SkyPilot Config
Configs included here are references for running AIPerf baselines with violin-tuner against the selected Qwen-Bailian TraceA/B1 window.
| Model | Role | Model path | GPU | AIPerf dataset | Objective | Config |
|---|---|---|---|---|---|---|
| Qwen3-8B | Primary Qwen-family baseline for Qwen-Bailian replay | Qwen/Qwen3-8B | 1 x H100 | bailian_trace | e2e_output_token_throughput | qwen3_8b_qwen_bailian_traceA_B1_baseline.yaml |
| gpt-oss-120b | Cross-model serving comparison, not the primary fidelity baseline | openai/gpt-oss-120b | 1 x H100 | bailian_trace | e2e_output_token_throughput | gpt_oss_120b_qwen_bailian_traceA_B1_baseline.yaml |
Suggested Qwen Baseline Launch
cd /path/to/violin-tuner
python skypilot_launcher/launch.py benchmark_configs/skypilot/baselines/local/qwen3_8b_qwen_bailian_traceA_B1_baseline.yaml --template benchmark_configs/skypilot/templates/baseline_exec_in_image.yaml.j2 --infra nebius/eu-north1The Qwen3-8B config is the primary fidelity baseline for Qwen-Bailian. The GPT-OSS config is included only as a cross-model comparison reference.