Add MiniMax-M3 Nightly Perf & Accuracy Regression Testing on MI355X with EAGLE3 spec decode

[functionstackx]

Summary

Adds workloads/minimax_m3_mi355x.yaml — a new recipe running MiniMax-M3-MXFP8 on MI355X (CDNA4) with EAGLE3 speculative decode.

+viz @hongxiayang @andyluo7 @chunfangamd @micah-wil

Serving config

• Model MiniMaxAI/MiniMax-M3-MXFP8 (MXFP8 runs natively on CDNA4 — OCP fp8 + MX matrix cores), num_gpus: 8.
• --tensor-parallel-size 8, no expert parallelism (EP's all-to-all overhead hurts the low-latency decode path that speculation targets).
• No --max-model-len — uses the model default, consistent with the other MI355X recipes.
• minimax_m3 tool-call + reasoning parsers with --enable-auto-tool-choice, --trust-remote-code.
• env: VLLM_USE_BREAKABLE_CUDAGRAPH=0 per @hongxiayang MiniMax-M3: set VLLM_USE_BREAKABLE_CUDAGRAPH=0 on AMD MI300X/MI325X/MI355X recipes#539

EAGLE3 speculative decode

• --speculative-config.method eagle3, drafter --speculative-config.model Inferact/MiniMax-M3-EAGLE3, --speculative-config.num_speculative_tokens 3.

Evals

• lm_eval gsm8k (5-shot) as the spec-decode correctness gate: speculative decoding is output-equivalent to non-spec greedy, so the score must match the no-spec MiniMax-M3 baseline.
• vllm_bench — a single 8k-in/1k-out config at conc-128 (MI355X-family baseline) for the decode-throughput uplift, using random + backend: openai (not speed_bench per Switch vllm_bench from speed_bench to random + ignore_eos #20).

⛔ Blocked — draft / do not merge yet

nightly: true (runs in the nightly schedule), but this is blocked on vllm-project/vllm#45381 (MiniMax-M3 support) landing on main, which carries the ROCm/AMD EAGLE3 enablement (#45546). Until that merges, the vllm-openai-rocm image cannot serve M3 + AMD EAGLE3, so the nightly run will fail at server bring-up until #45381 is on main (red nights are expected). Kept as a GitHub draft until it goes green.

Green checklist

• [Model] Add MiniMax M3 support vllm#45381 merged to main (includes [Bug Fix] [MiniMax-M3] Implement EAGLE3 support on the AMD MiniMax M3 vllm#45546 ROCm EAGLE3)
• wait for @micah-wil 's PR to merge [ROCm] Don't use VLLM_IMAGE on AMD unless it is a ROCm image #23
• Mark PR ready for review

Notes

• KV cache deliberately left at the model default (bf16). MiniMax-M3-MXFP8 ships no calibrated KV scales, so --kv-cache-dtype fp8 silently falls back to scale 1.0 and corrupts output ([Bug]: ROCm MI300X FP8 KV cache MiniMax-M3-MXFP8 accuracy issues vllm#45562). Don't enable fp8 KV here without a calibrated checkpoint or a self-calibrating dtype.
• Perf bench uses random + openai, not speed_bench. speed_bench requires --skip-tokenizer-init, which caps every request at 1 output token so throughput reads ~0; this recipe follows Switch vllm_bench from speed_bench to random + ignore_eos #20 (and drops the now-vestigial speed_bench_* fields the other MI355X recipes still carry).
• No README/schema change — the pipeline auto-discovers workloads/*.yaml.

Validation

• Parser smoke test (per CLAUDE.md, stubbed lm_eval registry) passes: name/image/model/serve_args/env and the lm_eval + vllm_bench TSVs all resolve correctly. Device=mi355x, TP=8, precision=fp8.
• No real GPU run yet (blocked, see above).

AI assistance

This PR was authored with assistance from Claude Code (AI-assisted). The commit carries a Co-Authored-By: Claude trailer.

[functionstackx]

chatted with kevin and got buidlkite perms to test this PR on buildkite before merging it