# ==========================================
# ACTUAL TEST CASES
# ==========================================
test_cases:
  - name: "Qwen3.5-397B-A17B-w8a8-64k-1k-90-50-single"
    model: "Eco-Tech/Qwen3.5-397B-A17B-w8a8-mtp"
    envs:
      VLLM_USE_MODELSCOPE: "true"
      PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
      HCCL_BUFFSIZE: "2048"
      OMP_NUM_THREADS: "1"
      TASK_QUEUE_ENABLE: "1"
      VLLM_ASCEND_ENABLE_FUSED_MC2: "0"
      VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
      VLLM_ASCEND_BALANCE_SCHEDULING: "1"
      SERVER_PORT: "DEFAULT_PORT"
    server_cmd:
      - "--port"
      - "$SERVER_PORT"
      - --data-parallel-size
      - "1"
      - --tensor-parallel-size
      - "16"
      - --quantization
      - "ascend"
      - --allowed-local-media-path
      - "/"
      - --seed
      - "1024"
      - --enable-prefix-caching
      - --enable-expert-parallel
      - --max-model-len
      - "140000"
      - --max-num-batched-tokens
      - "16384"
      - --max-num-seqs
      - "32"
      - --gpu-memory-utilization
      - "0.9"
      - --compilation-config
      - '{"cudagraph_capture_sizes":[1,4,8,16,24,32,48,64,96,108,128,192],"cudagraph_mode":"FULL_DECODE_ONLY"}'
      - --speculative_config
      - '{"method": "qwen3_5_mtp","num_speculative_tokens": 3}'
      - --additional-config
      - '{"enable_cpu_binding":true}'
      - --trust-remote-code
      - --async-scheduling
      - --mm-processor-cache-gb
      - "0"
      - --mm-encoder-tp-mode
      - "data"

    benchmarks:
      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
        batch_size: 32
        baseline: 90
        threshold: 10
        temperature: 0.6
        top_p: 0.95
        top_k: 20
        min_p: 0.0
        presence_penalty: 0.0
        repetition_penalty: 1.0
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs1000_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 96
        max_out_len: 1024
        batch_size: 24
        request_rate: 5
        baseline: 374.5353
        threshold: 0.97
      perf_128k:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs100_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 80
        max_out_len: 1024
        batch_size: 20
        request_rate: 5
        baseline: 345.3004
        threshold: 0.97