main
yaml 6 lines 275 Bytes
Raw
1 model_name: "Qwen/Qwen3-4B"
2 accuracy_threshold: 0.75
3 num_questions: 1319
4 num_fewshot: 5
5 # Triton serves the unquantized boundary layers with TurboQuant-compatible layout.
6 server_args: "--kv-cache-dtype turboquant_3bit_nc --max-model-len 4096 --attention-backend TRITON_ATTN"