| 1 | model_name: "Qwen/Qwen3-30B-A3B" |
| 2 | accuracy_threshold: 0.89 |
| 3 | tolerance: 0.03 |
| 4 | num_questions: 1319 |
| 5 | num_fewshot: 5 |
| 6 | server_args: >- |
| 7 | --max-model-len 4096 |
| 8 | --tensor-parallel-size 2 |
| 9 | --gpu-memory-utilization 0.35 |
| 10 | --moe-backend aiter |
| 11 | --quantization mxfp4 |
| 12 | --quantization-config '{"linear":{"activation":null,"weight":null}}' |
| 13 | env: |
| 14 | VLLM_ROCM_USE_AITER: "1" |
| 15 | ENABLE_CK: "0" # Avoid AITER CK-based MHA JIT compilation to save time |