| 1 | model_name: "nvidia/Qwen3-Next-80B-A3B-Instruct-NVFP4" |
| 2 | accuracy_threshold: 0.85 |
| 3 | num_questions: 1319 |
| 4 | num_fewshot: 5 |
| 5 | max_concurrency: 16 |
| 6 | rocm_request_timeout_seconds: 1800 |
| 7 | min_acceptance_length: 1.5 |
| 8 | server_args: >- |
| 9 | --enforce-eager |
| 10 | --max-model-len 4096 |
| 11 | --max-num-seqs 16 |
| 12 | --max-num-batched-tokens 32768 |
| 13 | --gpu-memory-utilization 0.5 |
| 14 | --tensor-parallel-size 2 |
| 15 | --enable-expert-parallel |
| 16 | --all2all-backend allgather_reducescatter |
| 17 | --attention-backend TRITON_ATTN |
| 18 | --no-async-scheduling |
| 19 | --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}' |