[CI][ROCm][Disagg] Enable breakable CUDA graphs for MiniMax-M3-MXFP8 decode

[avininjamay8]

Purpose

Follow-up to #54782. After that merge, MiniMax-M3-MXFP8 decode fails in our ROCm disagg CI:

RuntimeError: piecewise CUDA graphs (cudagraph_mode=FULL_AND_PIECEWISE) unavailable ...
Set VLLM_USE_BREAKABLE_CUDAGRAPH=1

Prefill runs eager; decode uses piecewise graphs. Set VLLM_USE_BREAKABLE_CUDAGRAPH: "1" in models.yaml so bring-up and GSM8K pass on nightly.

Changes

.buildkite/amd-disagg/models.yaml only:

• MiniMax-M3-MXFP8 env: VLLM_USE_BREAKABLE_CUDAGRAPH: "0" → "1"

No launcher, MoRIIO, or harness script changes.

Test Plan

MoRIIO 1P1D TP8 disaggregated serving, ROUTER_TYPE=proxy, RUN_AFTER_HEALTH=accuracy, image vllm/vllm-openai-rocm:nightly.

MiniMax-M3-MXFP8 — 1P1D TP8

export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_MOE=1
export VLLM_ROCM_USE_AITER_RMSNORM=1
export VLLM_USE_BREAKABLE_CUDAGRAPH=1
export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT6

# PREFILL — node0
vllm serve /data/models2/MiniMax-M3-MXFP8 \
    --host $PREFILL_IP --port 8100 -tp 8 \
    --trust-remote-code \
    --attention-backend TRITON_ATTN \
    --block-size 128 \
    --language-model-only \
    --kv-cache-dtype fp8 \
    --gpu-memory-utilization 0.85 \
    --enforce-eager \
    --kv-transfer-config '{"kv_connector":"MoRIIOConnector","kv_role":"kv_producer","kv_port":"14579",
      "kv_connector_extra_config":{"proxy_ip":"'$PREFILL_IP'","proxy_port":"10001",
      "proxy_ping_port":"36367","http_port":"8100","local_ping_port":"61555",
      "handshake_port":"6301","notify_port":"61005"}}'

# DECODE — node1
vllm serve /data/models2/MiniMax-M3-MXFP8 \
    --host $DECODE_IP --port 8200 -tp 8 \
    --trust-remote-code \
    --attention-backend TRITON_ATTN \
    --block-size 128 \
    --language-model-only \
    --kv-cache-dtype fp8 \
    --gpu-memory-utilization 0.85 \
    --kv-transfer-config '{"kv_connector":"MoRIIOConnector","kv_role":"kv_consumer","kv_port":"14579",
      "kv_connector_extra_config":{"proxy_ip":"'$PREFILL_IP'","proxy_port":"10001",
      "proxy_ping_port":"36367","http_port":"8200","local_ping_port":"61555",
      "handshake_port":"6301","notify_port":"61005"}}'

GSM8K evaluation

lm_eval --model local-completions \
    --tasks gsm8k \
    --model_args "model=/data/models2/MiniMax-M3-MXFP8,base_url=http://127.0.0.1:10001/v1/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,trust_remote_code=True,timeout=7200" \
    --limit 250

Test Result

MiniMax-M3-MXFP8, MoRIIO 1P1D TP8 (proxy). GSM8K flexible-extract, threshold 0.90.

Variant	Health	GSM8K exact_match
#54782 + upstream models.yaml (VLLM_USE_BREAKABLE_CUDAGRAPH=0)	Fail (decode piecewise CUDA graphs unavailable)	—
#54782 + this fix	Pass	0.9128 (PASS)

[avininjamay8]

@lcskrishna were you mentioning abt this issue in Thursday's evening meeting , will close this incase you addressed already.

[avininjamay8]

Closing stuck draft PR; reopening as ready-for-review (non-draft).

[avininjamay8]

Closing stuck draft; replacing with ready-for-review PR from fix/mm3-breakable-cudagraph-open.

[avininjamay8]

Superseded by #56075 (non-draft replacement).