test_name: "test DeepSeek-V3.2-W8A8 on A3"
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
num_nodes: 2
npu_per_node: 16
env_common:
  HCCL_OP_EXPANSION_MODE: "AIV"

  VLLM_USE_MODELSCOPE: true
  HCCL_BUFFSIZE: 1024
  SERVER_PORT: 8080
  OMP_PROC_BIND: false
  OMP_NUM_THREADS: 1
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  VLLM_ASCEND_ENABLE_MLAPO: 1
  VLLM_ASCEND_ENABLE_FLASHCOMM1: 1
  ASCEND_A3_EBA_ENABLE: 1

# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case

deployment:
  -
    server_cmd: >
      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
      --host 0.0.0.0
      --port $SERVER_PORT
      --data-parallel-size 4
      --data-parallel-size-local 2
      --data-parallel-address $LOCAL_IP
      --data-parallel-rpc-port 13399
      --tensor-parallel-size 8
      --quantization ascend
      --seed 1024
      --enable-expert-parallel
      --max-num-seqs 16
      --max-model-len 68000
      --max-num-batched-tokens 4096
      --no-enable-prefix-caching
      --gpu-memory-utilization 0.85
      --trust-remote-code
      --speculative-config '{"num_speculative_tokens": 3, "method":"deepseek_mtp"}'
      --compilation-config '{"cudagraph_capture_sizes": [8, 16, 24, 32, 40, 48], "cudagraph_mode": "FULL_DECODE_ONLY"}' 
      --additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}'
      --tokenizer-mode deepseek_v32
      --reasoning-parser deepseek_v3

  -
    server_cmd: >
      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
      --headless
      --data-parallel-size 4
      --data-parallel-rpc-port 13399
      --data-parallel-size-local 2
      --data-parallel-start-rank 2
      --data-parallel-address $MASTER_IP
      --tensor-parallel-size 8
      --quantization ascend
      --seed 1024
      --enable-expert-parallel
      --max-num-seqs 16
      --max-model-len 68000
      --max-num-batched-tokens 4096
      --no-enable-prefix-caching
      --gpu-memory-utilization 0.85
      --trust-remote-code
      --speculative-config '{"num_speculative_tokens": 3, "method":"deepseek_mtp"}'
      --compilation-config '{"cudagraph_capture_sizes": [8, 16, 24, 32, 40, 48], "cudagraph_mode": "FULL_DECODE_ONLY"}' 
      --additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}'
      --tokenizer-mode deepseek_v32
      --reasoning-parser deepseek_v3
benchmarks:
  perf_short_warmup:
    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: 1
    max_out_len: 3000
    batch_size: 512
    request_rate: 11.2
    baseline: 1253.8466 
    threshold: 0.97

  perf_long_warmup:
    case_type: performance
    dataset_path: vllm-ascend/GSM8K-in64000-bs2800
    request_conf: vllm_api_stream_chat
    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
    num_prompts: 1
    max_out_len: 3000
    batch_size: 1
    request_rate: 11.2
    baseline: 1253.8466 
    threshold: 0.97

  perf_short:
    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: 3000
    batch_size: 1
    request_rate: 11.2
    baseline: 148 # after switch vllm to 0.15.0, the baseline reduced significantly, need to confirm if it's a regression or just a more strict measurement
    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

  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: 80000
    batch_size: 32
    baseline: 40
    threshold: 7