Backends#

A backend bundles the pattern lists that drive each pipeline stage. When transform() is called without an explicit patterns argument, the active backend supplies the conversion and fusion patterns automatically. The quantization pipeline similarly reads the backend’s quantized_patterns and smooth_patterns.

Built-in backends#

Backend	Package	Patterns
`"tensorrt"`	`embedl_deploy._internal.tensorrt`	Full set: conversions, fusions, smooth, quantized. Ships as `embedl-deploy-tensorrt`.
`"lattice"`	`embedl_deploy._internal.lattice`	Conversions (fusions, smooth, quantized TBA). Ships as `embedl-deploy-lattice`.

Selecting a backend#

from embedl_deploy.backend import set_backend

set_backend("tensorrt")

After this call, transform(model, (example_input,)) uses the TensorRT conversion and fusion patterns — equivalent to passing patterns=TENSORRT_PATTERNS explicitly.

If only one backend is installed, get_backend() auto-selects it. With multiple backends installed you must call set_backend() before any pipeline function that relies on implicit pattern selection.

API reference#

from embedl_deploy.backend import (
    Backend,
    discover_backends,
    get_backend,
    reset_backends,
    set_backend,
)

set_backend(name)#

Select the active backend by name. Raises ValueError if the name does not match any discovered backend.

get_backend()#

Return the active Backend. If none has been set, auto-discovers backends and selects the only one found. Raises RuntimeError when zero or multiple backends are installed without an explicit set_backend() call.

discover_backends()#

Scan for installed backend packages and return a list of Backend instances. Results are cached after the first call.

reset_backends()#

Clear the cached discovery results and deselect the active backend. Useful in test suites to isolate backend state between tests.

Backend#

@dataclass(frozen=True)
class Backend:
    name: str
    conversion_patterns: Sequence[type[Pattern]] = ()
    fusion_patterns: Sequence[type[Pattern]] = ()
    smooth_patterns: Sequence[type[Pattern]] = ()
    quantized_patterns: Sequence[type[Pattern]] = ()

Each field holds pattern classes for the corresponding pipeline stage. Validation in __post_init__ ensures every pattern’s phase matches the field it is assigned to.