| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | model_name: "Qwen/Qwen3.8-Flash-Next-FP8" |
| 4 | datasets: |
| 5 | gsm8k: |
| 6 | metric_threshold: 0.98 |
| 7 | tolerance: 0.02 |
| 8 | aime25: |
| 9 | metric_threshold: 0.90 |
| 10 | tolerance: 0.05 |
| 11 | eval_batch_size: 32 |
| 12 | startup_max_wait_seconds: 1800 |
| 13 | generation_config: |
| 14 | seed: 1236 |
| 15 | do_sample: true |
| 16 | temperature: 0.7 |
| 17 | top_p: 0.8 |
| 18 | top_k: 20 |
| 19 | repetition_penalty: 1.0 |
| 20 | presence_penalty: 1.5 |
| 21 | max_tokens: 32768 |
| 22 | stream: true |
| 23 | extra_body: |
| 24 | chat_template_kwargs: |
| 25 | enable_thinking: true |
| 26 | server_args: "--tensor-parallel-size 4" |