enginex-ascend-910-vllm/tests/e2e/models/configs/Qwen3-30B-A3B.yaml

model_name: "Qwen/Qwen3-30B-A3B"
model_type: "vllm"
hardware: "Atlas A2 Series"

serve:
  tensor_parallel_size: 2
  dtype: auto
  max_model_len: 4096
  gpu_memory_utilization: 0.6
  trust_remote_code: false
  enable_expert_parallel: true

tasks:
- name: "gsm8k"
  metrics:
  - name: "exact_match,strict-match"
    value: 0.89
  - name: "exact_match,flexible-extract"
    value: 0.85
- name: "ceval-valid"
  metrics:
  - name: "acc,none"
    value: 0.84

num_fewshot: 5
apply_chat_template: false
fewshot_as_multiturn: false
batch_size: "auto"