[Feature]: Add DBO support for Ascend NPU ModelRunnerV2

[jiangkuaixue123]

Feature, motivation, and pitch

Ascend NPU ModelRunnerV2 support landed in #257, but the current path explicitly rejects enable_dbo=true and use_ubatching=true. This leaves DBO workloads on ModelRunnerV1 and prevents the Ascend ModelRunnerV2 path from reaching the DBO milestone described in Phase 3 of #240.

Add native DBO support to AFDNPUAttentionModelRunnerV2 for the synchronous CAMP2pAFDConnector path. The implementation should preserve native vLLM/vLLM-Ascend ownership of ModelRunnerV2 request state, input preparation, per-uBatch Attention metadata, output merging, NPU streams/events, and ACL graph lifecycle.

Related work:

• [RFC]: Add Ascend NPU ModelRunnerV2 support for AFD #240 — Ascend NPU ModelRunnerV2 RFC, Phase 3
• add MRV2 for NPU without DBO #257 — Ascend ModelRunnerV2 implementation without DBO

Proposed change

• Enable enable_dbo / native ubatching for the Ascend ModelRunnerV2 Attention path.
• Prefer an AscendUBatchRunner owned by vLLM-Ascend, exposed through a narrow factory or extension hook.
• Integrate AFD transaction, stage, token-shape, DP, and connector-control metadata per uBatch without copying the ModelRunnerV2 execution loop.
• Initially support exactly two uBatches with synchronous CAMP2pAFDConnector.
• Deliver eager DBO first; qualify FULL_DECODE_ONLY DBO only after eager correctness is established.
• Keep ModelRunnerV1 DBO and ModelRunnerV2 no-DBO behavior unchanged.
• Fail closed for unsupported connectors, graph modes, topologies, or uBatch counts.

vLLM compatibility and extension points

Preferred extension point:

• A vLLM-Ascend-owned AscendUBatchRunner and an overridable uBatch-runner construction hook in the native NPU ModelRunnerV2 path.
• Narrow AFD composition/hooks for connector lifecycle, control publication, AFD forward-context metadata, profiling, and cleanup.

Compat shim needed:

• A version-pinned shim may be used if the pinned vLLM-Ascend runtime does not yet expose the required factory or lifecycle hooks.

Monkey patch needed:

• Avoid if possible. If unavoidable, keep it exact-version, minimal, and documented with an upstream/removal plan.
• Do not copy native execute_model, input preparation, Attention metadata construction, output merging, or ACL graph-manager lifecycle into afd-plugin.

The implementation targets the repository's pinned vLLM/vLLM-Ascend runtime pair and must not modify either upstream source tree from this repository.

Alternatives considered

• Keep DBO on ModelRunnerV1 only: lowest implementation risk, but leaves Ascend MRV2 feature parity incomplete.
• Port the existing NPU V1 DBO loop into ModelRunnerV2: rejected because it duplicates upstream execution ownership and creates a large upgrade surface.
• Copy the GPU UBatchRunner into afd-plugin: rejected because CUDA-specific stream/graph assumptions do not define the Ascend runtime contract.
• Implement eager and ACL graph DBO together: increases the debugging surface; eager-first provides a smaller correctness gate.

Validation plan

Focused tests:

• Configuration accepts the supported two-uBatch MRV2 DBO setup and rejects unsupported combinations before connector rendezvous or weight loading.
• Per-uBatch AFD metadata preserves real and padded token lengths, stage identity, transaction identity, and DP rank agreement.
• Connector control is published before each uBatch data transfer and outside ACL graph bodies.
• Failure paths clean up sibling execution, streams/events, forward contexts, graph state, and connector transactions.

Ascend hardware E2E:

• DeepSeek-V2-Lite, synchronous CAMP2pAFDConnector, 1A1F, DP1/TP1, eager DBO.
• Decode, prefill, mixed batches, uneven splits, padding-only tails, repeated requests, profiling, and clean shutdown.
• Deterministic output comparison against approved native/no-DBO or ModelRunnerV1 references.
• DP2/TP1 and TP2/DP1 coverage before broad topology claims.
• Full non-accuracy NPU E2E gate, then GSM8K-7 accuracy.
• After eager acceptance, FULL_DECODE_ONLY capture/replay, graph-key separation by uBatch count/shape, graph misses, and shape changes.

Acceptance criteria:

• Exactly two uBatches execute correctly through native Ascend ModelRunnerV2 DBO ownership.
• No copy of the native MRV2 execution loop or Ascend Attention/graph internals is introduced.
• AFD connector payloads and control metadata are correct for every uBatch.
• Eager correctness, cleanup, and accuracy gates pass on Ascend hardware.
• Graph DBO is not claimed until its separate capture/replay gate passes.
• Existing Ascend MRV2 no-DBO and ModelRunnerV1 DBO tests remain green.

Before submitting

• I searched existing issues and RFCs.
• I identified whether this belongs in plugin-owned code, compat helpers, or compat patches.