[None][perf] Autotune MiniMax-M3 MXFP8 decode backends

[zheyuf]

@coderabbitai summary

Description

Factorize MiniMax-M3's MXFP8 generation-graph backend tuning:

Stage	Before this PR	This PR
Activation quantization	Fixed TRT-LLM native kernel; no backend tuning	TRT-LLM native vs. FlashInfer CuTeDSL per graph shape bucket
GEMM	FlashInfer CUTLASS; only its internal tactics were tuned during eager warmup	FlashInfer CUTLASS vs. FlashInfer CuTeDSL per graph shape bucket

The first generation-graph warmup profiles the two candidates for each stage; the second captures the winners. The search is factorized (2 + 2), not a four-way cross product, and CUDA-graph replay has no profiling or Python dispatch.

Activation is intentionally narrow: MiniMax-M3, PP1, whole-model generation CUDA graphs, no torch.compile, and no explicit backend override. Other models, PP>1, graph-disabled execution, torch.compile, explicit backend choices, and missing CuTeDSL retain their existing path and autotuning lifecycle. This adds no public option, dependency, or M <= 32 heuristic.

Performance

Same-node A-H-H-A on Alex's 12 B300 AGG Pareto points; all 48 arms passed correctness and dispatch validation, and both head arms improved both InferenceX axes at every point.

Point	Output tok/s/GPU (X)	Output tok/s/user (Y)
TEP8 C1	+7.02%	+7.70%
TP4 C1	+7.90%	+7.67%
TP4 C2	+10.26%	+11.74%
TP4 C4	+9.06%	+10.03%
TP4 C8	+8.33%	+9.38%
TP4 C16	+7.44%	+8.64%
TP4 C32	+4.66%	+6.03%
DEP8 C128	+4.10%	+4.94%
DEP8 C256	+1.99%	+3.32%
DEP8 C512	+1.74%	+2.60%
DEP8 C1024	+0.68%	+2.83%
DEP4 C1024	+0.12%	+0.83%

X = (sum(input tokens) + sum(actual output tokens)) / benchmark duration / deployment GPUs.

Y = 1000 / mean(TPOT_ms), where each request's TPOT = (E2EL - TTFT) / (actual output tokens - 1).

Every run used 8,192 input and 1,024 output tokens per request. Each entry is the mean of the two wall-clock-interpolated H-vs-A gains. The measured serving path is unchanged by the final lifecycle narrowing in this commit.

Autotuning overhead

Measurement	Added time
First model instance, warm same-image/JIT cache	+4-26 seconds
Second model instance	Usually +0-3 seconds; worst about +5 seconds
Serving after RPC readiness	0 seconds

With an empty CuTeDSL JIT cache, the factorized graph-tuning window itself is 93-96 seconds, including one-time compilation. That is an absolute cold-start window, not a +93-96 second incremental cost: on MiniMax-M3 it replaces the old eager FlashInfer tuning phase. All tuning finishes before RPC readiness, so it is outside AgentX benchmark duration and request TTFT; cold-start/autoscaling readiness is the tradeoff.

Test Coverage

• 42 focused unit tests passed on the latest Alex image.
• Full pre-commit suite and git diff --check passed.
• B300 SM103 tuned and replayed real CUDA graphs at M=1, 24, 40, and 64; eager and replay outputs matched the TRT-LLM native reference with cosine 1.0.
• A current-image TP4 C1 server completed factorized tuning, captured winners, and reached readiness; the second model instance reused the caches without retuning.

PR Checklist

• The description explains the change and its performance impact.
• New paths have tests and follow the coding guidelines.
• No public API, dependency, documentation, or CODEOWNERS change is required.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.