# ==========================================
# Shared Configurations
# ==========================================

_envs: &envs
  HCCL_BUFFSIZE: "512"
  SERVER_PORT: "DEFAULT_PORT"
  OMP_PROC_BIND: "false"
  OMP_NUM_THREADS: "1"
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  HCCL_OP_EXPANSION_MODE: "AIV"
  VLLM_ASCEND_BALANCE_SCHEDULING: "1"
  VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE: "1"
  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"

_server_cmd: &server_cmd
  - "--enable-expert-parallel"
  - "--tensor-parallel-size"
  - "8"
  - "--data-parallel-size"
  - "2"
  - "--port"
  - "$SERVER_PORT"
  - "--max-model-len"
  - "8192"
  - "--max-num-batched-tokens"
  - "8192"
  - "--max-num-seqs"
  - "16"
  - "--async-scheduling"
  - "--quantization"
  - "ascend"
  - "--trust-remote-code"
  - "--gpu-memory-utilization"
  - "0.9"
  - "--speculative-config"
  - '{"num_speculative_tokens": 3, "method":"mtp"}'
  - "--additional-config"
  - '{"enable_shared_expert_dp": true, "ascend_fusion_config": {"fusion_ops_gmmswigluquant": false}}'

_benchmarks: &benchmarks
  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: 8
    baseline: 95
    threshold: 5

# ==========================================
# ACTUAL TEST CASES
# ==========================================

test_cases:
  - name: "GLM-4.7-TP8-DP2-decodegraph"
    model: "Eco-Tech/GLM-4.7-W8A8-floatmtp"
    envs:
      <<: *envs
    server_cmd: *server_cmd
    server_cmd_extra:
      - "--compilation-config"
      - '{"cudagraph_capture_sizes": [1,2,4,8,16,32,64,128,256,512], "cudagraph_mode": "FULL_DECODE_ONLY"}'
    benchmarks:
      <<: *benchmarks