o ei~@sRddlZddlZddlZddlmZddlZddlmZddlm Z ddl m Z m Z ddl mZddlmZmZmZdd lmZdd lmZdd lmZdd lmZmZmZdd lmZmZddl m!Z!ddl"m#Z#deej$j%Bde&e'efddfddZ(dej$j%ddfddZ)deddfddZ*dej$j%ddfddZ+  dBdede,dee&e'efBdBde e&e'efBdBdef ddZ-    dCdej$j%d!e!e&e'efBde,d"e.ed#fd$ee&e'efBdBd%e!e&e'efBdBde e&e'efBdBd&e,defd'd(Z/  dBdej$j%d!e!e&e'efBde,d"e.ed#fd$ee&e'efBdBde e&e'efBdBdefd)d*Z0  dBdej$j%dee&e'efBdBde e&e'efBdBdefd+d,Z1e2e#   dDdej$j%d!e!e&e'efBd"e.ed#fd$ee&e'efBdBd%e!e&e'efBdBde e&e'efBdBdefd-d.Z3e2e#  dBdej$j%d!e!e&e'efBd"e.ed#fd$ee&e'efBdBde e&e'efBdBdef d/d0Z4  1   dEd2ed3e,d4ee&e'efBdBd&e,d5e,d!e!e&e'efBdBde e&e'efBdBd6e,d7e,defd8d9Z5e2e#  1   dFd2ed4ee&e'efBdBd5e,d!e!e&e'efBdBde e&e'efBdBd7e,defd:d;Z6  1  dGd2ed4ee&e'efBdBd5e,d!e!e&e'efBdBde e&e'efBdBdef dd?Z8 dHd2ed3e,d4ee&e'efBdBdefd@dAZ9dS)IN)Any) GraphModule)_USER_PRESERVED_ATTRIBUTES_KEY) BackendConfigget_tensorrt_backend_config)convert)ConvertCustomConfigFuseCustomConfigPrepareCustomConfig)fuse)ObservedGraphModule)prepare)QuantizationTracerScopeScopeContextManager)get_custom_module_class_keys#get_skipped_module_name_and_classes)QConfigMapping)DEPRECATION_WARNINGmodelpreserved_attrsreturncCs8t||jt<|jtD] \}}t|||qdS)zXStore preserved attributes to the model.meta so that it can be preserved during deepcopyN)copymetaritemssetattr)rr attr_nameattrrk/var/www/addictedbytheproject.nl/epg/venv/lib/python3.10/site-packages/torch/ao/quantization/quantize_fx.pyattach_preserved_attrs_to_modelsr!cCs*t|tstdtt|dddS)Nz,input model must be a GraphModule, Got type:z Please make zsure to follow the tutorials.) isinstancer ValueErrorstrtyperrrr _check_is_graph_module%s  r'cCs"|jjD] }t|dsi|_qdS)aAttach meta field to all nodes of the graph if it does not exist, meta field is a field stores some meta information about the node, such as dtype and shape information for output of the node, this only exists if the program is captured by make_fx (used in quantize_pt2e flow), if the program is captured by torch.fx symbolic tracing, this field may not exist, so we add it here to avoid checking this all over the places rN)graphnodeshasattrr)rnoderrr !_attach_meta_to_node_if_not_exist0s  r,cCsfg}|D]\}}t|tjjjjr||qt|q|D]}|j |=tjjj |j |<q dS)z+Swap FloatFunctional with FXFloatFunctionalN) named_childrenr"torchaonn quantizedFloatFunctionalappend_swap_ff_with_fxff_modulesFXFloatFunctional)rmodules_to_swapnamemodulerrr r4=s  r4is_qatfuse_custom_configbackend_configcCst|t||||S)zInternal helper function to fuse modules in preparation for quantization Args: model: GraphModule object from symbolic tracing (torch.fx.symbolic_trace) )r'r )rr:r;r<rrr _fuse_fxKs r=Fqconfig_mappingexample_inputs.prepare_custom_config_equalization_configis_standalone_modulec s|durt}|durt}t|tr tjdtddt|}tt ||\}} |j } fdd| D} t || } t | } t| t|j }t| |||} t| ||| j|||||d }t|| |S)aZInternal helper function for prepare_fx Args: `model`, `qconfig_mapping`, `prepare_custom_config`, `_equalization_config`: see docs for :func:`~torch.ao.quantization.prepare_fx` `is_standalone_module`: a boolean flag indicates whether we are quantizing a standalone module or not, a standalone module is a submodule of the parent module that is not inlined in the forward graph of the parent module, the way we quantize standalone module is described in: :func:`~torch.ao.quantization._prepare_standalone_module_fx` NzPassing a prepare_custom_config_dict to prepare is deprecated and will not be supported in a future version. Please pass in a PrepareCustomConfig instead. stacklevelc"i|] }t|r|t|qSrr*getattr.0rr&rr  z_prepare_fx..)r?r@rAr<rB)r rr"dictwarningswarn FutureWarning from_dictr4rpreserved_attributesrrtracer,r set_preserved_attributesr=rnode_name_to_scoper!)rr>r:r?r@rAr<rBskipped_module_namesskipped_module_classespreserved_attr_namesrtracer graph_moduler;preparedrr&r _prepare_fxZsN     r\c Cst||||||ddS)a[Internal use only] Prepare a standalone module, so that it can be used when quantizing the parent module. standalone_module means it a submodule that is not inlined in parent module, and will be quantized separately as one unit. How the standalone module is observed is specified by `input_quantized_idxs` and `output_quantized_idxs` in the prepare_custom_config for the standalone module Returns: * model(GraphModule): prepared standalone module. It