[KV Connector][NIXL] Support MRV2 prefill PCP replicas

[LucasWilkinson]

Note: This PR is stacked on vllm-project#49294 and should be retargeted to main after it merges.

Purpose

Support NIXL P/D disaggregation when an MRV2 prefill producer uses Prefill Context Parallelism (PCP) and the decode consumer remains PCP1/DCP1.

MRV2 PCP gathers prefill K/V inputs before cache insertion, so PCP replicas at the same TP/PP position own the same complete KV shard. NIXL handshake metadata is keyed by PP/TP only; publishing every PCP replica therefore overwrites the same identity and makes completion aggregation wait for or process duplicate producer events.

This change:

• selects PCP rank 0 as the canonical producer KV replica;
• excludes non-canonical PCP replicas from producer handshake publication;
• aggregates producer completion over canonical TP x PP workers;
• suppresses duplicate finished_sending events from non-canonical replicas while still allowing their local cleanup; and
• extends the NIXL accuracy launcher with PREFILLER_PCP_SIZE.

The validated topology is:

Prefill producer: TP1 + PCP2 + EP2 + PP1
Decode consumer:  TP1 + PCP1/DCP1
Connector:        NixlPullConnector

Related work and duplicate check

Live searches for NIXL PCP, prefill context parallel NIXL, replica selection, and completion handling found no open PR implementing this NIXL ownership fix.

• [KV Connector][NIXL][Mooncake] CP-scaled scheduler block accounting for PD disaggregation with aligned DCP/PCP vllm-project/vllm#45340 fixes CP-scaled scheduler block accounting for aligned P/D CP and explicitly leaves the transfer path unchanged. This PR instead selects one canonical replica for replicated prefill PCP with a PCP1 decoder.
• feat(nixl,dcp): Supports MLA DCP for PD disaggregation with nixl connector vllm-project/vllm#38433 implements general sharded DCP topology, including unequal P/D DCP sizes and sequence-overlap mapping. PCP here is replicated, so publishing every PCP rank would create duplicate sources rather than distinct shards.
• [Core][KV Connector][Mooncake] Support MRV2 prefill PCP in Mooncake vllm-project/vllm#49109 applies the same canonical-owner principle to Mooncake bootstrap and lifecycle code. This PR applies it to NIXL handshake publication and worker completion aggregation.

Test plan and results

Focused PCP/NIXL regression coverage:

.venv/bin/python -m pytest tests/v1/kv_connector/unit/test_nixl_connector.py -k 'pcp_producer or pcp_consumer or pcp_non_publishing' -q

Result: 5 passed, 68 deselected.

Full NIXL connector unit file:

.venv/bin/python -m pytest tests/v1/kv_connector/unit/test_nixl_connector.py -q

Result: pytest reported 73 passed in 176.91 seconds. The process subsequently segfaulted during UCX shared-library teardown in rtnl_tc_unregister, after the completed pytest result.

DeepSeek-V2-Lite-Chat NIXL P/D GSM8K:

MODEL_NAMES=deepseek-ai/DeepSeek-V2-Lite-Chat PREFILLER_PCP_SIZE=2 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=1 GPU_MEMORY_UTILIZATION=0.3 PREFILL_BLOCK_SIZE=128 DECODE_BLOCK_SIZE=128 VLLM_SERVE_EXTRA_ARGS=--max-model-len,4096,--max-num-seqs,128,--max-num-batched-tokens,32768,--trust-remote-code,--enable-expert-parallel,--moe-backend,flashinfer_cutlass bash tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh

Result: all 1,319 questions completed, with strict-match accuracy 0.6535253980; expected 0.65 ± 0.05. The pytest evaluation passed in 112.14 seconds. The shell wrapper received a cleanup signal while terminating the prefill/decode servers after the passing result.

Pre-commit passed on all four changed files, including ruff, mypy, shellcheck, sign-off, and repository policy hooks. git diff --check passed. A static GPU-sync audit found no new host-device synchronization calls.

AI assistance disclosure

OpenAI Codex assisted with investigation, implementation, testing, evaluation, and drafting this PR.

• The human submitter has reviewed every changed line, understands the change end-to-end, and can defend it during review.

[LucasWilkinson]

Superseded by reopened upstream draft vllm-project#49342, now rebased directly onto current main.