i7 `SSKrSSKJr \"S5(aSSKJr SSKJrJr SSKJr SSK J r J r SS K J r JrJr SS KJrJrJrJr SS KJr SS KJr SS KJr /SQrS\S\S\4SjrS\S\S\4Sjr\R8R:R<R>\R8R:R<R@\R8R:RBR>\R8RDRF/r$S\S\%4Sjr&Sr'SS\S\%S\%S\4Sjjr(g)N)torch_version_at_least2.7.0) constant_fold) GraphModuleNode) PassManager)_fold_conv_bn_qat_fuse_conv_bn_qat)DuplicateDQPassPortNodeMetaForQDQ Quantizer)_disallow_eval_train_fuse_conv_bn__fuse_linear_bn__get_node_name_to_scope)#_convert_to_reference_decomposed_fx)prepare) reference_representation_rewrite) prepare_pt2eprepare_qat_pt2e convert_pt2emodel quantizerreturncX[RRS5 URn[ U5n[ U5 [ U5 URU5nURU5 URU5 [UUSURS9nURRU5 [U5nURRHnUR S:XdMUR"[R$R&R(LdMHUR*Sn[-U[.5(deUR S:Xde[-UR"[05(deUR3UR"5n[5Xa5n[7XR"U5 M URRHnUR S:XdMUR"[R$R&R8LdMHUR*Sn[-U[.5(deUR S:Xde[-UR"[05(deUR3UR"5n [5X5n [7XR"U 5 M U$)aPrepare a model for post training quantization Args: * `model` (torch.fx.GraphModule): a model captured by `torch.export.export` API. * `quantizer`: A backend specific quantizer that conveys how user want the model to be quantized. Tutorial for how to write a quantizer can be found here: https://pytorch.org/tutorials/prototype/pt2e_quantizer.html Return: A GraphModule with observer (based on quantizer annotation), ready for calibration Example:: import torch from torchao.quantization.pt2e.quantize_pt2e import prepare_pt2e from torchao.quantization.pt2e.quantizer import ( XNNPACKQuantizer, get_symmetric_quantization_config, ) class M(torch.nn.Module): def __init__(self) -> None: super().__init__() self.linear = torch.nn.Linear(5, 10) def forward(self, x): return self.linear(x) # initialize a floating point model float_model = M().eval() # define calibration function def calibrate(model, data_loader): model.eval() with torch.no_grad(): for image, target in data_loader: model(image) # Step 1. program capture # NOTE: this API will be updated to torch.export API in the future, but the captured # result shoud mostly stay the same m = torch.export.export(m, *example_inputs).module() # we get a model with aten ops # Step 2. quantization # backend developer will write their own Quantizer and expose methods to allow # users to express how they # want the model to be quantized quantizer = XNNPACKQuantizer().set_global(get_symmetric_quantization_config()) m = prepare_pt2e(m, quantizer) # run calibration # calibrate(m, sample_inference_data) z&torchao.quantization.pt2e.prepare_pt2eFis_qatobs_or_fq_callback call_functionrget_attrr)torch_C_log_api_usage_oncemetarrrtransform_for_annotationannotatevalidaterprepare_obs_or_fq_callbackupdatergraphnodesoptargetops higher_orderscanargs isinstancerstr get_submodulersetattr while_loop) rroriginal_graph_metanode_name_to_scopenodescan_combine_fn_nodescan_combine_fnprepared_scan_combine_fnwhile_loop_body_fn_nodewhile_loop_body_fnprepared_while_loop_body_fns f/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/torchao/quantization/pt2e/quantize_pt2e.pyrr)s0t HH  !IJ**075U  . .u 5E u u  $??  E  JJ)*  'E  !! 77o %$++9O9O9T9T*T#'99Q< 2D99 99'**j8 882993?? ??#112F2M2MNO'3O'O $ E668P Q" !! GG & uyy55@@@&*iil #5t<< <<*--; ;;5< None: super().