[Bugfix] Generalize PLE graph and runtime compatibility - #466

Purpose

Generalize PLE CPU-offload acceleration beyond the safe C1-only CUDA Graph workaround.

• Bind each MRV2 request to its own D2H completion event.
• Stage MRV2 inputs on the model stream so CUDA Graph batch-shape transitions cannot retarget a shared background event.
• Size the MRV2 transport queue from the framework concurrent-batch contract.
• Preserve the existing MRV1 CPU-mirror path.
• Do not bind the fix to a model name, checkpoint, TP size, KV dtype, max-num-seqs, or chunked-prefill setting.
• Generalize SM70 graph coverage, kernel warmup, PLE NUMA residency, NVFP4 routed-row admission, and grouped QSA Page4 dispatch.

This follows upstream vLLM PLE PR #53899 / commit 4e8b849b8d971f65a264bd61846e846b1134c00d, adapted to the dual MRV1/MRV2 support in 1Cat.

Integration base: ca73a34cd7f3f95bd1f016ed7b8a03c499c10cf4
Validated head: f71300a2a5e4ba45a3aaded9187d688fbbe922dd

Duplication check

This is the integration PR for the PLE/graph/runtime compatibility scope. PR #469 is a complementary pre-Ampere Triton fallback launch profile; PR #471 is an optional disk-backed PLE capacity mode. Neither duplicates the concurrent input-staging and graph-shape fixes here.

Test Plan and Result

• Focused CPU tests:
  • pytest -q tests/v1/worker/test_gpu_warmup_blocks.py tests/v1/worker/test_ple_offload_worker.py
  • 42 passed
• Targeted pre-commit, mypy 3.12, and forbidden-import checks passed.
• TP4 V100 CUDA Graph transition stress:
  • C4 → C3 → C2 → C1 → C2 → C3 → C4 → C1, repeated twice.
  • 16/16 phases completed; all request token SHA256 values matched their single-request references.
  • Warm aggregate decode reached 315.92–321.88 tok/s at C4 and 99.41–101.93 tok/s at C1.
• Production-shaped startup smoke:
  • official FlashNext NVFP4 checkpoint, TP4, MTP4, 262144 max length,
    chunked prefill 4096, max-num-seqs 4, prefix cache, Mamba align,
    Flash-V100, target and MTP CUDA Graph.
• Additional C4 → C3 → C2 → C1 API smoke after the CI repair: all 10 requests returned HTTP 200.
• Grouped QSA Page4 numerical microbenchmark:
  • FP16 30.105 → 4.308 ms.
  • E4M3 34.850 → 5.435 ms.
  • Relative L2 error remained within approximately 2.8e-4.
• Task-owned service was shut down after validation and GPU memory was released.

Raw artifacts:

• /data/minimax-h3/task-cache/v100-flashnext-pr466-local-20260903/logs/transition-stress-old-profile.txt
• /data/minimax-h3/task-cache/v100-flashnext-pr466-local-20260903/logs/launcher.log

AI assistance

This change was AI-assisted by OpenAI Codex. The human repository owner requested and reviewed the integration scope, authorized the merge, and the commits include DCO sign-off.