[Spec Decode][ROCm][Perf] Default EAGLE3 Llama drafter to AITER unified attention

[tanpinsiang]

This PR changes the default attention backend for ROCm EAGLE3 Llama draft models from the platform auto-selected ROCM_ATTN path to ROCM_AITER_UNIFIED_ATTN.

Why

On ROCm, EAGLE3 Llama drafter attention currently resolves to ROCM_ATTN by default. For MiniMax-M3 EAGLE3 serving, that path is much slower than the available AITER unified-attention backend. The default should select the faster backend for this drafter family while still preserving explicit user overrides in speculative_config.attention_backend.

ROCM_AITER_UNIFIED_ATTN is also the most robust fast option here: ROCM_AITER_FA does not currently work for this stack because MiniMax-M3 requires block size 128, while the backend only advertises [16, 32]. Forcing 128 experimentally allowed startup but failed at runtime in AITER paged_attention_v1, so it is not a valid replacement.

Benchmark

Benchmark stack: upstream main + #47391 + this PR.

The code change in this PR does not depend on #47391, but the benchmark includes #47391 so MiniMax-M3 EAGLE3 acceptance and throughput are measured on the representative stack.

Note: These benchmark numbers were collected before recipes#621 clarified the nested text_config override. The command below is updated to the working form; the benchmark comparison remains comparable because all rows used the same config, but absolute tok/s should be rerun with nested index-topk reuse.

Server setup:

• Model: amd/MiniMax-M3-MXFP4
• Draft model: Inferact/MiniMax-M3-EAGLE3
• TP: 4
• Block size: 128
• Target attention backend: TRITON_ATTN
• Speculative tokens: 2
• KV cache dtype: fp8
• Dataset: random 8k input / 1k output
• Concurrency: 256
• Prompts: 2560

Drafter backend	Status	Output tok/s/GPU	Total tok/s/GPU	Mean TPOT ms	P99 TPOT ms	Acceptance	Improvement vs ROCM_ATTN
ROCM_ATTN	pass	247.045	2328.647	253.811	392.013	68.84%	baseline
TRITON_ATTN	pass	843.506	7950.873	71.576	108.118	67.91%	+241.44%
ROCM_AITER_FA	do not support Block size: 128	n/a	n/a	n/a	n/a	n/a	n/a
ROCM_AITER_UNIFIED_ATTN	pass	858.428	8091.533	70.158	107.299	67.89%	+247.48%

ROCM_AITER_UNIFIED_ATTN is 3.47x the current ROCM_ATTN default and slightly faster than TRITON_ATTN on this workload.

Test command

vllm serve amd/MiniMax-M3-MXFP4 \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --block-size 128 \
  --no-enable-prefix-caching \
  --language-model-only \
  --max-model-len 32768 \
  --attention-backend TRITON_ATTN \
  --moe-backend aiter \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice \
  --reasoning-parser minimax_m3 \
  --kv-cache-dtype fp8 \
  --hf-overrides '{"text_config": {"use_index_cache": true, "index_topk_freq": 4}}' \
  --speculative-config '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3","num_speculative_tokens":2}' \
  --no-async-scheduling

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[hongxiayang]

We can use TRITON_ATTN, which do not need to depend on aiter to be installed.
Also I used TRITON_ATTN for the minimax-m3 eagle3 recipe in this PR: vllm-project/recipes#615 which has shown very good performance.

[tanpinsiang]

Superseded by the MiniMax-M3 recipe update in vllm-project/recipes#615.
Closing this PR.