o
    °ƒðj(  ã                   @   sä   d dl Z d dlZd dlZddlmZ e e¡Zdd„ Zedkrpeƒ Z	e	j
r4e	jr4e	jr4e d¡ e ¡  e	j
sCe	jrCe d¡ e ¡  e d	e	j¡ e d
e	j¡ ee	je	je	j
e	je	je	je	je	je	je	je	je	je	jƒ dS dS )é    Né   )Úquant_pre_processc                  C   sø   t jdd�} | jdddd� | jdddd� | jd	td
dd� | jdtd
dd� | jdtd
dd� | jdddd
d� | jddtdd� | jdddd
d� | jddtdd� | jdddd
d� | jd d!dd
d� | jd"d#d d$� | jd%d&td'd� |  ¡ S )(NaÜ  Model optimizer and shape inferencer, in preparation for quantization,
Consists of three optional steps:
1. Symbolic shape inference (best for transformer models).
2. Model optimization.
3. ONNX shape inference.

Model quantization with QDQ format, i.e. inserting QuantizeLinear/DeQuantizeLinear on
the tensor, requires tensor shape information to perform its best. Currently, shape inferencing
works best with optimized model. As a result, it is highly recommended to run quantization
on optimized model with shape information. This is the tool for optimization and shape
inferencing.

Essentially this tool performs the following three (skippable) steps:

1. Symbolic shape inference.
2. Model optimization
3. ONNX shape inference)Údescriptionz--inputTzPath to the input model file)ÚrequiredÚhelpz--outputzPath to the output model filez--skip_optimizationFz°Skip model optimization step if true. It's a known issue that ORT optimization has difficulty with model size greater than 2GB, rerun with this option to get around this issue.)ÚtypeÚdefaultr   z--skip_onnx_shapezåSkip ONNX shape inference. Symbolic shape inference is most effective with transformer based models. Skipping all shape inferences may reduce the effectiveness of quantization, as a tensor with unknown shape can not be quantized.z--skip_symbolic_shapezéSkip symbolic shape inference. Symbolic shape inference is most effective with transformer based models. Skipping all shape inferences may reduce the effectiveness of quantization, as a tensor with unknown shape can not be quantized.z--auto_mergez:Automatically merge symbolic dims when confliction happensÚ
store_true)r   Úactionr   z	--int_maxzGmaximum value for integer to be treated as boundless for ops like sliceiÿÿÿ)r   r   r   z--guess_output_rankz;guess output rank to be the same as input 0 for unknown opsz	--verbosezHPrints detailed logs of inference, 0: turn off, 1: warnings, 3: detailedr   z--save_as_external_dataz%Saving an ONNX model to external dataz--all_tensors_to_one_filez(Saving all the external data to one filez--external_data_locationz+The file location to save the external file)r   r   z--external_data_size_thresholdz$The size threshold for external datai   )ÚargparseÚArgumentParserÚadd_argumentÚboolÚintÚ
parse_args)Úparser© r   ú¥/tmp/claude-1000/-var-www-html-academy/3f42efd1-c2b7-4d08-9b05-30d86bb35479/scratchpad/piper/venv/lib/python3.10/site-packages/onnxruntime/quantization/preprocess.pyÚparse_arguments   sŽ   ÿüü	ü	üüüüüüýür   Ú__main__z9Skipping all three steps, nothing to be done. Quitting...z:ORT model optimization does not support external data yet!zinput model: %szoutput model: %s)r   ÚloggingÚsysÚshape_inferencer   Ú	getLoggerÚ__name__Úloggerr   ÚargsÚskip_optimizationÚskip_onnx_shapeÚskip_symbolic_shapeÚerrorÚexitÚsave_as_external_dataÚinfoÚinputÚoutputÚ
auto_mergeÚint_maxÚguess_output_rankÚverboseÚall_tensors_to_one_fileÚexternal_data_locationÚexternal_data_size_thresholdr   r   r   r   Ú<module>   s@   
c

óô