[AMD][ROCm]Minimax-M3 Eagle3 MTP attn backend override for perf

[hongxiayang]

For Minimax-M3, override the drafter's attention backend to TRITON_ATTN.

When using eagle3 MTP:
Using

    --speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"TRITON_ATTN\"}" \

has much better performance than without it, like below:

    --speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS} \

For 8k/1k, con1 to con256, perf improved from 4% to 151%

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jul 6, 2026 10:33pm

[gemini-code-assist]

Code Review

This pull request introduces hardware overrides for AMD (ROCm) in the MiniMax-M3 model configuration, specifying the use of the 'TRITON_ATTN' attention backend instead of 'FLASH_ATTN' for speculative decoding to improve performance. There are no review comments, so I have no feedback to provide.