experimental_experiment.torch_bench._bash_bench_model_runner

class experimental_experiment.torch_bench._bash_bench_model_runner.MakeConfig(**kwargs)

Creates a dictionary where keys are attributes.

class experimental_experiment.torch_bench._bash_bench_model_runner.ModelRunner(model: Any, inputs: Any, kw_inputs: Dict[str, Any] | None, device: str, dtype: dtype, warmup: int, repeat: int, suite: str, autocast: bool = False, wrap_kind: None = None, nvtx: bool = False, model_name: str | None = None, export_options: Dict[str, Any] | None = None)

Wrappers around a model. Makes it easier to load, run inference.

Parameters:

• model – torch model

• inputs – example of inputs

• kw_inputs – example of keyword inputs

• device – device

• dtype – if the model needs to be converted

• warmup – number of iteration to warmup the model

• repeat – number of iteration to repeat the model

• suite – model suite

• wrap_kind – to wrap the model and tuple as much as possible, None is default behavior, ‘nowrap’ to explicit avoid wrapping

• nvtx – enable nvtx events

• model_name – model name

• export_options – additional options when exporting if the default options never work

classmethod allowed_configuration(exporter: str, optimization: str | None = None) → bool

Defines the allowed configurations.

compute_weight_size() → int

Returns the weight size.

dump_std(filename: str)

Dumps some information in the given filename.

export_as(exporter: str, name: str, dynamic: bool, fake_tensor: bool, no_grad: bool, optimization: str, verbose: int, target_opset: int) → Tuple[ModelProto, Dict[str, Any] | None]

Converts a model into onnx.

Parameters:

• exporter – exporter

• name – filename

• dynamic – use dynamic shape

• fake_tensor – use fake_tensor

• no_grad – use no_grad

• optimization – defines the optimizations

• verbose – verbosity

• target_opset – target opset

Returns:

the model proto with or without weights, statistics

get_dynamic_shapes(dynamic: bool = False, input_names: List[str] | None = None) → Dict[str, Any] | Tuple[Any] | List[Any] | None

Returns dynamic shapes specifying the first dimension as dynamic.

Parameters:

• dynamic – make it dynamic or not

• input_names – to overwrite the input names, (not used)

get_input_shapes(dynamic: bool = False, export: bool = False, inputs: Any | None = None) → Any

Returns the input shapes.

Parameters:

• dynamic – dynamic, yes or no?

• inputs – existing inputs or None to use self.inputs

• export – returns the shapes for the inputs used for export

Returns:

new inputs

is_lm() → bool

Returns True if the model is a language model. In that case, the dynamic dimensions with the two first ones. This test relies on the model name.

make_dynamic_inputs()

Creates dynamic inputs based on the static ones by changing the dynamic according to the definition of the dynamic_shapes.

make_export_inputs(dynamic: bool = False, inputs: Tuple[Any, ...] | None = None, kw_inputs: Dict[str, Any] | None = None, int_to_tensor: bool = False) → Tuple[Tuple[Any, ...], Dict[str, Any]]

Creates the new inputs for the benchmarks. torch.export.export() fails when a dimension is dynamic and the value for this dimension is 1. This function expands the input on that dimension to make it 2 if it is 1. These inputs should only be used at export time.

Parameters:

• dynamic – dynamic, yes or no?

• inputs – existing inputs or None to use self.inputs

• int_to_tensor – converts integers or float to tensors

Returns:

new inputs

make_feeds(exporter: str, filename: str | None = None, dynamic: bool = False)

Creates feed inputs.

parameters_dtype() → str

Returns the unique dtypes of all parameters.

class experimental_experiment.torch_bench._bash_bench_model_runner.UseDefaultValue(value)

Defines if the exporter may use the default value.

• FALSE: no default value

• TRUE: there is a default value and the input is not specified

• BOTH: there is a default and one input

class experimental_experiment.torch_bench._bash_bench_model_runner.WrappedModelBase(model)

Wrapper around a module.

forward(*args, **kwargs)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

parameters()

Return an iterator over module parameters.

This is typically passed to an optimizer.

Args:

recurse (bool): if True, then yields parameters of this module

and all submodules. Otherwise, yields only parameters that are direct members of this module.

Yields:

Parameter: module parameter

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for param in model.parameters():
>>>     print(type(param), param.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

class experimental_experiment.torch_bench._bash_bench_model_runner.WrappedModelToTuple(model)

Wrapper around a module, flattens outputs and outputs so that every exporter can use it.

forward(*args, **kwargs)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

experimental_experiment.torch_bench._bash_bench_model_runner.download_retry_decorator(retry: int = 5) → Callable

Decorator function for applying retry logic to a download function.

The wrapped function will be called up to 5 times and raises an exception if the function fails each time. After each unsuccessful attempt, there is a delay before the next attempt, which is increased linearly with the number of tries.

Parameters:

retry – number of times to retry

Usage:

@download_retry_decorator(retry=5)
def download_function(model_name: str):
    # download logic goes here
    # ...

experimental_experiment.torch_bench._bash_bench_model_runner.get_dynamo_stats() → Dict[str, float]

Returns statistics on memory as a dictionary.

experimental_experiment.torch_bench._bash_bench_model_runner.get_peak_memory()

Retuns the memory peak.