Installation#

Requirements#

  • Python ≥ 3.10

  • PyTorch (the only runtime dependency)

Install#

Wheels are published to the Embedl package index at https://deploy.embedl.com/downloads/, a PEP 503 simple index, not to PyPI. Access needs an Embedl Hub account: put your hub API key (https://hub.embedl.com/profile) in the index URL as the password.

Install the latest core package and TensorRT backend from the index and let pip fill in dependencies such as torch from PyPI:

pip install --index-url https://hub:<API_KEY>@deploy.embedl.com/downloads/ --upgrade --no-deps embedl-deploy embedl-deploy-tensorrt && pip install "embedl-deploy[tensorrt]"

The first half resolves the Embedl packages against the Embedl index only; the second finds them already installed and fetches the rest from PyPI. --upgrade moves an existing installation to the newest release. To install a specific one, replace the package names with embedl-deploy==X.Y.Z embedl-deploy-tensorrt==X.Y.Z; releases are listed at https://deploy.embedl.com/downloads/embedl-deploy/. The backend depends on the core at exactly its own version, so the second command rejects a mismatched pair instead of leaving broken internals.

Variations#

--extra-index-url lets pip search PyPI and the Embedl index together, so everything installs in one pip install without --no-deps:

pip install --extra-index-url https://hub:<API_KEY>@deploy.embedl.com/downloads/ "embedl-deploy[tensorrt]"

To keep the key out of shell history and process arguments, put it in ~/.netrc and drop hub:<API_KEY>@ from the URL; pip reads netrc automatically:

machine deploy.embedl.com
login hub
password <API_KEY>

Extras#

  • search pulls in Optuna for the quantization search tooling. Add it to the second half of the install command: pip install "embedl-deploy[tensorrt,search]".

Backends#

A compilation backend is a companion wheel installed alongside the base package. The TensorRT backend, for NVIDIA GPUs, is installed by the command above. Drop embedl-deploy-tensorrt and the [tensorrt] extra to install the base package alone, which is enough to build models in PyTorch and export to ONNX.

See Backends for how the active backend is selected at runtime.

For model compilation, you may need additional tools such as TensorRT (trtexec), but these are not required to create deployment-ready models in PyTorch and export them to ONNX.