[Minimax M3] Enable AR + RMSNorm + quant fusion

[Phi-C]

Motivation

Support allreduce + rmsnorm + quant fusion in Minimax M3

Technical Details

Test Plan

model_path=/shared/data/amd_int/models/MiniMax-M3-MXFP8/
export AITER_QUICK_REDUCE_QUANTIZATION=INT4
export AITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0
export ATOM_FORCE_ATTN_TRITON=1
 
python -m atom.entrypoints.openai_server \
  --model "$model_path" \
  --tensor-parallel-size 4 \
  --server-port 8013 \
  --trust-remote-code \
  --gpu-memory-utilization 0.85 \
  --block-size 128 \
  --max-model-len 32768 \
  --no-enable_prefix_caching \
  --max-num-seqs 256 \
  --kv_cache_dtype fp8 \
  --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"]}' \
  --max-num-batched-tokens 32768 2>&1 | tee m3-mxfp8-server-tp4.log

Test Result

Without fusion

With fusion

Submission Checklist

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