Tune MiniMax MI355X vLLM scheduling thresholds

[jiacao-amd]

Summary

Tune the MiniMax-M2.5 FP8 MI355X vLLM launch policy for better throughput across the 1k/1k and 8k/1k sweep points.

• Default path: block-size=32, shuffled KV cache disabled, async scheduling enabled.
• 1k/1k: use block-size=16 with shuffled KV cache; disable async scheduling through c128.
• 1k/1k TP8/EP8 c2: keep block-size=32 and shuffled KV cache disabled, but disable async scheduling.
• TP8/EP8 fallback path: keep block-size=32 and shuffled KV cache disabled.
• 8k/1k: disable async scheduling through c64; keep c32 on block-size=32 with shuffled KV cache disabled, and enable block-size=16 with shuffled KV cache from c64 upward.

Throughput Comparison

Metric: tput_per_gpu only.

• Validation run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/25346292897
• Baseline run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24006177731
• Compared benchmark points: 33/33
• Improved: 32/33
• Regressed: 1/33
• Average delta: +6.5%
• Median delta: +6.6%

ISL/OSL	TP/EP	Conc	Baseline	Validation	Delta
1k/1k	2/2	2	175.6	197.4	+12.4%
1k/1k	2/2	4	316.6	352.1	+11.2%
1k/1k	2/2	8	551.1	591.0	+7.2%
1k/1k	2/2	16	912.4	985.4	+8.0%
1k/1k	2/2	32	1512.4	1621.4	+7.2%
1k/1k	2/2	64	2283.1	2268.4	-0.6%
1k/1k	2/2	128	3745.8	3845.6	+2.7%
1k/1k	2/2	256	5459.7	5781.6	+5.9%
1k/1k	2/2	512	8080.0	8362.1	+3.5%
1k/1k	4/4	4	173.0	195.2	+12.8%
1k/1k	4/4	8	329.8	355.5	+7.8%
1k/1k	4/4	16	554.2	597.7	+7.9%
1k/1k	4/4	32	976.6	1037.3	+6.2%
1k/1k	4/4	64	1574.5	1679.0	+6.6%
1k/1k	4/4	128	2620.1	2707.7	+3.3%
1k/1k	4/4	256	3846.2	3945.2	+2.6%
1k/1k	8/8	2	47.7	53.1	+11.4%
8k/1k	2/2	2	712.6	831.1	+16.6%
8k/1k	2/2	4	1320.6	1357.8	+2.8%
8k/1k	2/2	8	2162.1	2187.6	+1.2%
8k/1k	2/2	16	3378.7	3530.0	+4.5%
8k/1k	2/2	32	4645.2	5091.0	+9.6%
8k/1k	2/2	64	6495.4	6760.2	+4.1%
8k/1k	2/2	128	8601.5	8900.6	+3.5%
8k/1k	2/2	256	10391.0	10689.6	+2.9%
8k/1k	4/4	4	730.2	807.2	+10.6%
8k/1k	4/4	8	1291.4	1394.4	+8.0%
8k/1k	4/4	16	2076.3	2250.0	+8.4%
8k/1k	4/4	32	3314.3	3533.6	+6.6%
8k/1k	4/4	64	4741.3	5074.6	+7.0%
8k/1k	4/4	128	6719.6	6882.4	+2.4%
8k/1k	4/4	256	8156.8	8409.5	+3.1%
8k/1k	4/4	512	9136.0	9873.0	+8.1%

Testing

• bash -n benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh
• git diff --check
• Compared results_bmk artifacts from the validation and baseline runs above.

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sgl-cookbook

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[jiacao-amd]

/sweep test-config --config-files .github/configs/amd-master.yaml --config-keys minimaxm2.5-fp8-mi355x-vllm

[github-actions]

@jiacao-amd Kicking off a sweep.

Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/25346292897
Command: test-config --config-files .github/configs/amd-master.yaml --config-keys minimaxm2.5-fp8-mi355x-vllm
Pinned ref: a9a3cef
Approval: not required (trusted collaborator).