[atom-vllm] enable minimax_m3 for atom vllm

[lirui927]

Motivation

Enable MiniMax-M3 to run through the ATOM vLLM plugin path, including dense attention and sparse attention support, so MiniMax-M3 can use vLLM serving, KV cache management, and cudagraph execution.

Technical Details

• Add MiniMax-M3 vLLM attention adapters:
  • MiniMaxM3DenseAttentionForVllm
  • MiniMaxM3SparseAttentionForVllm
• Route MiniMax-M3 dense attention through vLLM Attention while keeping ATOM-owned projection, Q/K norm, RoPE, and checkpoint layout unchanged.
• Add MiniMax-M3 sparse attention support under the ATOM vLLM attention backend instead of relying on model-local/community backend code.
• Support fp8 KV cache scale layout for MiniMax-M3 sparse attention.

Test Plan

• Launch MiniMax-M3 vLLM service with ATOM adapters enabled.
• Run GSM8K accuracy validation:
  • MIXFP8
  • MIXFP4
  • MIXFP8 + kv_fp8
  • MIXFP4 + kv_fp8

MODEL=/shared/data/amd_int/models/MiniMax-M3-MXFP8

vllm serve "$MODEL" \
    --dtype auto \
    --load-format auto \
    --host localhost \
    --port 8001 \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 32768 \
    --max-num-batched-tokens 32768 \
    --block-size 128 \
    --no-async-scheduling \
    --kv-cache-dtype fp8 \
    --no-enable-prefix-caching \
    --language-model-only \
    --no-trust-remote-code \
    --additional-config '{"online_quant_config": {"global_quant_config": "ptpc_fp8", "exclude_layer": ["lm_head", "model.embed_tokens", "vision_tower", "multi_modal_projector", "patch_merge_mlp", "*block_sparse_moe"]}}' \
    --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
    2>&1 | tee log_m3_mxfp8_vllm.log

MODEL=/shared/data/amd_int/models/MiniMax-M3-MXFP4

vllm serve "$MODEL" \
    --dtype auto \
    --load-format auto\
    --host localhost \
    --port 8001 \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 32768 \
    --max-num-batched-tokens 32768 \
    --block-size 128 \
    --kv-cache-dtype fp8 \
    --no-enable-prefix-caching \
    --language-model-only \
    --no-trust-remote-code \
    --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
    2>&1 | tee log_m3_mxfp4_vllm_0625.log

Test Result

FP8-KV16

FP8-KV8

FP4-KV16

FP4-KV8

每组多次测试精度结果

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.