test_name: "multi-node-kimi-k2.5-w4a8-ep-128k-1k-TPOT50"
model: "Eco-Tech/Kimi-K2.5-W4A8"
num_nodes: 4
npu_per_node: 16

routing:
  type: "disaggregated_prefill"
  groups:
    prefiller: [ 0, 1 ]
    decoder: [ 2, 3 ]

config:
  - node_index: 0
    port_start: 7100
    dp_rpc_port: 12321
    dp_size: 4
    dp_size_local: 4
    dp_rank_start: 0
    tp_size: 4
    dp_address: "${NODE_0_IP}"

  - node_index: 1
    port_start: 7100
    dp_rpc_port: 12321
    dp_size: 4
    dp_size_local: 4
    dp_rank_start: 0
    tp_size: 4
    dp_address: "${NODE_1_IP}"

  - node_index: 2
    port_start: 7100
    dp_rpc_port: 12321
    dp_size: 8
    dp_size_local: 4
    dp_rank_start: 0
    tp_size: 4
    dp_address: "${NODE_2_IP}"

  - node_index: 3
    port_start: 7100
    dp_rpc_port: 12321
    dp_size: 8
    dp_size_local: 4
    dp_rank_start: 4
    tp_size: 4
    dp_address: "${NODE_2_IP}"

env_common: &env_common
  SERVER_PORT: "${PORT}"
  VLLM_RPC_TIMEOUT: "3600000"
  VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
  OMP_PROC_BIND: "false"
  OMP_NUM_THREADS: "1"
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  TASK_QUEUE_ENABLE: "1"
  HCCL_OP_EXPANSION_MODE: "AIV"
  ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
  VLLM_USE_MODELSCOPE: "true"

env_prefill: &env_prefill
  <<: *env_common
  HCCL_BUFFSIZE: "1536"
  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"

env_decode: &env_decode
  <<: *env_common
  VLLM_ASCEND_ENABLE_MLAPO: "1"
  HCCL_BUFFSIZE: "800"

templates:
  - node_index: 0
    envs:
      <<: *env_prefill
    server_cmd_template:
      - --host
      - "0.0.0.0"
      - --port
      - $SERVER_PORT
      - --data-parallel-size
      - ${DP_SIZE}
      - --data-parallel-rank
      - ${DP_RANK}
      - --data-parallel-address
      - ${DP_ADDRESS}
      - --data-parallel-rpc-port
      - ${DP_RPC_PORT}
      - --tensor-parallel-size
      - ${TP_SIZE}
      - --enable-expert-parallel
      - --seed
      - "1024"
      - --max-model-len
      - "133120"
      - --max-num-batched-tokens
      - "16384"
      - --max-num-seqs
      - "1"
      - --enforce-eager
      - --trust-remote-code
      - --gpu-memory-utilization
      - "0.92"
      - --no-enable-prefix-caching
      - --quantization
      - "ascend"
      - --additional-config
      - '{"enable_shared_expert_dp": true}'
      - --kv-transfer-config
      - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'

  - node_index: 1
    envs:
      <<: *env_prefill
    server_cmd_template:
      - --host
      - "0.0.0.0"
      - --port
      - $SERVER_PORT
      - --data-parallel-size
      - ${DP_SIZE}
      - --data-parallel-rank
      - ${DP_RANK}
      - --data-parallel-address
      - ${DP_ADDRESS}
      - --data-parallel-rpc-port
      - ${DP_RPC_PORT}
      - --tensor-parallel-size
      - ${TP_SIZE}
      - --enable-expert-parallel
      - --seed
      - "1024"
      - --max-model-len
      - "133120"
      - --max-num-batched-tokens
      - "16384"
      - --max-num-seqs
      - "1"
      - --enforce-eager
      - --trust-remote-code
      - --gpu-memory-utilization
      - "0.92"
      - --no-enable-prefix-caching
      - --quantization
      - "ascend"
      - --additional-config
      - '{"enable_shared_expert_dp": true}'
      - --kv-transfer-config
      - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'

  - node_index: 2
    envs:
      <<: *env_decode
    server_cmd_template:
      - --host
      - "0.0.0.0"
      - --port
      - $SERVER_PORT
      - --data-parallel-size
      - ${DP_SIZE}
      - --data-parallel-rank
      - ${DP_RANK}
      - --data-parallel-address
      - ${DP_ADDRESS}
      - --data-parallel-rpc-port
      - ${DP_RPC_PORT}
      - --tensor-parallel-size
      - ${TP_SIZE}
      - --enable-expert-parallel
      - --seed
      - "1024"
      - --max-model-len
      - "133720"
      - --max-num-batched-tokens
      - "32"
      - --max-num-seqs
      - "8"
      - --trust-remote-code
      - --gpu-memory-utilization
      - "0.91"
      - --quantization
      - "ascend"
      - --no-enable-prefix-caching
      - --speculative-config
      - '{"method": "eagle3", "model": "lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
      - --compilation-config
      - '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
      - --additional-config
      - '{"recompute_scheduler_enable": true}'
      - --kv-transfer-config
      - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'

  - node_index: 3
    envs:
      <<: *env_decode
    server_cmd_template:
      - --host
      - "0.0.0.0"
      - --port
      - $SERVER_PORT
      - --data-parallel-size
      - ${DP_SIZE}
      - --data-parallel-rank
      - ${DP_RANK}
      - --data-parallel-address
      - ${DP_ADDRESS}
      - --data-parallel-rpc-port
      - ${DP_RPC_PORT}
      - --tensor-parallel-size
      - ${TP_SIZE}
      - --enable-expert-parallel
      - --seed
      - "1024"
      - --max-model-len
      - "133720"
      - --max-num-batched-tokens
      - "32"
      - --max-num-seqs
      - "8"
      - --trust-remote-code
      - --gpu-memory-utilization
      - "0.91"
      - --quantization
      - "ascend"
      - --no-enable-prefix-caching
      - --speculative-config
      - '{"method": "eagle3", "model": "lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
      - --compilation-config
      - '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
      - --additional-config
      - '{"recompute_scheduler_enable": true}'
      - --kv-transfer-config
      - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'

benchmarks:
  perf:
    case_type: performance
    dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs1000_kimi
    request_conf: vllm_api_stream_chat
    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
    trust_remote_code: true
    num_prompts: 512
    max_out_len: 1024
    batch_size: 128
    request_rate: 3
    baseline: 1
    threshold: 0.95
  acc_aime2025:
    case_type: accuracy
    dataset_path: vllm-ascend/aime2025
    request_conf: vllm_api_general_chat
    dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
    max_out_len: 65536
    temperature: 0.0
    top_p: 1
    top_k: -1
    repetition_penalty: 1.0
    batch_size: 32
    baseline: 95
    threshold: 10
  acc_GPQA:
    case_type: accuracy
    dataset_path: vllm-ascend/gpqa
    request_conf: vllm_api_general_chat
    dataset_conf: gpqa/gpqa_gen_0_shot_cot_chat_prompt
    max_out_len: 65536
    batch_size: 32
    baseline: 87.6
    threshold: 5