import torch from torch.fx import GraphModule from torch.fx.graph import ( Node, Graph, ) from ..observer import ( default_affine_fixed_qparams_observer, default_symmetric_fixed_qparams_observer, ) from ..quantization_mappings import ( get_static_quant_module_class, get_dynamic_quant_module_class, ) from ..utils import ( _parent_name, get_swapped_custom_module_class, activation_is_statically_quantized, activation_is_int8_quantized, weight_is_statically_quantized, get_qconfig_dtypes, activation_dtype, get_qparam_dict, ) from torch.ao.quantization.quantize import ( is_activation_post_process, ) from .pattern_utils import ( register_quant_pattern, get_default_output_activation_post_process_map, Pattern, ) from .utils import ( all_node_args_have_no_tensors, quantize_node, get_per_tensor_qparams, get_linear_prepack_op_for_dtype, create_qparam_nodes, get_qconv_prepack_op, get_qconv_op, create_node_from_old_node_preserve_meta, ) from ..qconfig import QConfigAny from abc import ABC import operator import warnings from typing import Any, Callable, Dict, Union, Optional, Tuple, List # ------------------------- # Pattern Registrations # ------------------------- # 1. Post Training Static Quantization and Quantization Aware Training Patterns # Base Pattern Handler class QuantizeHandler(ABC): """ Base handler class for the quantizer patterns """ def __init__(self, node: Node, modules: Dict[str, torch.nn.Module]): """ Records pattern information in __init__, which will be used in convert """ # this is an indicator of whether all the inputs are Node or not # since some op might be quantized differently depending on whether # all inputs are tensors or not, e.g. add/mul self.num_tensor_args = len(node.args) self.all_node_args_are_tensors = True # the last node of the matched pattern self.last_node = node def _maybe_get_last_node_only_observer( self, modules: Dict[str, torch.nn.Module] ) -> Optional[torch.nn.Module]: """ If the last node of the pattern is observed, return the observer instance. Otherwise, return None. """ for maybe_obs_node, _ in self.last_node.users.items(): if maybe_obs_node.op == 'call_module': maybe_obs = modules[str(maybe_obs_node.target)] if is_activation_post_process(maybe_obs): return maybe_obs return None def input_output_observed(self) -> bool: """ Returns True if the pattern matched to this qhandler could be be observed, and False it it should not be observed. """ return True def is_general_tensor_value_op(self) -> bool: """ Returns True if the operator works for both floating point and quantized input, and does some computation based on the input Tensor, so we need to insert observer/fake_quant for the output of the operator since the distribution of values is different for input and output Tensors (for HistogramObserver) while they share the same quantization parameters Example: avgpool2d """ return False def is_general_tensor_shape_op(self) -> bool: """ Similar to is_general_tensor_value_op, this is a check for ops that works for both floating point and quantized input, that only re-arranges the Tensor values or query some metadata about the Tensor We don't insert observer/fake_quant for the output of these operators Example: reshape, transpose, maxpool2d """ return False def should_insert_observer_for_output( self, qconfig: Any, model_is_training: bool, ) -> bool: """ Returns true if an observer should be inserted for the output of the pattern matched to this QuantizeHandler instance during the prepare step. """ # TODO(future PR): potentially clean up and deduplicate these # mappings. return self.all_node_args_are_tensors and self.input_output_observed() def should_mark_output_quantized_from_input_quantized_status( self, qconfig: QConfigAny ) -> bool: """ Returns true if after convert, the output of the matched pattern is quantized iff the first