neural_compressor.adaptor.torch_utils.util

Util Class and Functions.

Module Contents

Functions

move_input_device(input[, device])	Auto mapping input to device for all kinds of format.
forward_wrapper(model, input)	Model forward with device auto mapping.
get_embedding_contiguous(model)	This is a helper function for nn.Embedding, and it will get input contiguous.
is_fused_module(module)	This is a helper function for _propagate_qconfig_helper to detect if this module is fused.
collate_torch_preds(results)	Fetch collated results.
input2tuple(input)	This is a helper function to converting a inputting dict values or a list to a tuple.
append_attr(fx_model, model[, fx_white_list])	This is a helper method to append attributes for the symbolic traced model.
generate_activation_observer(scheme, algorithm[, ...])	This is a helper method to generate an activation observer.
check_cfg_and_qconfig(tune_cfg, cfgs, ...[, smooth_quant])	Check configs and quantization configs.
paser_cfgs(cfgs)	Parse configs.
get_quantizable_ops_from_cfgs(ops_name, ...)	Get quantizable ops from configs, combine fused ops as one op.
update_sq_scale(ipex_config_path, smoothquant_scale_info)	Update ipex_config.json with smoothquant scale info generated by our algorithm.
auto_copy(module)	Get an IPEX prepared model and return a fp32 model.
fetch_module(model, op_name)	Get module with a given op name.
set_module(model, op_name, new_module)	Set module with a given op name.
simple_inference(model, input)	Record model output tensor.
get_example_input(dataloader[, i])	Get the example input.
get_fallback_order(adaptor, fp32_model, dataloader, ...)	Get the fall back order for strategy.
get_mse_order_per_fp32(adaptor, model, example_inp, ...)	This is a helper method to check the mse influence to last module after QDQ(quant/dequant).
get_mse_order_per_int8(adaptor, fp32_model, ...)	This is a helper method to check the mse influence to last module after QDQ(quant/dequant).
get_torch_version()	Get torch version.
match_datatype_pattern(datatype[, pattern])	Check the datatype pattern.
calculate_quant_min_max(unsigned, num_bits)	Calculate the qmin and qmax according to the datatype.
get_depth(→ int)	Query the depth of the dict.
get_dict_at_depth(d, target_depth, result[, depth])	Get all sub-dicts that are at a specified depth in a nested dict.
get_element_under_depth(d, ops_lst)	Get all values in a nested dict.
get_op_type_by_name(op_name, quantizable_ops)	Get op type by op name.
collect_weight_info(model, q_config)	Collect weight info from q_config for dumping into qconfig.json.
get_module_input_output(model[, module_hook_config, ...])	A help function to get input and output tensor of modules in module_name_list.
get_absorb_layers(model, example_inputs[, ...])	Get absorb_to_layer and no_absorb_layer.
get_block_prefix(model)	Get prefix and number of blocks.
calibration(model[, dataloader, n_samples, calib_func])	Calibration with dataloader or calib_func.
get_hidden_states(model[, dataloader, n_samples, ...])	Get the input args and kwargs of first block.

neural_compressor.adaptor.torch_utils.util.move_input_device(input, device='cpu')

Auto mapping input to device for all kinds of format.

Parameters:

• input (torch.tensor) – input data

• device (str, optional) – target device. Defaults to “cpu”.

Returns:

input data on target device

Return type:

input (torch.tensor)

neural_compressor.adaptor.torch_utils.util.forward_wrapper(model, input)

Model forward with device auto mapping.

Parameters:

• model (torch.nn.Module) – input model

• input (torch.tensor) – input data

Returns:

output data

Return type:

output

neural_compressor.adaptor.torch_utils.util.get_embedding_contiguous(model)

This is a helper function for nn.Embedding, and it will get input contiguous.

Parameters:

model (object) – the input model

Returns:

None

neural_compressor.adaptor.torch_utils.util.is_fused_module(module)

This is a helper function for _propagate_qconfig_helper to detect if this module is fused.

Parameters:

module (object) – the input module

Returns:

is fused or not

Return type:

(bool)

neural_compressor.adaptor.torch_utils.util.collate_torch_preds(results)

Fetch collated results.

Parameters:

result (list) – input result

Returns:

collated results

Return type:

collate_results (list)

neural_compressor.adaptor.torch_utils.util.input2tuple(input)

This is a helper function to converting a inputting dict values or a list to a tuple.

Parameters:

input (list or dict) –

Returns:

A tuple.

neural_compressor.adaptor.torch_utils.util.append_attr(fx_model, model, fx_white_list=[])

This is a helper method to append attributes for the symbolic traced model.