has these attributes in model.meta: * `standalone_module_input_quantized_idxs(List[Int])`: a list of indexes for the graph input that is expected to be quantized, same as input_quantized_idxs configuration provided for the standalone module * `standalone_module_output_quantized_idxs(List[Int])`: a list of indices for the graph output that is quantized same as input_quantized_idxs configuration provided for the standalone module T)r<rB)r\)rr>r:r?r@r<rrr _prepare_standalone_module_fxsr]cs|durt}t|trtjdtddt|}tj d|j }fdd|D}tj }t |t|d||}t|||S) aFuse modules like conv+bn, conv+bn+relu etc, model must be in eval mode. Fusion rules are defined in torch.ao.quantization.fx.fusion_pattern.py Args: * `model` (torch.nn.Module): a torch.nn.Module model * `fuse_custom_config` (FuseCustomConfig): custom configurations for fuse_fx. See :class:`~torch.ao.quantization.fx.custom_config.FuseCustomConfig` for more details Example:: from torch.ao.quantization import fuse_fx m = Model().eval() m = fuse_fx(m) NzPassing a fuse_custom_config_dict to fuse is deprecated and will not be supported in a future version. Please pass in a FuseCustomConfig instead.rDz$quantization_api.quantize_fx.fuse_fxcrFrrGrIr&rr rKrLzfuse_fx..F)r r"rMrNrOrPrQr._C_log_api_usage_oncerRfxsymbolic_tracer,r=r!)rr;r<rXrrZrr&r fuse_fxs&      rccCs tjdt||d||||S)a Prepare a model for post training quantization Args: * `model` (torch.nn.Module): torch.nn.Module model * `qconfig_mapping` (QConfigMapping): QConfigMapping object to configure how a model is quantized, see :class:`~torch.ao.quantization.qconfig_mapping.QConfigMapping` for more details * `example_inputs` (Tuple[Any, ...]): Example inputs for forward function of the model, Tuple of positional args (keyword args can be passed as positional args as well) * `prepare_custom_config` (PrepareCustomConfig): customization configuration for quantization tool. See :class:`~torch.ao.quantization.fx.custom_config.PrepareCustomConfig` for more details * `_equalization_config`: config for specifying how to perform equalization on the model * `backend_config` (BackendConfig): config that specifies how operators are quantized in a backend, this includes how the operators are observed, supported fusion patterns, how quantize/dequantize ops are inserted, supported dtypes etc. See :class:`~torch.ao.quantization.backend_config.BackendConfig` for more details Return: A GraphModule with observer (configured by qconfig_mapping), ready for calibration Example:: import torch from torch.ao.quantization import get_default_qconfig_mapping from torch.ao.quantization.quantize_fx import prepare_fx class Submodule(torch.nn.Module): def __init__(self) -> None: super().__init__() self.linear = torch.nn.Linear(5, 5) def forward(self, x): x = self.linear(x) return x class M(torch.nn.Module): def __init__(self) -> None: super().__init__() self.linear = torch.nn.Linear(5, 5) self.sub = Submodule() def forward(self, x): x = self.linear(x) x = self.sub(x) + x return 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) # qconfig is the configuration for how we insert observers for a particular # operator # qconfig = get_default_qconfig("fbgemm") # Example of customizing qconfig: # qconfig = torch.ao.quantization.QConfig( # activation=MinMaxObserver.with_args(dtype=torch.qint8), # weight=MinMaxObserver.with_args(dtype=torch.qint8)) # `activation` and `weight` are constructors of observer module # qconfig_mapping is a collection of quantization configurations, user can # set the qconfig for each operator (torch op calls, functional calls, module calls) # in the model through qconfig_mapping # the following call will get the qconfig_mapping that works best for models # that target "fbgemm" backend qconfig_mapping = get_default_qconfig_mapping("fbgemm") # We can customize qconfig_mapping in different ways. # e.g. set the global qconfig, which means we will use the same qconfig for # all operators in the model, this can be overwritten by other settings # qconfig_mapping = QConfigMapping().set_global(qconfig) # e.g. quantize the linear submodule with a specific qconfig # qconfig_mapping = QConfigMapping().set_module_name("linear", qconfig) # e.g. quantize all nn.Linear modules with a specific qconfig # qconfig_mapping = QConfigMapping().set_object_type(torch.nn.Linear, qconfig) # for a more complete list, please see the docstring for :class:`torch.ao.quantization.QConfigMapping` # argument # example_inputs is a tuple of inputs, that is used to infer the type of the # outputs in the model # currently it's not used, but please make sure model(*example_inputs) runs example_inputs = (torch.randn(1, 3, 224, 224),) # TODO: add backend_config after we split the backend_config for fbgemm and qnnpack # e.g. backend_config = get_default_backend_config("fbgemm") # `prepare_fx` inserts observers in the model based on qconfig_mapping and # backend_config. If the configuration for an operator in qconfig_mapping # is supported in the backend_config (meaning it's supported by the target # hardware), we'll insert observer modules according to the qconfig_mapping # otherwise the configuration in qconfig_mapping will be ignored # # Example: # in qconfig_mapping, user sets linear module to be quantized with quint8 for # activation and qint8 for weight: # qconfig = torch.ao.quantization.QConfig( # observer=MinMaxObserver.with_args(dtype=torch.quint8), # weight=MinMaxObserver.with-args(dtype=torch.qint8)) # Note: current qconfig api does not support setting output observer, but # we may extend this to support these more fine grained control in the # future # # qconfig_mapping = QConfigMapping().set_object_type(torch.nn.Linear, qconfig) # in backend config, linear module also supports in this configuration: # weighted_int8_dtype_config = DTypeConfig( # input_dtype=torch.quint8, # output_dtype=torch.quint8, # weight_dtype=torch.qint8, # bias_type=torch.float) # linear_pattern_config = BackendPatternConfig(torch.nn.Linear) \ # .set_observation_type(ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT) \ # .add_dtype_config(weighted_int8_dtype_config) \ # ... # backend_config = BackendConfig().set_backend_pattern_config(linear_pattern_config) # `prepare_fx` will check that the setting requested by suer in qconfig_mapping # is supported by the backend_config and insert observers and fake quant modules # in the model prepared_model = prepare_fx(float_model, qconfig_mapping, example_inputs) # Run calibration calibrate(prepared_model, sample_inference_data) z'quantization_api.quantize_fx.prepare_fxFr.r_r`r\)rr>r?r@rAr<rrr prepare_fxs recCs tjdt||d|||dS)aPrepare a model for quantization aware training Args: * `model` (torch.nn.Module): torch.nn.Module model * `qconfig_mapping` (QConfigMapping): see :func:`~torch.ao.quantization.prepare_fx` * `example_inputs` (Tuple[Any, ...]): see :func:`~torch.ao.quantization.prepare_fx` * `prepare_custom_config` (PrepareCustomConfig): see :func:`~torch.ao.quantization.prepare_fx` * `backend_config` (BackendConfig): see :func:`~torch.ao.quantization.prepare_fx` Return: A GraphModule with fake quant modules (configured by qconfig_mapping and backend_config), ready for quantization aware training Example:: import torch from torch.ao.quantization import get_default_qat_qconfig_mapping from torch.ao.quantization.quantize_fx import prepare_qat_fx class Submodule(torch.nn.Module): def __init__(self) -> None: super().__init__() self.linear = torch.nn.Linear(5, 5) def forward(self, x): x = self.linear(x) return x class M(torch.nn.Module): def __init__(self) -> None: super().__init__() self.linear = torch.nn.Linear(5, 5) self.sub = Submodule() def forward(self, x): x = self.linear(x) x = self.sub(x) + x return x # initialize a floating point model float_model = M().train() # (optional, but preferred) load the weights from pretrained model # float_model.load_weights(...) # define the training loop for quantization aware training def train_loop(model, train_data): model.train() for image, target in data_loader: ... # qconfig is the configuration for how we insert observers for a particular # operator # qconfig = get_default_qconfig("fbgemm") # Example of customizing qconfig: # qconfig = torch.ao.quantization.QConfig( # activation=FakeQuantize.with_args(observer=MinMaxObserver.with_args(dtype=torch.qint8)), # weight=FakeQuantize.with_args(observer=MinMaxObserver.with_args(dtype=torch.qint8))) # `activation` and `weight` are constructors of observer module # qconfig_mapping is a collection of quantization configurations, user can # set the qconfig