test_name: "test DeepSeek-V3.2-W8A8-EP disaggregated_prefill"
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
num_nodes: 4
npu_per_node: 16
env_common:
  HCCL_OP_EXPANSION_MODE: "AIV"
  VLLM_USE_MODELSCOPE: true
  SERVER_PORT: 8080
  OMP_PROC_BIND: false
  OMP_NUM_THREADS: 10
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  HCCL_BUFFSIZE: 1024
  VLLM_TORCH_PROFILER_WITH_STACK: 0
  ASCEND_AGGREGATE_ENABLE: 1
  ASCEND_TRANSPORT_PRINT: 1
  ACL_OP_INIT_MODE: 1
  ASCEND_A3_ENABLE: 1
  VLLM_NIXL_ABORT_REQUEST_TIMEOUT: 300000
  VLLM_ENGINE_READY_TIMEOUT_S: 1800
  HCCL_CONNECT_TIMEOUT: 1200
  HCCL_INTRA_PCIE_ENABLE: 1
  HCCL_INTRA_ROCE_ENABLE: 0

disaggregated_prefill:
  enabled: true
  prefiller_host_index: [0, 1]
  decoder_host_index: [2, 3]


deployment:
  -
    envs:
      VLLM_ASCEND_ENABLE_FLASHCOMM1: 1
    server_cmd: >
      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
          --host 0.0.0.0
          --port $SERVER_PORT
          --data-parallel-size 2
          --data-parallel-start-rank 0
          --data-parallel-size-local 1
          --data-parallel-address $LOCAL_IP
          --tensor-parallel-size 16
          --enable-expert-parallel
          --speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}' 
          --seed 1024
          --quantization ascend
          --max-num-seqs 64
          --max-model-len 68000 
          --max-num-batched-tokens 16384 
          --trust-remote-code
          --gpu-memory-utilization 0.82
          --enforce-eager
          --no-enable-prefix-caching
          --additional-config '{"enable_cpu_binding" : false, "enable_sfa_cp":false,"layer_sharding": ["q_b_proj", "o_proj"]}' 
          --tokenizer-mode deepseek_v32 
          --reasoning-parser deepseek_v3
          --kv-transfer-config
          '{"kv_connector": "MooncakeConnectorV1",
          "kv_role": "kv_producer",
          "kv_port": "30000",
          "engine_id": "0",
          "kv_connector_extra_config": {
                    "prefill": {
                            "dp_size": 2,
                            "tp_size": 16
                    },
                    "decode": {
                            "dp_size": 8,
                            "tp_size": 4
                    }
              }
          }'


  -
    envs:
      VLLM_ASCEND_ENABLE_FLASHCOMM1: 1
    server_cmd: >
      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
          --host 0.0.0.0
          --headless
          --port $SERVER_PORT
          --data-parallel-size 2
          --data-parallel-start-rank 1
          --data-parallel-size-local 1
          --data-parallel-address $MASTER_IP
          --tensor-parallel-size 16
          --enable-expert-parallel
          --speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}' 
          --seed 1024
          --quantization ascend
          --max-num-seqs 64
          --max-model-len 68000 
          --max-num-batched-tokens 16384 
          --trust-remote-code
          --gpu-memory-utilization 0.82
          --enforce-eager
          --no-enable-prefix-caching
          --additional-config '{"enable_cpu_binding" : false, "enable_sfa_cp":false,"layer_sharding": ["q_b_proj", "o_proj"]}' 
          --tokenizer-mode deepseek_v32 
          --reasoning-parser deepseek_v3
          --kv-transfer-config
          '{"kv_connector": "MooncakeConnectorV1",
          "kv_role": "kv_producer",
          "kv_port": "30000",
          "engine_id": "0",
          "kv_connector_extra_config": {
                    "prefill": {
                            "dp_size": 2,
                            "tp_size": 16
                    },
                    "decode": {
                            "dp_size": 8,
                            "tp_size": 4
                    }
              }
          }'
  -
    envs:
      VLLM_ASCEND_ENABLE_MLAPO: 1
      TASK_QUEUE_ENABLE: 1
    server_cmd: >
      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
          --host 0.0.0.0
          --port $SERVER_PORT
          --data-parallel-size 8
          --data-parallel-size-local 4
          --data-parallel-start-rank 0
          --data-parallel-address $LOCAL_IP
          --data-parallel-rpc-port 13389
          --tensor-parallel-size 4
          --enable-expert-parallel
          --speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
          --seed 1024
          --quantization ascend
          --max-model-len 68000 
          --max-num-batched-tokens 12
          --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3, 6, 9, 12]}' 
          --trust-remote-code
          --max-num-seqs 4
          --gpu-memory-utilization 0.95 
          --no-enable-prefix-caching 
          --async-scheduling 
          --additional-config '{"enable_cpu_binding" : false,"recompute_scheduler_enable" : true}' 
          --tokenizer-mode deepseek_v32
          --reasoning-parser deepseek_v3
          --kv-transfer-config 
          '{"kv_connector": "MooncakeConnectorV1",
          "kv_role": "kv_consumer",
          "kv_port": "30200",
          "engine_id": "1",
          "kv_connector_extra_config": {
                      "prefill": {
                              "dp_size": 2,
                              "tp_size": 16
                      },
                      "decode": {
                              "dp_size": 8,
                              "tp_size": 4
                      }
              }
          }'

  -
    envs:
      VLLM_ASCEND_ENABLE_MLAPO: 1
      TASK_QUEUE_ENABLE: 1
    server_cmd: >
      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
          --host 0.0.0.0
          --headless
          --port $SERVER_PORT
          --data-parallel-size 8
          --data-parallel-size-local 4
          --data-parallel-start-rank 4
          --data-parallel-address $MASTER_IP
          --data-parallel-rpc-port 13389
          --tensor-parallel-size 4
          --enable-expert-parallel
          --speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
          --seed 1024
          --quantization ascend
          --max-model-len 68000 
          --max-num-batched-tokens 12
          --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3, 6, 9, 12]}' 
          --trust-remote-code
          --max-num-seqs 4 
          --gpu-memory-utilization 0.95 
          --no-enable-prefix-caching 
          --async-scheduling 
          --additional-config '{"enable_cpu_binding" : false,"recompute_scheduler_enable" : true}' 
          --tokenizer-mode deepseek_v32
          --reasoning-parser deepseek_v3
          --kv-transfer-config 
          '{"kv_connector": "MooncakeConnectorV1",
          "kv_role": "kv_consumer",
          "kv_port": "30200",
          "engine_id": "1",
          "kv_connector_extra_config": {
                      "prefill": {
                              "dp_size": 2,
                              "tp_size": 16
                      },
                      "decode": {
                              "dp_size": 8,
                              "tp_size": 4
                      }
              }
          }'
benchmarks:
  perf:
    case_type: performance
    dataset_path: vllm-ascend/GSM8K-in3500-bs2800
    request_conf: vllm_api_stream_chat
    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
    num_prompts: 512
    max_out_len: 1500
    batch_size: 512
    request_rate: 11.2
    baseline: 1146
    threshold: 0.97
  acc:
    case_type: accuracy
    dataset_path: vllm-ascend/gsm8k-lite
    request_conf: vllm_api_general_chat
    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
    max_out_len: 4096
    batch_size: 64
    baseline: 95
    threshold: 5