torch_tensorrt.ts.ptq¶
These components are legacy quantization utilities designed to work with the TorchScript Frontend. They have been replaced by the TensorRT Model Optimizer toolkit which can be used with the dynamo frontend:
Classes¶
- class torch_tensorrt.ts.ptq.DataLoaderCalibrator(*args: Any, **kwargs: Any)[source]¶
Constructs a calibrator class in TensorRT and uses pytorch dataloader to load/preprocess data which is passed during calibration.
- Parameters
dataloader (torch.utils.data.DataLoader) – an instance of pytorch dataloader which iterates through a given dataset.
algo_type (CalibrationAlgo) – choice of calibration algorithm.
cache_file (str) – path to cache file.
use_cache (bool) – flag which enables usage of pre-existing cache.
device (Device) – device on which calibration data is copied to.
- class torch_tensorrt.ts.ptq.CacheCalibrator(*args: Any, **kwargs: Any)[source]¶
Constructs a calibrator class in TensorRT which directly uses pre-existing cache file for calibration.
- Parameters
cache_file (str) – path to cache file.
algo_type (CalibrationAlgo) – choice of calibration algorithm.
Enums¶
- class torch_tensorrt.ts.ptq.CalibrationAlgo(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)[source]¶
- ENTROPY_CALIBRATION = <CalibrationAlgo.ENTROPY_CALIBRATION: 1>¶
- ENTROPY_CALIBRATION_2 = <CalibrationAlgo.ENTROPY_CALIBRATION_2: 2>¶
- LEGACY_CALIBRATION = <CalibrationAlgo.LEGACY_CALIBRATION: 0>¶
- MINMAX_CALIBRATION = <CalibrationAlgo.MINMAX_CALIBRATION: 3>¶
