# ==========================================
# ACTUAL TEST CASES
# ==========================================
_envs: &envs
  HCCL_BUFFSIZE: "1024"
  HCCL_OP_EXPANSION_MODE: "AIV"
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  OMP_NUM_THREADS: "1"
  TASK_QUEUE_ENABLE: "1"
  VLLM_USE_MODELSCOPE: "true"
  SERVER_PORT: "DEFAULT_PORT"
test_cases:
  - name: "Qwen3.5-27B-w8a8-in3.5k-140-35"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--tensor-parallel-size"
      - "2"
      - "--data-parallel-size"
      - "1"
      - "--port"
      - "$SERVER_PORT"
      - "--max-num-seqs"
      - "128"
      - "--quantization"
      - "ascend"
      - "--max-model-len"
      - "128000"
      - "--max-num-batched-tokens"
      - "16384"
      - "--trust-remote-code"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--mm-processor-cache-gb"
      - "0"
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K-in3500-bs8000-qwen3
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 140
        max_out_len: 1500
        batch_size: 35
        request_rate: 0
        baseline: 610.22
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in16k-56-14"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "18432"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
      - "--no-enable-prefix-caching"
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs200_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 56
        max_out_len: 1024
        batch_size: 14
        request_rate: 0
        baseline: 221.22
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in32k-56-14"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "34304"
      - "--enable-prefix-caching"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix90_in32768_bs100_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 56
        max_out_len: 512
        batch_size: 14
        request_rate: 0
        baseline: 197.00
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in64k-16-4"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "67584"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,6,12,18,24,30,36,42,48,54,72,78,84,90,96,102,108,144,192], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 5, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
      - "--no-enable-prefix-caching"
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix0_in65536_bs100_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 16
        max_out_len: 1024
        batch_size: 4
        request_rate: 0
        baseline: 56.17
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in64k-48-12"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "67584"
      - "--enable-prefix-caching"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs1000_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 48
        max_out_len: 1024
        batch_size: 12
        request_rate: 0
        baseline: 195.87
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in128k-28-7"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "133120"
      - "--enable-prefix-caching"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
    benchmarks:
      perf:
        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: 28
        max_out_len: 1024
        batch_size: 7
        request_rate: 0
        baseline: 101.12
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in16k-16-4"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "18432"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,6,12,18,24,30,36,42,48,54,72,78,84,90,96,102,108,144,192], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 5, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
      - "--no-enable-prefix-caching"
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs200_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 16
        max_out_len: 1024
        batch_size: 4
        request_rate: 0
        baseline: 163.77
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in32k-8-2"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "34304"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,6,12,18,24,30,36,42,48,54,72,78,84,90,96,102,108,144,192], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 5, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
      - "--no-enable-prefix-caching"
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs100_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 8
        max_out_len: 512
        batch_size: 2
        request_rate: 0
        baseline: 53.82
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in32k-16-4"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "34304"
      - "--enable-prefix-caching"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix90_in32768_bs100_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 16
        max_out_len: 512
        batch_size: 4
        request_rate: 0
        baseline: 149.23
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in64k-8-2"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "67584"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,6,12,18,24,30,36,42,48,54,72,78,84,90,96,102,108,144,192], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 5, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
      - "--no-enable-prefix-caching"
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix0_in65536_bs100_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 8
        max_out_len: 1024
        batch_size: 2
        request_rate: 0
        baseline: 48.52
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in64k-8-2-90"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "67584"
      - "--enable-prefix-caching"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
    benchmarks:
      perf:
        case_type: performance
        dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs1000_qwen
        request_conf: vllm_api_stream_chat
        dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
        num_prompts: 8
        max_out_len: 1024
        batch_size: 2
        request_rate: 0
        baseline: 87.3
        threshold: 0.97
  - name: "Qwen3.5-27B-w8a8-in128k-4-1"
    model: "Eco-Tech/Qwen3.5-27B-w8a8-mtp"
    envs:
      <<: *envs
    server_cmd:
      - "--data-parallel-size"
      - "1"
      - "--tensor-parallel-size"
      - "2"
      - "--port"
      - "$SERVER_PORT"
      - "--max-model-len"
      - "133120"
      - "--enable-prefix-caching"
      - "--max-num-batched-tokens"
      - "16384"
      - "--max-num-seqs"
      - "128"
      - "--gpu-memory-utilization"
      - "0.9"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[1,4,8,12,16,24,32,48,56,64,72,84,96,108,112,128,160,172,196,200,212,232,256,272,288,312,328,344,360,384,400,416,432,448,480,512], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative_config"
      - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
      - "--trust-remote-code"
      - "--allowed-local-media-path"
      - "/"
      - "--quantization"
      - "ascend"
      - "--mm-processor-cache-gb"
      - "0"
      - "--additional-config"
      - '{"enable_cpu_binding":true}'
    benchmarks:
      perf:
        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: 4
        max_out_len: 1024
        batch_size: 1
        request_rate: 0
        baseline: 36.44
        threshold: 0.97