quark.torch.quantization.model_transformation

Module Contents

Functions

quark.torch.quantization.model_transformation.process_model_transformation(model: torch.nn.Module, config: quark.torch.quantization.config.config.Config) → torch.nn.Module

Replaces modules to be quantized by their quantized equivalent (e.g. nn.Linear by QuantLinear), based on the provided global config.

quark.torch.quantization.model_transformation.setup_config_per_layer(config: quark.torch.quantization.config.config.Config, named_modules: Dict[str, torch.nn.Module], module_configs: Dict[str, quark.torch.quantization.config.config.QuantizationConfig]) → None

Retrieves the QuantizationConfig used for each layer, based on the config’s global_quant_config, layer_quant_config and layer_type_quant_config.

quark.torch.quantization.model_transformation.in_place_replace_layer(model: torch.nn.Module, config: quark.torch.quantization.config.config.Config, named_modules: Dict[str, torch.nn.Module], module_configs: Dict[str, quark.torch.quantization.config.config.QuantizationConfig]) → None

Replaces nn.Linear, nn.Conv2d, etc. marked for quantization in module_configs by their quantized module equivalent.