[Core][PCP] Support data parallelism with MRV2 PCP

[LucasWilkinson]

Summary

• allow MRV2 prefill context parallelism (PCP) with data parallelism (DP)
• allocate each local DP engine's TP/PP/PCP device range correctly
• rebuild DP token metadata after PCP partitions the scheduled batch
• compose AgRs MoE dispatch across DP then PCP, reversing the order for combine
• replicate sequence-parallel token sizes across PCP ranks in flattened EP order
• mask empty MLA context chunks with -inf LSE before merging

Related to #46358.

Why this is not a duplicate

I checked issue #46358 and searched open PRs by issue number and DP/PCP/MLA
keywords. No open PR enables generic MRV2 DP+PCP composition. #49109 is scoped to
Mooncake KV-connector PCP replica selection, while #43809 is model-specific
DeepSeek-V4 hybrid-KV/MegaMoE PCP work. Neither changes the generic DP supervisor,
post-PCP DP metadata, or hierarchical AgRs path implemented here.

Validation

.venv/bin/python -m pytest \
  tests/test_config.py \
  tests/entrypoints/openai/test_dp_supervisor.py -q

Result: 178 passed.

Commit-time pre-commit hooks passed, including Ruff, formatting, mypy, SPDX,
configuration validation, and forbidden-import checks.

Runtime configuration (4 GPUs):

CUDA_VISIBLE_DEVICES=0,1,2,3 \
VLLM_USE_V2_MODEL_RUNNER=1 \
.venv/bin/vllm serve nvidia/GLM-5.2-NVFP4 \
  --enforce-eager \
  --max-model-len 4096 \
  --max-num-batched-tokens 32768 \
  --safetensors-load-strategy prefetch \
  --moe-backend flashinfer_cutlass \
  --data-parallel-size 2 \
  --prefill-context-parallel-size 2 \
  --enable-expert-parallel \
  --kv-cache-dtype fp8 \
  --trust-remote-code

Both DP ranks returned the expected answer (18) when addressed with
X-data-parallel-rank.

100-sample GSM8K smoke evaluation:

.venv/bin/python tests/evals/gsm8k/gsm8k_eval.py \
  --num-questions 100 \
  --num-shots 5 \
  --max-tokens 256 \
  --max-concurrency 100 \
  --port 8000

Deployment	Accuracy	Invalid responses
DP2+PCP2+EP4	94/100	0/100

The same rank-pinned runtime probe passed with VLLM_GPU_SYNC_CHECK=error, with
no synchronization detected during model execution.

AI assistance

AI assistance was used to diagnose and implement this draft. This PR should
remain a draft until the human submitter reviews every changed line, confirms
the test/evaluation results, and can explain and defend the change end-to-end.

[yiminghub2024]

this is very userful pr for prefill , but i want to know , if we can pcp not include in worldsize ? fg :
CUDA_VISIBLE_DEVICES=0,1,2,3
VLLM_USE_V2_MODEL_RUNNER=1
.venv/bin/vllm serve nvidia/GLM-5.2-NVFP4
--enforce-eager
--max-model-len 4096
--max-num-batched-tokens 32768
--safetensors-load-strategy prefetch
--moe-backend flashinfer_cutlass
--data-parallel-size 2
--prefill-context-parallel-size 2
--enable-expert-parallel
--kv-cache-dtype fp8
--trust-remote-code

world size not 4 is 2 , only compute tp/pp/dp or only use pcp if setting pcp ,setting tp/pp/dp to 1 ,so worldsize always 8 with pcp 8 @LucasWilkinson

refer:
#47846

[mergify]

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

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