[avininjamay8]

Purpose

Add MiniMax-M3-MXFP8 to the ATOMesh disaggregated CI model catalog for 2P+1D TP8 and 2P+1D DPA TP8 Mooncake PD runs.

Changes

• .github/benchmark/models_atomesh.yaml — add minimax-m3-fp8-2p1d-tp8 and minimax-m3-fp8-2p1d-dpa-tp8 nightly entries (2P+1D TP8; DPA/TBO on prefill for DPA case; GSM8K run_eval: true, eval_concurrency: [128, 768])
• .github/workflows/atomesh-benchmark.yaml — add both case names to manual dispatch MiniMax list

Test Plan

ATOMesh 2P+1D, Mooncake KV, rocm/atom-dev:latest.

MiniMax-M3-MXFP8 — 2P+1D TP8

export AITER_QUICK_REDUCE_QUANTIZATION=INT4
export ATOM_FORCE_ATTN_TRITON=1

# PREFILL — TP8 (each of 2 prefill nodes)
python3 -m atom.entrypoints.openai_server \
    --model ${ATOMESH_MODEL_ROOT}/MiniMax-M3-MXFP8 \
    --host 0.0.0.0 --server-port 8010 \
    --trust-remote-code \
    --tensor-parallel-size 8 \
    --kv_cache_dtype fp8 \
    --block-size 128 \
    --gpu-memory-utilization 0.8 \
    --max-model-len 32768 \
    --max-num-seqs 256 \
    --max-num-batched-tokens 32768 \
    --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"]}' \
    --kv-transfer-config '{"kv_role":"kv_producer","kv_connector":"mooncake","handshake_port":6301}' \
    --no-enable_prefix_caching \
    --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'

# DECODE — TP8
python3 -m atom.entrypoints.openai_server \
    --model ${ATOMESH_MODEL_ROOT}/MiniMax-M3-MXFP8 \
    --host 0.0.0.0 --server-port 8020 \
    --trust-remote-code \
    --tensor-parallel-size 8 \
    --kv_cache_dtype fp8 \
    --block-size 128 \
    --gpu-memory-utilization 0.8 \
    --max-model-len 32768 \
    --max-num-seqs 1024 \
    --max-num-batched-tokens 32768 \
    --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"]}' \
    --kv-transfer-config '{"kv_role":"kv_consumer","kv_connector":"mooncake","handshake_port":6301}' \
    --no-enable_prefix_caching \
    --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'

# ROUTER
atomesh launch \
    --host 0.0.0.0 --port 8000 \
    --pd-disaggregation \
    --prefill "http://${PREFILL_IP_1}:8010" \
    --prefill "http://${PREFILL_IP_2}:8010" \
    --decode  "http://${DECODE_IP}:8020" \
    --policy random \
    --backend atom \
    --log-level info \
    --disable-circuit-breaker \
    --prometheus-port 29100

MiniMax-M3-MXFP8 — 2P+1D DPA TP8

Same as above, plus on prefill: --enable-dp-attention --enable-tbo prefill; on decode: --enable-dp-attention; and DPA MXFP8 online_quant_config exclude list (*.gate.*, *.block_sparse_moe.experts*) on both roles.

GSM8K evaluation

lm_eval --model local-chat-completions \
    --tasks gsm8k \
    --model_args "model=${ATOMESH_MODEL_ROOT}/MiniMax-M3-MXFP8,base_url=http://127.0.0.1:8000/v1/chat/completions,num_concurrent=128,max_retries=3,max_gen_toks=16384" \
    --num_fewshot 5 \
    --batch_size 65 \
    --apply_chat_template \
    --fewshot_as_multiturn

Report exact_match,flexible-extract.

Test Result

Spur MI350X ATOMesh nightly cells, rocm/atom-dev:latest, GSM8K flexible-extract @ num_concurrent=128.

Model / config	GSM8K exact_match
MiniMax-M3-MXFP8 2P+1D TP8	0.951
MiniMax-M3-MXFP8 2P+1D DPA TP8	0.949

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 1 --add-label <label>