[AMD] feat: MiniMax M3 day-zero benchmark for MI300X

[cquil11]

Summary

• add minimaxm3-fp8-mi300x-vllm using the dedicated vllm/vllm-openai-rocm:minimax-m3 image already used by MI355X
• add a MI300X fixed-sequence launcher with the MI355X M3 serving flags: block size 128, prefix caching disabled, TRITON_ATTN, FP8 KV cache, language-model-only, and eager execution
• use the exact H100 non-MTP TP8/TEP8 search space across 1k1k and 8k1k
• fix the MI300X Slurm exclusion to use the current short node name (chi-mi300x-049)
• explicitly mount /dev/kfd and /dev/dri into the rootless Pyxis/Enroot container
• disable prefix caching to match the H100 MiniMax M3 benchmark behavior

Validation

• bash -n benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi300x.sh runners/launch_mi300x-amds.sh
• generated MI300X sweep matches H100 exactly: 14 TP/EP/concurrency configurations
• python -m pytest utils/matrix_logic/ -q (156 passed)
• dedicated ROCm image imported successfully to the shared runner squash cache (~29 GB)
• MiniMaxAI/MiniMax-M3-MXFP8 staged successfully in node-local /raid/hf-hub-cache (~414 GB)

Hardware Validation

Single-job GHA smoke: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27480947060

The TP8, EP1, 1k1k, concurrency-4 smoke completed successfully on chi-mi300x-054, including result processing and artifact upload:

• total throughput: 11 tok/s/GPU
• output throughput: 5 tok/s/GPU
• mean TTFT: 287 ms
• mean TPOT: 92 ms

The model loaded on all 8 MI300X GPUs, the vLLM API became healthy, all requests completed successfully, and the benchmark artifact was uploaded. The earlier Slurm, cache-directory, and rootless GPU-device permission blockers are resolved.

The slow throughput is expected for this image on MI300X: vLLM reports no native MXFP8 MoE backend on gfx942, so the MoE runs through the BF16 emulation path with eager execution.

Sweep

full-sweep-fail-fast runs the complete 14-point H100-aligned concurrency sweep. Keep this PR in draft while the sweep is running.

---

Note

Low Risk
Benchmark matrix, launcher script, and Slurm/container mount tweaks only; no changes to auth, data handling, or core application logic.

Overview
Adds a day-zero minimaxm3-fp8-mi300x-vllm matrix entry for MiniMax-M3 MXFP8 on MI300X, wired to the same vllm/vllm-openai-rocm:minimax-m3 image as MI355X and an H100-aligned fixed-seq sweep (TP8 and TP8+EP8 on 1k1k and 8k1k).

Introduces minimaxm3_fp8_mi300x.sh to serve with the MI355X-validated flags (block size 128, no prefix caching, TRITON_ATTN, FP8 KV, eager, MiniMax M3 parsers) and run the standard serving benchmark path.

MI300X runner fixes: Slurm --exclude now uses the short node name chi-mi300x-049, and Pyxis/Enroot mounts /dev/kfd and /dev/dri for rootless GPU access. Documents the recipe in perf-changelog.yaml.

^{Reviewed by Cursor Bugbot for commit d43b005. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node 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/sglang/tree/main/docs_new

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.

[claude]

Claude finished @cquil11's task in 1m 30s —— View job

---

Reviewing PR #1746

• Read repository guidelines
• Read all changed files
• Validate master config changes
• Validate benchmark script
• Validate launch script changes
• Validate perf-changelog entry
• Post review findings

Review

LGTM — no blocking issues found.

Config (amd-master.yaml): The minimaxm3-fp8-mi300x-vllm entry is appended at the end, uses a publicly accessible Docker Hub image (vllm/vllm-openai-rocm:minimax-m3), and the search space is byte-identical to the H100 non-MTP config in nvidia-master.yaml (14 sweep points confirmed).

Benchmark script (minimaxm3_fp8_mi300x.sh): Identical to the validated MI355X script. Server args are properly formatted on separate lines. Expert parallelism is correctly conditional on EP_SIZE (not hardcoded). Not an MTP script, so --use-chat-template is not required.

Launch script (launch_mi300x-amds.sh): The /dev/kfd and /dev/dri mounts are standard for rootless ROCm containers. The short hostname fix (chi-mi300x-049 vs FQDN) aligns with Slurm's default hostname resolution. The enroot import docker:// pattern is present for reproducibility.

Perf changelog: New entry is correctly appended at the end of the file.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27482347617
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27482347617