[wolfcomos]

Overview

Bound MiniMax-M3 indexer prefill splits/workspace, build reliable prefill metadata on the host, and precompile expected score variants using existing JIT-warmup controls.

Requested feature extraction from xinli-sw/vllm: maxmax at 3fcc2b022 into wolfcomos/vllm:main. Extraction baseline: 292da21f9d887e7cbd511a648971c7627d83c246. This is a fork-local draft for human review; AI assistance was used for extraction, main compatibility fixes, and validation notes.

Claims

• Includes split-prefill and host-metadata prerequisites together with the bounded planner and precompile source commits.
• Includes only this feature's default/control refinements from 5c353c3c, with no index-K sharding or decode-scoring CP.
• No newly measured speedup, GPU correctness, or model-accuracy claim.

Validation

• git diff --check extraction-main...HEAD: passed.
• Ruff 0.14.0 lint/format and Python syntax checks on changed files: passed.
• All applicable pre-commit hooks on the complete changed-file set, including mypy: passed.
• Isolated AST/numpy checks passed the 7 source parameter cases, 3,000 seeded randomized layouts, and 2 page-index reference checks. Full pytest collection of tests/kernels/attention/test_minimax_m3_indexer_prefill_split.py was blocked by missing Torch. No imported-module, SM100 score parity, serving, model evaluation, or performance run completed.

GPU/runtime/model validation remains outstanding; this draft is not qualified for upstream submission.

Details

The bounded planner defaults enabled. Precompile follows kernel_config.enable_jit_warmup, and host metadata is used when its request lengths are reliable. Disabling all split paths requires both attention_config.minimax_m3_indexer_prefill_kv_split=False and VLLM_MINIMAX_M3_INDEXER_PREFILL_KV_SPLIT=0, because the original split fallback remains available.

Includes the same a77c623f prerequisite as the independently based index-K draft; reconcile shared prefill edits when combining them. Source helper files and the source test match the pinned maxmax tip. Precompile currently runs in metadata-builder initialization; moving it into the current warmup registry remains a design review item.

Source commits: a77c623f0, 769102cb4, 37aee4554, 8c4363bbb, 5c353c3c0. Original authors and provenance are retained; mixed-feature follow-up commits are included only where specified above.

Duplicate-work check: no existing open PRs were present in wolfcomos/vllm before this extraction. Related upstream #59979 covers split-KV/sync-free metadata, while #60930 changes the prefill scorer to CuTe. This is the requested source-version extraction on the user's fork.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Human review of source design and listed tradeoffs is still required.

• Testing and Validation: Static/isolated host checks above passed; relevant runtime, GPU, model and performance validation remains outstanding.

• Code Quality and Style: Changed-file hooks pass; extraction adaptations are documented.

• Pull Request Contents: Source links, scope, duplicate-work context, and validation limits are included.