Parameters:

• fx_model (torch.fx.GraphModule) – The symbolic traced model.

• model (torch.nn.Module) – The original model.

Returns:

The symbolic traced model with additional attributes.

Return type:

fx_model (dir)

neural_compressor.adaptor.torch_utils.util.generate_activation_observer(scheme, algorithm, smooth_quant=False, smooth_quant_enable=False)

This is a helper method to generate an activation observer.

Parameters:

• scheme (str) – Quantization scheme to be used.

• algorithm (str) – What algorithm for computing the quantization parameters based on.

Returns:

An observer.

neural_compressor.adaptor.torch_utils.util.check_cfg_and_qconfig(tune_cfg, cfgs, op_infos_from_cfgs, output_tensor_ids_op_name, smooth_quant=False)

Check configs and quantization configs.

Parameters:

• tune_cfg (dict) – dictionary of quantization configuration.

• cfgs (dict) – the input configs.

• op_infos_from_cfgs (dict) – op infos from configs.

• output_tensor_ids_op_name (dict) – dictionary of output tensor op names.

Returns:

cfgs (dict).

neural_compressor.adaptor.torch_utils.util.paser_cfgs(cfgs)

Parse configs.

Parameters:

cfgs (dict) – the input configs.

Returns:

list of op names. tune_cfg (dict): dictionary of quantization configuration. op_infos_from_cfgs (dict): op infos from configs. output_tensor_ids_op_name (dict): dictionary of output tensor op names.

Return type:

ops_name (list)

neural_compressor.adaptor.torch_utils.util.get_quantizable_ops_from_cfgs(ops_name, op_infos_from_cfgs, input_tensor_ids_op_name)

Get quantizable ops from configs, combine fused ops as one op.

Parameters:

• ops_name (list) – list of op names.

• op_infos_from_cfgs (dict) – op infos from configs.

• input_tensor_ids_op_name (dict) – dictionary of input tensor op names.

Returns:

cfgs (dict).

neural_compressor.adaptor.torch_utils.util.update_sq_scale(ipex_config_path, smoothquant_scale_info)

Update ipex_config.json with smoothquant scale info generated by our algorithm.

Parameters:

• ipex_config_path (str) – a path to temporary ipex_config.json file.

• smoothquant_scale_info (dict) – a dict contains smoothquant scale info.

neural_compressor.adaptor.torch_utils.util.auto_copy(module)

Get an IPEX prepared model and return a fp32 model.

Parameters:

module (object) – IPEX prepared model.

Returns:

fp32 model.

neural_compressor.adaptor.torch_utils.util.fetch_module(model, op_name)

Get module with a given op name.

Parameters:

• model (object) – the input model.

• op_name (str) – name of op.

Returns:

module (object).

neural_compressor.adaptor.torch_utils.util.set_module(model, op_name, new_module)

Set module with a given op name.

Parameters:

• model (object) – the input model.

• op_name (str) – name of op.

• new_module (object) – the input model.

Returns:

module (object).

neural_compressor.adaptor.torch_utils.util.simple_inference(model, input)

Record model output tensor.

Parameters:

• model (object) – the input model.

• input (object) –

Returns:

output (object).

neural_compressor.adaptor.torch_utils.util.get_example_input(dataloader, i=1)

Get the example input.

Parameters:

dataloader (object) – calibration dataset.

Returns:

example_inp (object).

neural_compressor.adaptor.torch_utils.util.get_fallback_order(adaptor, fp32_model, dataloader, tune_cfg, confidence_batches, fallback=False, requantize_cfgs=None)

Get the fall back order for strategy.

Parameters:

• fp32_model (object) – the input model.

• dataloader (torch.utils.data.DataLoader) – The calibration dataloader.

• tune_cfg (dict) – dictionary of quantization configuration.

• confidence_batches (int) – number of confidence batches.

• fallback (bool) – if the order is fallback.

Returns:

The fallback order for strategy.

Return type:

ordered_ops (dict/list)

neural_compressor.adaptor.torch_utils.util.get_mse_order_per_fp32(adaptor, model, example_inp, tune_cfg)

This is a helper method to check the mse influence to last module after QDQ(quant/dequant).

Parameters:

• model (torch.fx.GraphModule/torch.nn.Module) – A torch model.

• example_inp (object) – example inputs.

• tune_cfg (dict) – dictionary of quantization configuration.

Returns:

The fallback order for strategy.

