_gpu_model_runner_v2.py tests/v1/worker/test_gpu_kv_connector.py

Fresh negative control on the current main: the two new FULL-context cases fail for DP sizes 1 and 2; both pass here. The tests check that pre_forward observes FULL mode without triggering another DP synchronization. Rebase preserves #60754's (scheduler_output, input_batch, attn_metadata) interface.

Exact-source FULL-graph PD/RDMA execution remains pending. Eager accuracy and a mocked replay boundary do not establish that result.

The eight independent PRs compose on this same main. The combined CPU suite passes 344 tests, with 33 skips (32 ROCm/aiter-gated connector cases and one GPU LLM case) and one model-loading deselection. Counts overlap with the independent suites.

Details

ActiveKVConnector starts synchronous loads during pre_forward(). Without a caller-provided context it constructs a default context. MoRI's READ path distinguishes FULL replay, which must finish all required reads before replay, from execution that can wait through per-layer hooks. The missing FULL context can therefore select the wrong existing wait path.

This changes only the FULL call site and its tests. It adds no connector capability, does not introduce a new READ barrier and does not depend on the MultiConnector ownership or speculative-readiness split. #57700 explicitly left this context correction to #58968; #48464 concerns MRV1 imperative lifecycle helpers, not this MRV2 replay call site.

Merge impact: reversible code change; MRV2 FULL pre-forward hooks. An existing connector's FULL-specific waiting behavior can now run at this call site. No throughput or latency equivalence is claimed.

#60808 changes MoRIIO to await all READs before any forward; if it lands first, MoRI no longer relies on FULL mode for that decision. This PR separately preserves the common MRV2 replay-context contract. #60637 handles output publication after draft writes, not input loading.

AI assistance was used for implementation, rebase, regression checks and this description. Human line-by-line review and the runtime acceptance items above remain pending. No performance improvement is claimed.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).
• AI assistance was used during the creation of this PR.
• Design Fit: Source/compatibility review completed; human line-by-line review remains pending.
• Testing and Validation: Rebased CPU regressions passed; the runtime acceptance gaps above remain open.
• Code Quality and Style: Changed-file pre-commit, mypy and test/CI tethering checks passed.
• Pull Request Contents: Current source, evidence, related PRs and remaining limitations documented.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @whx-sjtu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @whx-sjtu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork