[Klaud Cold] Add minimaxm3-fp4-mi355x-atom

[indianspeedster]

Summary

Adds the minimaxm3-fp4-mi355x-atom config — MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) on MI355X, single-node atom engine — for the 1k/1k and 8k/1k fixed-seq-len cells, TP4.

Follows the ROCm/ATOM MiniMax-M3 recipe (FP4 on 4×MI355 section).

• .github/configs/amd-master.yaml: new config entry + search space (TP4, conc 1→128, image rocm/atom-dev:M3).
• benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh: atom serve script — --block-size 128 (mandatory for MiniMax MSA), --gpu-memory-utilization 0.8, --trust-remote-code. KV cache left at the default dtype: this MXFP4 checkpoint ships no calibrated FP8 KV scales, so --kv_cache_dtype fp8 asserts (k_scale is None) in the MSA fused_qknorm kernel during init.
• runners/launch_mi355x-amds.sh: route amd/MiniMax-M3* weights to the NFS cache (alongside the existing MiniMaxAI/MiniMax-M3* rule).
• perf-changelog entry.

Validation

• generate_sweep_configs.py test-config → 16 configs: minimaxm3_1k1k and minimaxm3_8k1k, each TP4 at conc {1,2,4,8,16,32,64,128}; max-model-len = 2304 (1k1k) / 9472 (8k1k); framework atom.
• Smoke-tested on real MI355X hardware (TP4 / conc-1 / 1k1k): atom server came up across 4 ranks, served, and the benchmark wrote a well-formed result JSON.

🤖 Generated with Claude Code

---

Note

Low Risk
Benchmark and launch-routing changes only; no production serving or auth paths touched.

Overview
Adds day-zero fixed-seq-len benchmarking for MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) on MI355X using the ATOM engine (minimaxm3-fp4-mi355x-atom).

The new amd-master.yaml entry uses image rocm/atom-dev:M3, TP4, concurrency 1→128, and 1k/1k and 8k/1k cells per the ROCm/ATOM recipe. A new serve script starts atom.entrypoints.openai_server with --block-size 128 (MSA requirement), 0.8 GPU memory utilization, and default KV cache dtype (no FP8 KV — the MXFP4 checkpoint has no calibrated scales). launch_mi355x-amds.sh now NFS-mounts weights for amd/MiniMax-M3* as well as MiniMaxAI/MiniMax-M3*. perf-changelog documents the new config key.

^{Reviewed by Cursor Bugbot for commit a68c303. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9a2b0f4. Configure here.}

[functionstackx]

@indianspeedster thanks for the contribution, can u or one of ur teammates add create an upstream branch & add full-sweep-enabled PR validation

[andyluo7]

@functionstackx done — mirrored this to an upstream branch so the GPU sweep can run (fork PRs can't access the self-hosted runners): #1813 (feat/minimaxm3-fp4-mi355x-atom on the upstream repo, same commits, credits @indianspeedster). Added the full-sweep-enabled label and kicked off full-sweep PR validation there — it's running now across the full matrix (1k1k + 8k1k, TP4, conc 1→128) on MI355X: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27718177974

(Also: the earlier Cursor Bugbot MAX_MODEL_LEN comment is already resolved by commit a68c303, which uses the matrix $MAX_MODEL_LEN.)

[functionstackx]

thanks @andyluo7

superceded by #1813