Return type:

fallback_order (dict/list)

neural_compressor.adaptor.torch_utils.util.get_mse_order_per_int8(adaptor, fp32_model, example_input, tune_cfg)

This is a helper method to check the mse influence to last module after QDQ(quant/dequant).

Parameters:

• model (torch.fx.GraphModule/torch.nn.Module) – A torch model.

• example_inp (object) – example inputs.

• tune_cfg (dict) – dictionary of quantization configuration.

Returns:

The fallback order for strategy.

Return type:

fallback_order (dict/list)

neural_compressor.adaptor.torch_utils.util.get_torch_version()

Get torch version.

neural_compressor.adaptor.torch_utils.util.match_datatype_pattern(datatype, pattern=None)

Check the datatype pattern.

neural_compressor.adaptor.torch_utils.util.calculate_quant_min_max(unsigned, num_bits)

Calculate the qmin and qmax according to the datatype.

neural_compressor.adaptor.torch_utils.util.get_depth(d) → int

Query the depth of the dict.

neural_compressor.adaptor.torch_utils.util.get_dict_at_depth(d, target_depth, result, depth=0)

Get all sub-dicts that are at a specified depth in a nested dict.

neural_compressor.adaptor.torch_utils.util.get_element_under_depth(d, ops_lst)

Get all values in a nested dict.

neural_compressor.adaptor.torch_utils.util.get_op_type_by_name(op_name, quantizable_ops)

Get op type by op name.

neural_compressor.adaptor.torch_utils.util.collect_weight_info(model, q_config)

Collect weight info from q_config for dumping into qconfig.json.

qconfig.json example: ``` {

‘fc’: {

‘bits’: 4, ‘group_size’: 128, ‘scheme’: ‘asym’, ‘algorithm’: ‘RTN’

}

param q_config:

quantization configure

type q_config:

_type_

neural_compressor.adaptor.torch_utils.util.get_module_input_output(model, module_hook_config={}, dataloader=None, iters=-1, calib_func=None, input_func=None, output_func=None)

A help function to get input and output tensor of modules in module_name_list.

Parameters:

• model – torch model.

• module_hook_config (dict, optional) –

  required module name for input/output. Defaults to {}. For example:

  module_hook_config = {

  ‘fc1’: [‘output’], ‘fc2’: [‘input’, ‘output’]

  }

• dataloader – dataloader for model input.

• iters – iterations for inference.

• calib_func – a custom inference function to replace dataloader and iters.

• input_func – preprocess input for less memory usage

• output_func – preprocess output for less memory usage

Returns:

recorded input_values, output_values.

for example:

{‘fc1’:

{‘input’: [], ‘output’: []},

}

Return type:

total_values

neural_compressor.adaptor.torch_utils.util.get_absorb_layers(model, example_inputs, supported_layers=['Linear'], folding=False)

Get absorb_to_layer and no_absorb_layer.

Parameters:

• model (torch.nn.Module) – input model

• example_inputs – example_inputs

• supported_layers (list, optional) – supported_layers. Defaults to [‘Linear’].

• folding (bool, optional) – whether allow self-absorption. Defaults to False.

Returns:

dict of absorb_to_layer. eg. {absorb, [absorbed_1, xx]} no_absorb_layers: list of no_absorb_layers

Return type:

absorb_to_layer

neural_compressor.adaptor.torch_utils.util.get_block_prefix(model)

Get prefix and number of blocks.

Parameters:

model (torch.nn.Module) – input model

Returns:

block_list name in model block_num(int): number of block in block_list

Return type:

block_prefix(str)

neural_compressor.adaptor.torch_utils.util.calibration(model, dataloader=None, n_samples=128, calib_func=None)

Calibration with dataloader or calib_func.

Parameters:

• model (torch.nn.Module) – input model

• dataloader – dataloader. Defaults to None.

• n_samples (int, optional) – n_samples. Defaults to 128.

• calib_func – calib_func. Defaults to None.

neural_compressor.adaptor.torch_utils.util.get_hidden_states(model, dataloader=None, n_samples=128, calib_func=None)

Get the input args and kwargs of first block.

Parameters:

• model (torch.nn.Module) – input model

• dataloader (dataloader, optional) – input dataloader. Defaults to None.

• n_samples (int, optional) – number samples from dataloader. Defaults to 128.

• calib_func (func, optional) – a calib func to replace dataloader. Defaults to None.

Raises:

ValueError – to avoid inference of rest parts in model

Returns:

a list of input args of each batch total_block_kwargs(list): a list of input kwargs of each batch

Return type:

total_block_args(list)