sg)$ ddlZddlmZddlmZddlmZddlmZm Z m Z m Z m Z m Z mZddlmZmZddlmZdd lmZdd lmZmZdd lmZdd lmZmZmZdd lm Z gdZ!dede defdZ"dede defdZ#ejHjJjLjNejHjJjLjPejHjJjRjNgZ*dede+fdZ, ddede+de+defdZ-y)N) constant_fold)DuplicateDQPass)PortNodeMetaForQDQ)DerivedQuantizationSpecFixedQParamsQuantizationSpecQuantizationAnnotationQuantizationSpecQuantizationSpecBase QuantizerSharedQuantizationSpec) GraphModuleNode) PassManager)prepare)_fold_conv_bn_qat_fuse_conv_bn_qat) reference_representation_rewrite)_disallow_eval_train_fuse_conv_bn__get_node_name_to_scope)#_convert_to_reference_decomposed_fx) prepare_pt2eprepare_qat_pt2e convert_pt2emodel quantizerreturncVtjjd|j}t |}t ||j |}|j||j|t||d}|jj|t|}|S)aZPrepare a model for post training quantization Args: * `model` (torch.fx.GraphModule): a model captured by `torch.export` API in the short term we are using `torch._export.capture_pre_autograd_graph`, in the long term we'll migrate to some `torch.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 torch.ao.quantization.quantize_pt2e import prepare_pt2e from torch._export import capture_pre_autograd_graph from torch.ao.quantization.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 = capture_pre_autograd_graph(m, *example_inputs) # 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+quantization_api.quantize_pt2e.prepare_pt2eFis_qat) torch_C_log_api_usage_oncemetarrtransform_for_annotationannotatevalidaterupdaterrroriginal_graph_metanode_name_to_scopes V/var/www/html/venv/lib/python3.12/site-packages/torch/ao/quantization/quantize_pt2e.pyrrsz HH  !NO**075  . .u 5E u u E-e 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 = capture_pre_autograd_graph(m, *example_inputs) # 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/quantization_api.quantize_pt2e.prepare_qat_pt2eTr ) r"r#r$r%rr&r'r(rrr)rr*s r-rrlsr HH  !RS**07  . .u 5E u ue E-d ;E JJ)*  'E Lr.ncH|jdk(xr|jtvS)aTIf there is any pure ops between get_attr and quantize op they will be const propagated e.g. get_attr(weight) -> transpose -> quantize -> dequantize* (Note: dequantize op is not going to be constant propagated) This filter is added because we don't want to constant fold the things that are not related to quantization call_function)optarget _QUANT_OPS)r0s r-_quant_node_constraintr6s! 44? " =qxx:'==r.use_reference_representation fold_quantizectjjdt|tst d|d|j }t|}t|}ttg}||j}ttg}||j}|rt|t|r t|}|j j!|t#|}|S)aConvert a calibrated/trained model to a quantized model Args: * `model` (torch.fx.GraphModule): calibrated/trained model * `use_reference_representation` (bool): boolean flag to indicate whether to produce referece representation or not * `fold_quantize` (bool): boolean flag for whether fold the quantize op or not Returns: quantized model, either in q/dq representation or reference representation Example:: # prepared_model: the model produced by `prepare_pt2e`/`prepare_qat_pt2e` and calibration/training # `convert_pt2e` produces a quantized model that represents quantized computation with # quantize dequantize ops and fp32 ops by default. # Please refer to # https://pytorch.org/tutorials/prototype/pt2e_quant_ptq_static.html#convert-the-calibrated-model-to-a-quantized-model # for detailed explanation of output quantized model quantized_model = convert_pt2e(prepared_model) z+quantization_api.quantize_pt2e.convert_pt2ezjUnexpected argument type for `use_reference_representation`, please make sure you intend to pass argument z to convert_pt2e)r"r#r$ isinstancebool ValueErrorr%rrrr graph_modulerrr6rr)r)rr7r8r+pms r-rrs4 HH  !NO 2D 9 <rRs! ?HL':!@AUU< J JJJZF FFFT II""66>> II""66== II""77?? >d>t>*/2 2"&22 2r.