test_name: "Kimi-K2.7-W4A8-128k-1k-TPOT50-PD"
model: "Eco-Tech/Kimi-K2.7-Code-w4a8"
num_nodes: 2
npu_per_node: 16

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

config:
  - node_index: 0
    port_start: 7100
    dp_rpc_port: 12321
    dp_size: 2
    dp_size_local: 2
    dp_rank_start: 0
    tp_size: 8
    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}"

env_common: &env_common
  SERVER_PORT: "${PORT}"
  VLLM_RPC_TIMEOUT: "3600000"
  VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
  HCCL_EXEC_TIMEOUT: "204"
  HCCL_CONNECT_TIMEOUT: "120"
  OMP_PROC_BIND: "false"
  OMP_NUM_THREADS: "1"
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  TASK_QUEUE_ENABLE: "1"
  HCCL_OP_EXPANSION_MODE: "AIV"
  VLLM_ASCEND_ENABLE_MLAPO: "1"
  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
  ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
  VLLM_USE_MODELSCOPE: "true"
env_prefill: &env_prefill
  <<: *env_common
  HCCL_BUFFSIZE: "512"
  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"

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

templates:
  - node_index: 0
    envs:
      <<: *env_prefill
    server_cmd_template:
      - --allowed-local-media-path
      - "/"
      - --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}
      - --safetensors-load-strategy
      - 'prefetch'
      - --enable-expert-parallel
      - --seed
      - "1024"
      - --max-model-len
      - "133120"
      - --max-num-batched-tokens
      - "8192"
      - --max-num-seqs
      - "8"
      - --enforce-eager
      - --trust-remote-code
      - --gpu-memory-utilization
      - "0.9"
      - --speculative-config
      - '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
      - --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": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
  - node_index: 1
    envs:
      <<: *env_decode
    server_cmd_template:
      - --allowed-local-media-path
      - "/"
      - --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
      - --safetensors-load-strategy
      - 'prefetch'
      - --seed
      - "1024"
      - --max-model-len
      - "133120"
      - --max-num-batched-tokens
      - "256"
      - --max-num-seqs
      - "16"
      - --trust-remote-code
      - --gpu-memory-utilization
      - "0.92"
      - --speculative-config
      - '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 15}'
      - --compilation-config
      - '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
      - --additional-config
      - '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
      - --kv-transfer-config
      - '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'

benchmarks:
  acc:
    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: 32768
    batch_size: 8
    baseline: 93.33
    threshold: 10
    temperature: 1.0
    top_p: 1
  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
    num_prompts: 40
    max_out_len: 1024
    batch_size: 10
    request_rate: 0.3
    baseline: 220.503
    threshold: 0.97