[Bugfix] Guard DeepSeek V4 MRV1 piecewise CUDA graphs

[WoosukKwon]

Summary

Default DeepseekV4ForCausalLM to Model Runner V2 and reject only the known-broken configuration:

• DeepSeek V4
• Model Runner V1
• PIECEWISE or FULL_AND_PIECEWISE CUDA graphs

The validation runs after CUDA-graph mode resolution. MRV1 therefore remains available with eager/NONE, FULL, or FULL_DECODE_ONLY; other model architectures are unaffected.

Why

#51430 exposed a correctness problem in the legacy V1 model runner's breakable PIECEWISE CUDA-graph path. The same attention implementation retains normal GSM8K accuracy and MTP acceptance with Model Runner V2.

Failing this specific combination avoids silent output corruption while preserving MRV1 for configurations that do not exercise the affected path.

The error directs users to either select MRV2 or disable PIECEWISE CUDA graphs.

Relationship to #51750

This is an alternative to #51750, not a duplicate. #51750 exactly reverts #51430 and restores the wider MRV1 eager region. This PR leaves the attention implementation unchanged, defaults DeepSeek V4 to the already-correct MRV2 path, and fails closed if a user explicitly reaches the broken MRV1 PIECEWISE path.

The required duplicate searches found no other open PR adding this configuration guard.

Model validation

GPU validation ran through Slurm on a GB200 node with TP=2, FP4 indexer cache, MTP with two speculative tokens, and CUDA graphs enabled.

code	runner	questions	GSM8K	MTP draft acceptance
exact post-#51430 commit 635dd6aae	MRV2	256	0.9570	80.7–81.4%
unmodified current main	MRV2	1,319	0.9500	81.1–82.1%
exact post-#51430 commit 635dd6aae	MRV1 + PIECEWISE	1,319	0.0311	collapsed to 3.1–5.5%

Tests

env -u VLLM_USE_V2_MODEL_RUNNER \
  .venv/bin/python -m pytest tests/test_config.py \
  -k 'v2_model_runner or mrv1_piecewise' -q
# 29 passed

.venv/bin/pre-commit run ruff-check --files \
  vllm/config/vllm.py tests/test_config.py
.venv/bin/pre-commit run ruff-format --files \
  vllm/config/vllm.py tests/test_config.py
.venv/bin/pre-commit run mypy-3.12 --files \
  vllm/config/vllm.py tests/test_config.py --hook-stage manual

The tests cover both rejected PIECEWISE modes and the allowed MRV2, eager/NONE, FULL, FULL_DECODE_ONLY, and non-DeepSeek-V4 cases.

AI assistance

AI assistance was used to investigate the runner-specific regression, implement this change, run validation, and prepare this PR.

Before marking ready for review, the human submitter must:

• review and understand every changed line
• independently confirm the relevant test results

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[WoosukKwon]

/ci run

[github-actions]

✅ Triggered Buildkite CI #83314 for commit 28cfd2e196d8.