[WIP] Add MiniMax-M3 on MI355X with EAGLE3 spec decode

Adds workloads/minimax_m3_mi355x.yaml: MiniMax-M3-MXFP8 on MI355X (CDNA4),
TP=8 (no expert parallelism), EAGLE3 speculative decode (drafter
Inferact/MiniMax-M3-EAGLE3, 3 speculative tokens), env
VLLM_USE_BREAKABLE_CUDAGRAPH=0. gsm8k as the spec-decode correctness gate;
vllm_bench 8k-in/1k-out at conc-128 (MI355X-family baseline).

KV cache left at default (bf16): MiniMax-M3-MXFP8 ships no calibrated KV scales,
so --kv-cache-dtype fp8 silently corrupts output (vllm-project/vllm#45562).

nightly: true, but blocked on vllm-project/vllm#45381 (M3 support, which carries
the ROCm/AMD EAGLE3 enablement #45546) landing on main; it will fail at server
bring-up until then.

AI-assisted (Claude Code).

Co-Authored-By: Claude <noreply@anthropic.com>
Signed-off-by: functionstackx <47992694+functionstackx@users.noreply.github.com>

Author: functionstackx

workloads/minimax_m3_mi355x.yaml

@@ -0,0 +1,53 @@
+# MiniMax-M3 on MI355X (CDNA4) with EAGLE3 speculative decode
+#
+# MXFP8 checkpoint runs natively on CDNA4 (OCP fp8 + MX matrix cores).
+# EAGLE3 drafter: Inferact/MiniMax-M3-EAGLE3 (LlamaForCausalLMEagle3 head,
+# hidden_size 6144 / vocab 200064 matching the M3 target), drafting 3 tokens.
+#
+# KV cache is intentionally 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 (see vllm-project/vllm#45562); do not enable it
+# here without a calibrated checkpoint or a self-calibrating dtype.
+name: minimax_m3-mi355x
+gpu: MI355X
+num_gpus: 8
+nightly: true
+
+vllm:
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  env:
+    VLLM_USE_BREAKABLE_CUDAGRAPH: 0
+  serve_args: >-
+    --tensor-parallel-size 8
+    --tool-call-parser minimax_m3
+    --reasoning-parser minimax_m3
+    --enable-auto-tool-choice
+    --trust-remote-code
+    --speculative-config.method eagle3
+    --speculative-config.model Inferact/MiniMax-M3-EAGLE3
+    --speculative-config.num_speculative_tokens 3
+
+lm_eval:
+  # gsm8k here is a correctness gate: speculative decoding is output-equivalent
+  # to non-spec greedy, so the score must match the no-spec MiniMax-M3 baseline.
+  model_args:
+    tokenized_requests: false
+    tokenizer_backend: null
+    timeout: 6000
+  tasks:
+    - name: gsm8k
+      num_fewshot: 5
+      model_args:
+        num_concurrent: 64
+        max_length: 40960
+        max_gen_toks: 32768
+
+vllm_bench:
+  configs:
+    - name: 8k-in-1k-out-conc-128
+      backend: openai
+      dataset: random
+      input_len: 8192
+      output_len: 1024
+      num_prompts: 512
+      max_concurrency: 128