| 1 | model_name: "Qwen/Qwen3-4B-FP8" |
| 2 | num_questions: 1319 |
| 3 | num_fewshot: 5 |
| 4 | max_tokens: 256 |
| 5 | use_chat_completions: false |
| 6 | temperature: 1.0 |
| 7 | seed: 42 |
| 8 | |
| 9 | # Reference accuracy from the non-watermarked DSpark regression test. |
| 10 | accuracy_threshold: 0.801 |
| 11 | tolerance: 0.02 |
| 12 | # Measured TP1 acceptance length: 4.115; allow a 0.2 regression. |
| 13 | min_acceptance_length: 3.915 |
| 14 | |
| 15 | env: |
| 16 | VLLM_USE_V2_MODEL_RUNNER: "1" |
| 17 | server_args: >- |
| 18 | --tensor-parallel-size 2 |
| 19 | --max-model-len 4096 |
| 20 | --gpu-memory-utilization 0.92 |
| 21 | --no-enable-prefix-caching |
| 22 | --speculative-config.method dspark |
| 23 | --speculative-config.model deepseek-ai/dspark_qwen3_4b_block7 |
| 24 | --speculative-config.num_speculative_tokens 7 |
| 25 | --speculative-config.draft_sample_method probabilistic |
| 26 | --watermark-config.algorithm dual_key_gumbel |
| 27 | --watermark-config.key 42 |
| 28 | --watermark-config.context_width 4 |
| 29 | --watermark-config.deduplicate_contexts none |