__init__() self.linear = torch.nn.Linear(5, 10) def forward(self, x): return self.linear(x) # initialize a floating point model float_model = M().eval() # define the training loop for quantization aware training def train_loop(model, train_data): model.train() for image, target in data_loader: ... # Step 1. program capture # NOTE: this API will be updated to torch.export API in the future, but the captured # result shoud mostly stay the same m = torch.export.export(m, *example_inputs).module() # we get a model with aten ops # Step 2. quantization # backend developer will write their own Quantizer and expose methods to allow # users to express how they # want the model to be quantized quantizer = XNNPACKQuantizer().set_global(get_symmetric_quantization_config()) m = prepare_qat_pt2e(m, quantizer) # run quantization aware training train_loop(prepared_model, train_loop) z*torchao.quantization.pt2e.prepare_qat_pt2eTr) r"r#r$r%rr&r'r(r rr)r*r)rrr8r9s rArrsp HH  !MN**07  . .u 5E u ue  $??  E  JJ)*  'E LrBncTURS:H=(a UR[;$)a@If there is any pure ops between get_attr and quantize op they will be const propagated e.g. get_attr(weight) -> transpose -> quantize -> dequantize* (Note: dequantize op is not going to be constant propagated) This filter is added because we don't want to constant fold the things that are not related to quantization r )r-r. _QUANT_OPS)rDs rA_quant_node_constraintrGs! 44? " =qxx:'==rBcSSKJn SSKJn SSKJn SnSnUR 5Hupg[ U[RRRR5(dz[ U[RRRR5(d=[ U[RRRR5(aSn[ Xq5(d"[ Xs5(d[ Xr5(dMSnM U(aU(aS5eU(aU(aS5eU$)Nr) FakeQuantize)AffineQuantizedObserverBase) ObserverBaseFTz2Cannot be prepared using both torch.ao and torchao) 'torchao.quantization.pt2e.fake_quantizerI"torchao.quantization.pt2e.observerrJrK named_modulesr3r"ao quantization fake_quantizeobserver)rtorchao_FakeQuantize#torchao_AffineQuantizedObserverBasetorchao_ObserverBaseis_torch_ao_preparedis_torchao_prepared_ms rA1_is_torchao_prepared_do_not_use_outside_this_filerZsX ##% q%((//==JJ K K!UXX22;;HHII!UXX22;;WWXX#' q / /!22!AA"& && @ &' @ ' rBuse_reference_representation fold_quantizec[RRS5 [U[5(d[ SUS35eUR nURRHnURS:XdMUR[RRRLdMHURSn[U[5(deURS:Xde[UR[ 5(deUR#UR5n[%UUUS9n['XRU5 M URRHnURS:XdMUR[RRR(LdMHURSn[U[5(deURS:Xde[UR[ 5(deUR#UR5n [%U UUS9n ['XRU 5 M [+U5n[-U5n[/[15/5n U "U5R2n[/[55/5n U "U5R2nU(a [7S 5(a[9U[:5 U(a [=U5nUR R?U5 [AU5nU$) aConvert a calibrated/trained model to a quantized model Args: * `model` (torch.fx.GraphModule): calibrated/trained model * `use_reference_representation` (bool): boolean flag to indicate whether to produce referece representation or not * `fold_quantize` (bool): boolean flag for whether fold the quantize op or not Returns: quantized model, either in q/dq representation or reference representation Example:: # prepared_model: the model produced by `prepare_pt2e`/`prepare_qat_pt2e` and calibration/training # `convert_pt2e` produces a quantized model that represents quantized computation with # quantize dequantize ops and fp32 ops by default. # Please refer to # https://pytorch.org/tutorials/prototype/pt2e_quant_ptq_static.html#convert-the-calibrated-model-to-a-quantized-model # for detailed explanation of output quantized model quantized_model = convert_pt2e(prepared_model) z&torchao.quantization.pt2e.convert_pt2ezjUnexpected argument type for `use_reference_representation`, please make sure you intend to pass argument z to convert_pt2er rr!)r[r\rr)!r"r#r$r3bool ValueErrorr%r+r,r-r.r/r0r1r2rr4r5rr6r7rr r r graph_moduler rrrGrr*r) rr[r\r8r:r;r<converted_scan_combine_fnr>r?converted_while_loop_body_fnpms rArrs4 HH  !IJ 2D 9 9 <rtsH 0'"",&:T 9N g gggTJ JJJ\ II""66>> II""66== II""77?? II%%  >d>t>"N*/S S"&SS SrB