Init the Matmul object.

Calculate the adjustment scales. Args: absorb_to_layer (dict): A dictionary mapping absorb layers to smooth quantized layers.

Wrap a metric fn that returns scores and weights as an eval metric fn. The input metric_fn returns values for the current batch. The wrapper aggr

Get node struct from node name.

Apply activation and weight scales to the model.

Update min/max statistics during calibration. Args: inputs (jnp.ndarray): Input tensor to observe. mask (Optional[jnp

Apply all the transformers to fuse into int8 op.

Evaluates the given model using the evaluation function and arguments provided. Args: model: The model to be evaluated.

Model forward with device auto mapping. Args: model (torch.nn.Module): input model input (torch.tensor): input data Returns:

Generate the default smoothquant config. Returns: the default smoothquant config.

Get the input node name and output node name of the graph_def.

Get module from model by key name. Args: model (torch.nn.Module): original model key (str): module name to be replaced

Compute quantization scale and zero-point for a weight tensor. Args: orig_weight (jnp.ndarray): Weight tensor to analyze. dtype (

Returns whether hpex is available.

Move input to the spevific device.

Set the node's attr which data type is int list.

Set the tensor value to specified attribute field. Args: node (nodedef): the target nodedef object key (string): attr

Set new module into model by key name. Args: model (torch.nn.Module): original model key (str): module name to be replaced

Convert the config to a dictionary. Returns: The config as a dictionary.

Wrapper for inserting bytes features into Example proto.

Get the first input of the node.

Return True when model-free AutoRound is called with a string model reference.

Make the logger dict pretty.

Set batch_size and last_batch.

Validate that a tensor shape matches the sample data.

Compute a SHA-256 hash of the model's config and all variable values.

Get the device name.

Get module with a given op name. Args: model (object): the input model. op_name (str): name of op. Returns: module (

Get the number of dimensions for a Const node. Args: node_def: Const NodeDef. Returns: Number of dimensions for the Const no

Make sure Save&Load format is an Enum object.

Get the graphdef input/output node names. Sometimes, the configuration doesn't specifies the input/output names of the graph, but ten

Get information about the model. Args: model (torch.nn.Module): The model. Returns: List[Tuple[str, Callable

Get the original node name from input string. Args: node_name: input node's name in string Returns: node's n

Generate RTN double quant config set. Returns: List[RTNConfig]: a set of quant config

Return a sorted graphdef object. Sometimes the input graphdef was composed of the random nodedef objects, we reorder the graph to mak

Get the tensor by name. Considering the 'import' scope when model may be imported more then once, handle naming format like both name:0 and n

Check if the package exists in the environment without importing. Args: package_name (str): package name

Receives the parent module to quantize. Replaces its submodules with patched submodules that perform calibration and quantization. Returns the

Print the statistics.

Check whether the mode is for quantization.

Reads input graph file as GraphDef. :param in_graph: input graph file. :param in_graph_is_binary: whether input graph is binary, default True

Save alpha/scale or model weight. Args: save_scale_file: path to save alpha/scale with torch.save. save_state_dict_fi

Generate the configuration mapping based on the model information. Args: config_list (List[BaseConfig], optional): A list of Base

Extracts the values from a const NodeDef as a numpy ndarray. Args: node_def: Const NodeDef that has the values we want to access. Re

Fake quantization.

Load a TensorFlow session for the wrapped model. Args: model: Model path or graph object. **kwargs: Additional sessio

Populate global and local configs based on the whitelist settings.

Add rank suffix to name.

Apply Conv3D Add fusion. Dequantize + Conv3D + BiasAdd + QuantizeV2 Dequantize + Conv3D + Add + QuantizeV2 Dequantize + Conv3

Apply Conv2D BiasAdd AddN fusion. Dequantize + Conv2D + Add + Add + QuantizeV2 Dequantize + Conv2D + AddV2 + Add + QuantizeV2

Apply Conv2D BiasAdd fusion. Dequantize + Conv2D + Biasadd + QuantizeV2 Dequantize + DepthwiseConv2dNative + BiasAdd + QuantizeV2

Get Reshape input nodes.

Compute padding values based on padding type and stride. Args: padding: controls the amount of padding applied to the input.

Quant the model with mx data type. Args: model: torch module Returns: model: fake quantized torch module

Create a quantization specification based on the given configuration. Args: dtype (str): The desired data type for quantization. Valid op

Dump quantizable ops stats of model to user. Args: mode (object): quantization mode. tune_cfg (dict): quantization config

Unified export function for quantization. Args: model (torch.nn.Module): The model to be exported. example_inputs (Tuple[Any]): E

Get prefix and number of blocks. Args: model (torch.nn.Module): input model Returns: block_prefix(str): block_list name in m

Generate the default fake config. Returns: the default fake config.

Create a dequantization function for the specified dtype. Args: dtype (jnp.dtype): Output dtype after dequantization. asymmetric

Get the example input. Args: dataloader (object): calibration dataset. Returns: example_inp (object).

Return ipex version if ipex exists.

Get the device. Args: model (torch.nn.Module): the input model. Returns: device (str): a string.

Get the predefined configuration set. Returns: Dict[torch_utils.ProcessorType, "RTNConfig"]: The configuration of RTN.

Create a quantization function for the specified dtype. Args: dtype (jnp.dtype): Target quantization dtype. asymmetric (bool): Wh

Given a dictionary with 'weight' and 'activation' ScaleMethodConfig objects, returns the corresponding ScaleMethodString if it matches a know

Build session with tf.compat.v1.GraphDef. Args: model (tf.compat.v1.GraphDef): tf.compat.v1.GraphDef object. input_tensor_names (

Check whether intel_extension_for_pytorch is imported.

Check if the scale is quantized.

Checks if the given value is tunable based on the specified tunable type. Args: value (Any): The value to be checked for tunabili

Check if the zero is quantized.

Run the graph that have iterator integrated in the graph. Args: sess (tf.compat.v1.Session): the model sess to run the graph iter

Load a scale file and convert it to module configuration objects. Args: fname (str): Source file path. target_format (type): Desi

replace torch.nn.Module._load_from_state_dict

Convert string to unicode (required in Python 2).

Pack int weight.

The function that processes the inference outputs to prediction