| 1 | model_name: "nvidia/GLM-5.2-NVFP4" |
| 2 | accuracy_threshold: 0.90 |
| 3 | num_questions: 1319 |
| 4 | num_fewshot: 5 |
| 5 | max_concurrency: 100 |
| 6 | # Preserve KV-cache headroom after CUDA graph memory profiling and bound the |
| 7 | # DCP-aware FlashInfer workspace to 2.39 GiB at the 4,096-token scheduler cap. |
| 8 | server_args: >- |
| 9 | --compilation-config '{"cudagraph_mode":"PIECEWISE"}' |
| 10 | --gpu-memory-utilization 0.96 |
| 11 | --max-model-len 4096 |
| 12 | --max-num-batched-tokens 4096 |
| 13 | --safetensors-load-strategy prefetch |
| 14 | --moe-backend flashinfer_cutlass |
| 15 | --prefill-context-parallel-size 4 |
| 16 | --decode-context-parallel-size 4 |
| 17 | --dcp-comm-backend ag_rs |
| 18 | --enable-expert-parallel |
| 19 | --kv-cache-dtype fp8 |
| 20 | env: |
| 21 | PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True" |
| 22 | VLLM_LOGGING_LEVEL: "DEBUG" |
| 23 | VLLM_USE_V2_MODEL_RUNNER: "1" |