When adding new model recipes, ensure they follow the standard structure:
<model-name>/
├── model-cache/
│ ├── model-cache.yaml
│ └── model-download.yaml
├── <framework>/
│ └── <deployment-mode>/
│ ├── deploy.yaml
│ └── perf.yaml (optional)
└── README.md (optional)
The run.sh script expects this exact directory structure and will validate that the directories and files exist before deployment:
- Model directory exists in
recipes/<model>/ - Framework is one of the supported frameworks (vllm, sglang, trtllm)
- Framework directory exists in
recipes/<model>/<framework>/ - Deployment directory exists in
recipes/<model>/<framework>/<deployment>/ - Required files (
deploy.yaml) exist in the deployment directory - If present, performance benchmarks (
perf.yaml) will be automatically executed