[ROCm] Fix MiniMax-M3 EAGLE draft backend recipe

[tuukkjs]

Fix MiniMax-M3 ROCm EAGLE draft attention backend recipe.

Update the MiniMax-M3 recipe so AMD ROCm EAGLE3 speculative decoding uses TRITON_ATTN for the draft model attention backend. Keep the default CUDA/NVIDIA EAGLE3 draft backend as FLASH_ATTN.

AI assistance was used for this PR.

Motivation

The MI355X MiniMax-M3 MXFP8 recipe already selects TRITON_ATTN for the target model on ROCm, but the EAGLE3 speculative_config still selected FLASH_ATTN for the draft model:

vllm serve MiniMaxAI/MiniMax-M3-MXFP8 \
  --block-size 128 \
  --attention-backend TRITON_ATTN \
  --tensor-parallel-size 8 \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice \
  --reasoning-parser minimax_m3 \
  --speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3", "num_speculative_tokens": 3, "attention_backend": "FLASH_ATTN"}'

On ROCm, that draft backend selection failed during EAGLE/drafter startup:

AssertionError: FlashAttention version not detected.

Using TRITON_ATTN for the EAGLE draft backend allowed the same ROCm setup to start successfully:

vllm serve MiniMaxAI/MiniMax-M3-MXFP8 \
  --block-size 128 \
  --attention-backend TRITON_ATTN \
  --tensor-parallel-size 8 \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice \
  --reasoning-parser minimax_m3 \
  --speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3", "num_speculative_tokens": 3, "attention_backend": "TRITON_ATTN"}'

Validation

• Runtime startup probes on MI355X with vllm/vllm-openai-rocm:nightly-cb8104839c141609d99f1254459ef3a4f1bd4263
• Ran the documented recipe generator:

docker run --rm -v "$PWD":/work -w /work node:24-bookworm bash -lc 'corepack enable && pnpm install --frozen-lockfile && node scripts/build-recipes-api.mjs'

Result:

✓ JSON API: 153 models (130 with recommended_command, 670 default-hw alternatives, 1206 per-hw renderings), 139 promoted variants, 9 strategies, 2 kv-store deployments, 2 platforms (2 variant collisions skipped)

• Synthesized the MI355X + MXFP8 + EAGLE3 command and confirmed the draft speculative_config now uses TRITON_ATTN.
• Synthesized the B200 + NVFP4 + EAGLE3 command and confirmed the draft speculative_config still uses FLASH_ATTN.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Aug 4, 2026 10:02am

[gemini-code-assist]

Code Review

This pull request adds AMD hardware overrides for the eagle3 and eagle3_gqa speculative decoding configurations in models/MiniMaxAI/MiniMax-M3.yaml, switching the attention backend from FLASH_ATTN to TRITON_ATTN. There are no review comments, and I have no feedback to provide.