torch_to_nnef.inference_target

Targeted inference engine.

We mainly focus our effort to best support SONOS 'tract' inference engine.

Stricter Khronos NNEF specification mode also exist but is less extensively tested.

InferenceTarget

InferenceTarget(version: T.Union[SemanticVersion, str], check_io: bool = False)

Base abstract class to implement a new inference engine target.

Init InferenceTarget.

Each inference engine is supposed to have at least a version and a way to check output given an input.

has_dynamic_axes `property`

has_dynamic_axes: bool

Define if user request dynamic axes to be in the NNEF graph.

Some inference engines may not support it hence False by default.

post_export

post_export(model: nn.Module, nnef_graph: NGraph, args: T.List[T.Any], exported_filepath: Path, debug_bundle_path: T.Optional[Path] = None)

Get called after NNEF model asset is generated.

This is typically where check_io is effectively applied.

post_trace

post_trace(nnef_graph: NGraph, active_custom_extensions: T.List[str])

Get called just after PyTorch graph is parsed.

pre_trace

pre_trace(model: nn.Module, input_names: T.Optional[T.List[str]], output_names: T.Optional[T.List[str]])

Get called just before PyTorch graph is traced.

(after auto wrapper)

specific_fragments

specific_fragments(model: nn.Module) -> T.Dict[str, str]

Optional custom fragments to pass.

KhronosNNEF

KhronosNNEF(version: T.Union[SemanticVersion, str], check_io: bool = True)

Bases: InferenceTarget

Khronos Specification compliant NNEF asset build.

in case of check_io=True we perform evaluation against nnef_tool nnef to pytorch converter. And access original and reloaded pytorch model provide same outputs

post_export

post_export(model: nn.Module, nnef_graph: NGraph, args: T.List[T.Any], exported_filepath: Path, debug_bundle_path: T.Optional[Path] = None)

Check io via the Torch interpreter of NNEF-Tools.

TractNNEF

TractNNEF(version: T.Union[str, SemanticVersion], feature_flags: T.Optional[T.Set[TractFeatureFlag]] = None, check_io: bool = True, dynamic_axes: T.Optional[T.Dict[str, T.Dict[int, str]]] = None, specific_tract_binary_path: T.Optional[Path] = None, check_io_tolerance: TractCheckTolerance = TractCheckTolerance.APPROXIMATE, specific_properties: T.Optional[T.Dict[str, str]] = None, dump_identity_properties: bool = True, force_attention_inner_in_f32: bool = False, force_linear_accumulation_in_f32: bool = False, force_norm_in_f32: bool = False, reify_sdpa_operator: bool = False, upsample_with_debox: bool = False)

Bases: InferenceTarget

Tract NNEF inference target.

Init.

Parameters:

Name	Type	Description	Default
`version`	`Union[str, SemanticVersion]`	tract version targeted for export	required
`feature_flags`	`Optional[Set[TractFeatureFlag]]`	set of possibly added feature flags from tract (for example complex numbers)	`None`
`check_io`	`bool`	check between tract cli and Pytorch original model that given provided input, output is similar	`True`
`dynamic_axes`	`Optional[Dict[str, Dict[int, str]]]`	Optional specification of dynamic dimension By default the exported model will have the shapes of all input and output tensors set to exactly match those given in args. To specify axes of tensors as dynamic (i.e. known only at runtime) set dynamic_axes to a dict with schema: KEY (str): an input or output name. Each name must also be provided in input_names or output_names. VALUE (dict or list): If a dict, keys are axis indices and values are axis names. If a list, each element is an axis index.	`None`
`specific_tract_binary_path`	`Optional[Path]`	filepath of tract cli in case of custom non released version of tract (for testing purpose)	`None`
`check_io_tolerance`	`TractCheckTolerance`	TractCheckTolerance level of difference tolerance between original output values and those generated by tract (those are defined tract levels)	`APPROXIMATE`
`specific_properties`	`Optional[Dict[str, str]]`	custom tract_properties you wish to add inside NNEF asset (will be parsed by tract as metadata)	`None`
`dump_identity_properties`	`bool`	add tract_properties relative to user identity (host, username, OS...), helpfull for debug	`True`
`force_attention_inner_in_f32`	`bool`	`control if attention should be forced as f32 inside (even if inputs are all f16), usefull for unstable networks like qwen2.5`	`False`
`force_linear_accumulation_in_f32`	`bool`	usefull for f16 models to ensure that output of f16. f16 matmul become f32 accumulators.	`False`
`force_norm_in_f32`	`bool`	ensure that all normalization layers are in f32 whatever the original PyTorch modeling.	`False`
`reify_sdpa_operator`	`bool`	enable the conversion of scaled_dot_product_attention as a tract operator (intead of a NNEF fragment). Experimental feature.	`False`
`upsample_with_debox`	`bool`	use debox upsample operator instead of deconvolution. This should be faster. (if tract version support it). Experimental feature.	`False`

post_export

post_export(model: nn.Module, nnef_graph: NGraph, args: T.List[T.Any], exported_filepath: Path, debug_bundle_path: T.Optional[Path] = None)

Perform check io and build debug bundle if fail.

post_trace

post_trace(nnef_graph, active_custom_extensions)

Add dynamic axes in the NNEF graph.

pre_trace

pre_trace(model: nn.Module, input_names: T.Optional[T.List[str]], output_names: T.Optional[T.List[str]])

Check dynamic_axes are correctly formated.

specific_fragments

specific_fragments(model: nn.Module) -> T.Dict[str, str]

Optional custom fragments to pass.

torch_to_nnef.inference_target