for each operator (torch op calls, functional calls, module calls) # in the model through qconfig_mapping # the following call will get the qconfig_mapping that works best for models # that target "fbgemm" backend qconfig_mapping = get_default_qat_qconfig_mapping("fbgemm") # We can customize qconfig_mapping in different ways, please take a look at # the docstring for :func:`~torch.ao.quantization.prepare_fx` for different ways # to configure this # example_inputs is a tuple of inputs, that is used to infer the type of the # outputs in the model # currently it's not used, but please make sure model(*example_inputs) runs example_inputs = (torch.randn(1, 3, 224, 224),) # TODO: add backend_config after we split the backend_config for fbgemm and qnnpack # e.g. backend_config = get_default_backend_config("fbgemm") # `prepare_qat_fx` inserts observers in the model based on qconfig_mapping and # backend_config, if the configuration for an operator in qconfig_mapping # is supported in the backend_config (meaning it's supported by the target # hardware), we'll insert fake_quantize modules according to the qconfig_mapping # otherwise the configuration in qconfig_mapping will be ignored # see :func:`~torch.ao.quantization.prepare_fx` for a detailed explanation of # how qconfig_mapping interacts with backend_config prepared_model = prepare_qat_fx(float_model, qconfig_mapping, example_inputs) # Run training train_loop(prepared_model, train_loop) z+quantization_api.quantize_fx.prepare_qat_fxT)r<rd)rr>r?r@r<rrr prepare_qat_fxs frfTrZ is_referenceconvert_custom_config_remove_qconfig is_decomposedkeep_original_weightsc sz|durt}t|trtjdtddt|}t|j} fdd| D} t ||||||||d } t | | | S)z_`is_standalone_module`: see docs in :func:`~torch.ao.quantization.prepare_standalone_module_fx`NzPassing a convert_custom_config_dict to convert is deprecated and will not be supported in a future version. Please pass in a ConvertCustomConfig instead.rCrDcrFrrGrIrZrr rK rLz_convert_fx..)_remove_qconfig_flagr>r<rjrk) r r"rMrNrOrPrQr'rRrr!) rZrgrhrBrir>r<rjrkrXrr1rrlr _convert_fxs6     rnc Cs"tjdt|d|||||dS)a Convert a calibrated or trained model to a quantized model Args: * `graph_module` (torch.fx.GraphModule): A prepared and calibrated/trained model (GraphModule) * `convert_custom_config` (ConvertCustomConfig): custom configurations for convert function. See :class:`~torch.ao.quantization.fx.custom_config.ConvertCustomConfig` for more details * `_remove_qconfig` (bool): Option to remove the qconfig attributes in the model after convert. * `qconfig_mapping` (QConfigMapping): config for specifying how to convert a model for quantization. The keys must include the ones in the qconfig_mapping passed to `prepare_fx` or `prepare_qat_fx`, with the same values or `None`. Additional keys can be specified with values set to `None`. For each entry whose value is set to None, we skip quantizing that entry in the model:: qconfig_mapping = QConfigMapping .set_global(qconfig_from_prepare) .set_object_type(torch.nn.functional.add, None) # skip quantizing torch.nn.functional.add .set_object_type(torch.nn.functional.linear, qconfig_from_prepare) .set_module_name("foo.bar", None) # skip quantizing module "foo.bar" * `backend_config` (BackendConfig): A configuration for the backend which describes how operators should be quantized in the backend, this includes quantization mode support (static/dynamic/weight_only), dtype support (quint8/qint8 etc.), observer placement for each operators and fused operators. See :class:`~torch.ao.quantization.backend_config.BackendConfig` for more details Return: A quantized model (torch.nn.Module) Example:: # prepared_model: the model after prepare_fx/prepare_qat_fx and calibration/training # convert_fx converts a calibrated/trained model to a quantized model for the # target hardware, this includes converting the model first to a reference # quantized model, and then lower the reference quantized model to a backend # Currently, the supported backends are fbgemm (onednn), qnnpack (xnnpack) and # they share the same set of quantized operators, so we are using the same # lowering procedure # # backend_config defines the corresponding reference quantized module for # the weighted modules in the model, e.g. nn.Linear # TODO: add