MiniMax-M3-NVFP4 B200 recipe: validation numbers predate the NVFP4 correctness fix (vllm#48929); offer to re-validate + PR post-fix numbers

[stewtong]

The MiniMax-M3-NVFP4 B200 recipe's "validated benchmark" note (#605, merged 2026-07-09) predates the NVFP4 inference-correctness fix vllm-project/vllm#48929 (merged 2026-08-05). No stable release contains that fix (v0.26.0 was cut 2026-07-27), so the page's validation numbers were measured on a build carrying the pre-fix swiglu/indexer behavior.

We re-ran the recipe on a post-fix nightly (0.26.1rc1.dev457+gc810e5ee9) on 8x B200 with correctness gates in front of every measurement. Full report with methodology, raw-data conventions, and startup logs: vllm-project/vllm#51494. Highlights on the post-fix build:

Measurement	Result
Correctness (temp 0, content-checked 7-prompt suite)	7/7 pass; needle recovery 144/144 (envelope) + 15/15 (ladder, 128K-1M real prose)
1M real-prose TTFT, c=1	43.2s p50 (1,011,492 input tokens, 3 reps)
Aggregate prefill throughput	~54K input tok/s at 128K c=16
EAGLE3 (Inferact/MiniMax-M3-EAGLE3, 3 spec tokens)	2.1-2.3x decode-window gain at +5% TTFT (c=1)

Two proposals for the recipe page:

1. Re-validate on a post-#48929 build and state a minimum build date. We are happy to PR the post-fix numbers above directly, including the reproduction commands.
2. Add a note on the nightly window 2026-08-05 to 2026-08-07, where images fail at startup (TYPE_UNSTABLE_JOIN in the SM100 FMHA path, quack-kernels 0.6.3 under cutlass-dsl 4.6; fixed upstream by [Bugfix] Pin quack-kernels before FA4 AST regression vllm#51418). Anyone following the recipe on a nightly from that window needs quack-kernels<0.6.3.

Optional third item, since the data supports it: given the 2.1-2.3x decode gain at +5% TTFT and the EAGLE3 config's higher KV capacity (2,727,296 vs 2,558,976 tokens, from --language-model-only freeing the vision tower), the recipe could present EAGLE3 as the recommended pairing rather than an option. Details and acceptance counters in vllm-project/vllm#51494.

If maintainers agree on direction, tell us which shape you want and we'll open the PR.