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

_envs: &envs
  HCCL_BUFFSIZE: "512"
  SERVER_PORT: "DEFAULT_PORT"
  HCCL_OP_EXPANSION_MODE: "AIV"
  OMP_PROC_BIND: "false"
  OMP_NUM_THREADS: "1"
  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
  TASK_QUEUE_ENABLE: "1"
  VLLM_ASCEND_ENABLE_MLAPO: "1"
  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
  VLLM_ASCEND_ENABLE_NZ: "1"

_server_cmd: &server_cmd
  - "--enable-expert-parallel"
  - "--allowed-local-media-path"
  - "/"
  - "--tensor-parallel-size"
  - "4"
  - "--data-parallel-size"
  - "4"
  - "--port"
  - "$SERVER_PORT"
  - "--quantization"
  - "ascend"
  - "--trust-remote-code"
  - "--gpu-memory-utilization"
  - "0.9"
  - "--seed"
  - "42"
  - "--mm-processor-cache-gb"
  - "0"
  - "--mm-encoder-tp-mode"
  - "data"

_benchmark_TPOT50_32K_0_5K: &_benchmark_TPOT50_32K_0_5K
  case_type: performance
  dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs200_kimi
  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
  trust_remote_code: true
  request_rate: 0
  baseline: 1
  threshold: 0.97

_benchmark_TPOT50_32K_0_5K_pc90: &_benchmark_TPOT50_32K_0_5K_pc90
  case_type: performance
  dataset_path: vllm-ascend/GSM8K_prefix90_in32768_bs1000_kimi
  request_conf: vllm_api_stream_chat
  dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
  num_prompts: 40
  max_out_len: 512
  batch_size: 10
  trust_remote_code: true
  request_rate: 0
  baseline: 1
  threshold: 0.97

_benchmark_TPOT20_32K_0_5K: &_benchmark_TPOT20_32K_0_5K
  case_type: performance
  dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs200_kimi
  request_conf: vllm_api_stream_chat
  dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
  num_prompts: 4
  max_out_len: 512
  batch_size: 1
  trust_remote_code: true
  request_rate: 0
  baseline: 1
  threshold: 0.97

_benchmark_TPOT20_32K_0_5K_pc90: &_benchmark_TPOT20_32K_0_5K_pc90
  case_type: performance
  dataset_path: vllm-ascend/GSM8K_prefix90_in32768_bs1000_kimi
  request_conf: vllm_api_stream_chat
  dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
  num_prompts: 4
  max_out_len: 512
  batch_size: 1
  trust_remote_code: true
  request_rate: 0
  baseline: 1
  threshold: 0.97

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

test_cases:
  - name: "Kimi-K2.6-W4A8-TOPT50-32k-0.5k"
    model: "Eco-Tech/Kimi-K2.6-w4a8"
    envs:
      <<: *envs
      HCCL_BUFFSIZE: "450"
      VLLM_ASCEND_BALANCE_SCHEDULING: "1"
    server_cmd: *server_cmd
    server_cmd_extra:
      - "--no-enable-prefix-caching"
      - "--enable-chunked-prefill"
      - "--max-num-seqs"
      - "32"
      - "--max-model-len"
      - "33792"
      - "--max-num-batched-tokens"
      - "8192"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[16,32,48,64,80,96,112,128], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative-config"
      - '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
      - "--additional-config"
      - '{"enable_shared_expert_dp": true}'
    benchmarks:
      perf: *_benchmark_TPOT50_32K_0_5K
  - name: "Kimi-K2.6-W4A8-TOPT50-32k-0.5k-prefix-cache90"
    model: "Eco-Tech/Kimi-K2.6-w4a8"
    envs:
      <<: *envs
      HCCL_BUFFSIZE: "450"
      VLLM_ASCEND_BALANCE_SCHEDULING: "1"
    server_cmd: *server_cmd
    server_cmd_extra:
      - "--enable-prefix-caching"
      - "--enable-chunked-prefill"
      - "--max-num-seqs"
      - "32"
      - "--max-model-len"
      - "33792"
      - "--max-num-batched-tokens"
      - "8192"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[16,32,48,64,80,96,112,128], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative-config"
      - '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
      - "--additional-config"
      - '{"enable_shared_expert_dp": true}'
    benchmarks:
      perf: *_benchmark_TPOT50_32K_0_5K_pc90
  - name: "Kimi-K2.6-W4A8-TOPT20-32k-0.5k"
    model: "Eco-Tech/Kimi-K2.6-w4a8"
    envs:
      <<: *envs
      HCCL_BUFFSIZE: "450"
      VLLM_ASCEND_BALANCE_SCHEDULING: "1"
    server_cmd: *server_cmd
    server_cmd_extra:
      - "--no-enable-prefix-caching"
      - "--enable-chunked-prefill"
      - "--max-num-seqs"
      - "32"
      - "--max-model-len"
      - "33792"
      - "--max-num-batched-tokens"
      - "8192"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[16,32,48,64,80,96,112,128], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative-config"
      - '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
      - "--additional-config"
      - '{"enable_shared_expert_dp": true}'
    benchmarks:
      perf: *_benchmark_TPOT20_32K_0_5K
  - name: "Kimi-K2.6-W4A8-TOPT20-32k-0.5k-pc90"
    model: "Eco-Tech/Kimi-K2.6-w4a8"
    envs:
      <<: *envs
      HCCL_BUFFSIZE: "450"
      VLLM_ASCEND_BALANCE_SCHEDULING: "1"
    server_cmd: *server_cmd
    server_cmd_extra:
      - "--enable-prefix-caching"
      - "--enable-chunked-prefill"
      - "--max-num-seqs"
      - "32"
      - "--max-model-len"
      - "33792"
      - "--max-num-batched-tokens"
      - "8192"
      - "--compilation-config"
      - '{"cudagraph_capture_sizes":[16,32,48,64,80,96,112,128], "cudagraph_mode":"FULL_DECODE_ONLY"}'
      - "--speculative-config"
      - '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
      - "--additional-config"
      - '{"enable_shared_expert_dp": true}'
    benchmarks:
      perf: *_benchmark_TPOT20_32K_0_5K_pc90