backend_config after we split the backend_config for fbgemm and qnnpack # e.g. backend_config = get_default_backend_config("fbgemm") quantized_model = convert_fx(prepared_model) z'quantization_api.quantize_fx.convert_fxF)rgrhrir>r<rkr.r_r`rn)rZrhrir>r<rkrrr convert_fx6s :rpcCs tjdt|d||||dS)a}Convert a calibrated or trained model to a reference quantized model, see https://github.com/pytorch/rfcs/blob/master/RFC-0019-Extending-PyTorch-Quantization-to-Custom-Backends.md for more details, reference quantized model is a standard representation of a quantized model provided by FX Graph Mode Quantization, it can be further lowered to run on the target hardware, like accelerators Args: * `graph_module` (GraphModule): A prepared and calibrated/trained model (GraphModule) * `convert_custom_config` (ConvertCustomConfig): custom configurations for convert function. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. * `_remove_qconfig` (bool): Option to remove the qconfig attributes in the model after convert. * `qconfig_mapping` (QConfigMapping): config for specifying how to convert a model for quantization. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. * `backend_config` (BackendConfig): A configuration for the backend which describes how operators should be quantized in the backend. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. Return: A reference quantized model (GraphModule) Example:: # prepared_model: the model after prepare_fx/prepare_qat_fx and calibration/training # TODO: add backend_config after we split the backend_config for fbgemm and qnnpack # e.g. backend_config = get_default_backend_config("fbgemm") reference_quantized_model = convert_to_reference_fx(prepared_model) z4quantization_api.quantize_fx.convert_to_reference_fxT)rgrhrir>r<ro)rZrhrir>r<rrr convert_to_reference_fx|s 'rqc Cs"tjdt|d|d||ddS)aConvert a calibrated or trained model to a reference quantized model, with decomposed representation for quantized Tensor see https://github.com/pytorch/rfcs/blob/master/RFC-0019-Extending-PyTorch-Quantization-to-Custom-Backends.md for more details, reference quantized model is a standard representation of a quantized model provided by FX Graph Mode Quantization, it can be further lowered to run on the target hardware, like accelerators Note: this is not public API Args: * `graph_module` (GraphModule): A prepared and calibrated/trained model (GraphModule) * `convert_custom_config` (ConvertCustomConfig): custom configurations for convert function. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. * `_remove_qconfig` (bool): Option to remove the qconfig attributes in the model after convert. * `qconfig_mapping` (QConfigMapping): config for specifying how to convert a model for quantization. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. * `backend_config` (BackendConfig): A configuration for the backend which describes how operators should be quantized in the backend. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. Return: A reference quantized model (GraphModule) with operators working with decomposed quantized Tensor Example:: # prepared_model: the model after prepare_fx/prepare_qat_fx and calibration/training # TODO: add backend_config after we split the backend_config for fbgemm and qnnpack # e.g. backend_config = get_default_backend_config("fbgemm") reference_quantized_model = _convert_to_reference_decomposed_fx(prepared_model) z@quantization_api.quantize_fx._convert_to_reference_decomposed_fxTF)rgrhrir>r<rjro)rZrhr>r<rrr #_convert_to_reference_decomposed_fxs)rrcCst|||ddS)av[Internal use only] Convert a model produced by :func:`~torch.ao.quantization.prepare_standalone_module_fx` and convert it to a quantized model Returns a quantized standalone module, whether input/output is quantized is specified by prepare_custom_config, with input_quantized_idxs, output_quantized_idxs, please see docs for prepare_fx for details T)rB)rn)rZrgrhrrr _convert_standalone_module_fxs  rs)NN)NNNF)NNN)NFTNNFF)NTNNF)NTNN)FN):rtyping_extensionsrNtypingrr.torch.fxrtorch.fx.graph_modulerr<rr fx.convertrfx.custom_configr r r fx.fuser fx.graph_moduler fx.preparer fx.tracerrrrfx.utilsrrr>rutilsrr0ModulerMr$r!r'r,r4boolr=tupler\r]rc deprecatedrerfrnrprqrrrsrrrr s                  M  , 1    s  0 G 4 9