| 1 | model_name: "amd/DeepSeek-V4-Flash-NVFP4" |
| 2 | accuracy_threshold: 0.92 |
| 3 | num_questions: 1319 |
| 4 | num_fewshot: 8 |
| 5 | startup_max_wait_seconds: 1800 |
| 6 | server_args: >- |
| 7 | --tensor-parallel-size 8 |
| 8 | --max-model-len 4096 |
| 9 | --trust-remote-code |
| 10 | --kv-cache-dtype fp8 |
| 11 | --tokenizer-mode deepseek_v4 |
| 12 | --compilation-config '{"mode": 3, "cudagraph_mode": "FULL_DECODE_ONLY"}' |
| 13 | env: |
| 14 | VLLM_ROCM_USE_AITER: "1" |
| 15 | VLLM_ROCM_USE_AITER_MOE: "1" |