| 1 | model_name: "mgoin/Kimi-K3-pruned75" |
| 2 | accuracy_threshold: 0.33 |
| 3 | min_acceptance_length: 4.8 |
| 4 | num_questions: 1319 |
| 5 | num_fewshot: 5 |
| 6 | startup_max_wait_seconds: 1800 |
| 7 | server_args: >- |
| 8 | --trust-remote-code |
| 9 | --gpu-memory-utilization 0.90 |
| 10 | --tensor-parallel-size 4 |
| 11 | --load-format fastsafetensors |
| 12 | --max-model-len 1048576 |
| 13 | --max-num-batched-tokens 16384 |
| 14 | --max-num-seqs 128 |
| 15 | --kv-cache-dtype fp8 |
| 16 | --moe-backend aiter |
| 17 | --attention-config '{"backend":"ROCM_AITER_MLA","mla_prefill_backend":"ROCM_AITER_FA"}' |
| 18 | --enable-prefix-caching |
| 19 | --enable-auto-tool-choice |
| 20 | --tool-call-parser kimi_k3 |
| 21 | --reasoning-parser kimi_k3 |
| 22 | --speculative-config '{"method":"dspark", |
| 23 | "model":"RedHatAI/Kimi-K3-speculator.dspark", |
| 24 | "attention_backend":"ROCM_AITER_FA","num_speculative_tokens":8, |
| 25 | "draft_sample_method":"probabilistic", |
| 26 | "rejection_sample_method":"block"}' |
| 27 | env: |
| 28 | VLLM_ROCM_USE_AITER: "1" |