model_name: "vllm-ascend/Qwen3-VL-8B-Instruct-W8A8"
model_type: "vllm-vlm"
hardware: "Atlas A2 Series"

serve:
  tensor_parallel_size: 1
  dtype: auto
  max_model_len: 8192
  gpu_memory_utilization: 0.8
  trust_remote_code: false
  quantization: ascend

tasks:
- name: "mmmu_val"
  metrics:
  - name: "acc,none"
    value: 0.52

num_fewshot: 0
apply_chat_template: true
fewshot_as_multiturn: false
batch_size: 32