h(>SSKrSSKrSSKrSSKJr \R "\5rSr\S:XGao\"5r \ R(aD\ R(a3\ R(a"\RS5 \R"5 \ R(d3\ R(a"\RS5 \R"5 \R!S\ R"5 \R!S \ R$5 \"\ R"\ R$\ R\ R\ R\ R&\ R(\ R*\ R,\ R\ R.\ R0\ R25 gg) N)quant_pre_processcH[R"SS9nURSSSS9 URSSSS9 URS [S S S 9 URS [S SS 9 URS[S SS 9 URSSSS S9 URSS[SS9 URSSSS S9 URSS[SS9 URSSSS S9 URS S!SS S9 URS"S#SS$9 URS%S&[S'S9 UR 5$)(NaModel 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) descriptionz--inputTzPath to the input model file)requiredhelpz--outputzPath to the output model filez--skip_optimizationFzSkip 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.)typedefaultrz--skip_onnx_shapezSkip 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_shapezSkip 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 happens store_true)ractionr z --int_maxzGmaximum value for integer to be treated as boundless for ops like slicei)rr 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: detailedrz--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)rr z--external_data_size_thresholdz$The size threshold for external datai)argparseArgumentParser add_argumentboolint parse_args)parsers ]/var/www/fran/franai/venv/lib/python3.13/site-packages/onnxruntime/quantization/preprocess.pyparse_argumentsrs  $ $F(  D7UV  T8WX  1   '   4   I    V     J    W    ! 4   # 7   " :  ( 3      __main__z9Skipping all three steps, nothing to be done. Quitting...z:ORT model optimization does not support external data yet!zinput model: %szoutput model: %s)r loggingsysshape_inferencer getLogger__name__loggerrargsskip_optimizationskip_onnx_shapeskip_symbolic_shapeerrorexitsave_as_external_datainfoinputoutput auto_mergeint_maxguess_output_rankverboseall_tensors_to_one_fileexternal_data_locationexternal_data_size_thresholdrrr0s0 .   8 $`F z  D $"6"64;S;S PQ   " "(B(B QR   KK!4::. KK"DKK0            "" $$ ## ))r