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

serve:
  tensor_parallel_size: 2
  dtype: auto
  max_model_len: 128000
  gpu_memory_utilization: 0.7
  trust_remote_code: false
  enable_expert_parallel: true

tasks:
- name: "mmmu_val"
  metrics:
  - name: "acc,none"
    value: 0.58

num_fewshot: 0
apply_chat_template: true
fewshot_as_multiturn: false
batch_size: "auto"