input is also quantized. """ return False def get_activation_ctr( self, qconfig: Any, pattern: Pattern, is_training: bool, ) -> Optional[Callable]: """ Returns the constructor for the activation observer which should be used for the pattern matched to this handler. Some handlers override this to a different value than what is specified in the qconfig. """ return qconfig.activation def is_output_quantized(self, qconfig): """ Returns true if the output node of convert is quantized when is_reference is False, we would return float node when a certain dtype combination is not supported (since fbgemm/qnnpack only support certain dtype combinations), so the output may be float, but when is_reference is True, we support all dtype combinations so the output will always be quantized. TODO: This is fragile, whether output is quantized should not depend on `is_reference` since we want to make sure whether a Tensor is quantized should be the same in prepare and convert and is_reference is only available in convert currently """ return True def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: """ Convert the given node to a quantized node and insert it to the quantized graph """ return NotImplemented # Binary op configs # Supported combinations are: # quant_type | activation (compute_type) | weight # static quint8 qint8 # tuple (activation_dtype, weight_dtype, compute_dtype) # these are supported types for common binary ops like add/mul etc. all_dtypes = [ (torch.qint8, torch.qint8, None), (torch.quint8, torch.qint8, None), (torch.float16, torch.float16, None), ] fp16_dtypes = [ (torch.float16, torch.float16, None) ] int8_dtypes = [ (torch.qint8, torch.qint8, None), (torch.quint8, torch.qint8, None), ] binary_op_supported_dtypes : Dict[Union[Callable, str], List[Tuple[torch.dtype, torch.dtype, None]]] = { operator.add: all_dtypes, torch.add: all_dtypes, operator.mul: all_dtypes, torch.mul: all_dtypes, torch.bmm: fp16_dtypes, torch.sub: fp16_dtypes, operator.sub: fp16_dtypes, torch.div: fp16_dtypes, operator.truediv: fp16_dtypes, torch.matmul: int8_dtypes, } default_op_supported_dtypes = { torch.nn.ConvTranspose1d: int8_dtypes, torch.nn.ConvTranspose2d: int8_dtypes, torch.nn.ELU: int8_dtypes, torch.nn.LeakyReLU: int8_dtypes, torch.nn.Hardswish: int8_dtypes, torch.nn.InstanceNorm1d: int8_dtypes, torch.nn.InstanceNorm2d: int8_dtypes, torch.nn.InstanceNorm3d: int8_dtypes, torch.nn.LayerNorm: all_dtypes, torch.nn.SiLU: fp16_dtypes, torch.nn.Mish: fp16_dtypes, torch.nn.GELU: int8_dtypes, torch.nn.Dropout: int8_dtypes, torch.nn.Softmax: int8_dtypes, torch.nn.functional.elu: int8_dtypes, torch.nn.functional.hardswish: int8_dtypes, torch.nn.functional.instance_norm: int8_dtypes, torch.nn.functional.layer_norm: all_dtypes, torch.nn.functional.leaky_relu: int8_dtypes, torch.nn.functional.silu: fp16_dtypes, torch.nn.functional.mish: fp16_dtypes, torch.nn.functional.gelu: int8_dtypes, torch.nn.functional.softmax: int8_dtypes, torch.nn.functional.dropout: int8_dtypes, torch.sum: fp16_dtypes, } QAT_CONV_MODULE_CLASSES = \ (torch.nn.qat.Conv2d, torch.nn.qat.Conv3d, torch.nn.intrinsic.qat.ConvBn1d, torch.nn.intrinsic.qat.ConvBn2d, torch.nn.intrinsic.qat.ConvBn3d, torch.nn.intrinsic.qat.ConvBnReLU1d, torch.nn.intrinsic.qat.ConvBnReLU2d, torch.nn.intrinsic.qat.ConvBnReLU3d, torch.nn.intrinsic.qat.ConvReLU2d, torch.nn.intrinsic.qat.ConvReLU3d) ########################## # Helper Functions ########################## def _load_weight_qparams( self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): key = prefix + "_weight_qparams" if key in state_dict: self._weight_qparams = state_dict[key] state_dict.pop(key) def _save_weight_qparams(self, destination, prefix, keep_vars): for attr_name in dir(self): if "_weight_qparams" == attr_name and \ isinstance(getattr(self, attr_name), dict): weight_qparams = getattr(self, attr_name) destination[prefix + attr_name] = weight_qparams def _to_reference(float_module, weight_qparams): """ Make a weighted float module (e.g. conv and linear )a reference module by attaching _weight_qparams that records the qparams for weight and change the name for the module so that it's recognized when people print the model """ float_module._weight_qparams = weight_qparams float_module._register_state_dict_hook(_save_weight_qparams) float_module._register_load_state_dict_pre_hook(_load_weight_qparams, with_module=True) float_module_name = float_module._get_name() def _get_name(): return float_module_name + "(Reference)" float_module._get_name = _get_name @register_quant_pattern(operator.add) @register_quant_pattern(operator.sub) @register_quant_pattern(operator.mul) @register_quant_pattern(operator.truediv) @register_quant_pattern(torch.add) @register_quant_pattern(torch.sub) @register_quant_pattern(torch.mul) @register_quant_pattern(torch.div) @register_quant_pattern(torch.bmm) @register_quant_pattern((torch.nn.ReLU, operator.add)) @register_quant_pattern((torch.nn.ReLU, operator.mul)) @register_quant_pattern((torch.nn.ReLU, torch.add)) @register_quant_pattern((torch.nn.ReLU, torch.mul)) @register_quant_pattern((torch.nn.functional.relu, operator.add)) @register_quant_pattern((torch.nn.functional.relu, operator.mul)) @register_quant_pattern((torch.nn.functional.relu, torch.add)) @register_quant_pattern((torch.nn.functional.relu, torch.mul)) @register_quant_pattern((torch.relu, operator.add)) @register_quant_pattern((torch.relu, operator.mul)) @register_quant_pattern(torch.matmul) class BinaryOpQuantizeHandler(QuantizeHandler): def __init__( self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) self.relu_node = None if ( node.op == 'call_function' and node.target in (torch.nn.functional.relu, torch.relu) ) or ( node.op == 'call_module' and isinstance(modules[str(node.target)], torch.nn.ReLU) ): self.relu_node = node node = node.args[0] # type: ignore[assignment] self.binary_op_node = node self.binary_op = node.target # determine how many of the first two args are Tensors (versus scalars) # this distinguishes things like "x + y" from "x + 2" or "2 + x" self.num_tensor_args = 0 cache_for_no_tensor_check: Dict[Node, bool] = dict() for arg_idx in range(len(self.binary_op_node.args)): arg = self.binary_op_node.args[arg_idx] if isinstance(arg, Node) and (not all_node_args_have_no_tensors(arg, modules, cache_for_no_tensor_check)): self.num_tensor_args += 1 self.all_node_args_are_tensors = \ (self.num_tensor_args == len(self.binary_op_node.args)) def should_insert_observer_for_output( self, qconfig: Any, model_is_training: bool, ) -> bool: """ Returns true if an observer should be inserted for the output of the pattern matched to this QuantizeHandler instance during the prepare step. """ dtypes = get_qconfig_dtypes(qconfig) if not (self.binary_op in binary_op_supported_dtypes and dtypes in binary_op_supported_dtypes[self.binary_op]): return False if self.num_tensor_args == 1: return True elif self.all_node_args_are_tensors and self.input_output_observed(): return True else: return False def is_general_tensor_value_op(self) -> bool: return self.num_tensor_args == 1 def input_output_observed(self): # for x + y where x and y are scalars, we do not observe anything return self.num_tensor_args > 0 def is_output_quantized(self, qconfig): dtypes = get_qconfig_dtypes(qconfig) return self.binary_op in binary_op_supported_dtypes and \ dtypes in binary_op_supported_dtypes[self.binary_op] def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: if self.num_tensor_args == 0: # example: x + y, when x and y are scalars return quantized_graph.node_copy( node, load_arg(quantized=None)) dtypes = get_qconfig_dtypes(qconfig) act_dtype = activation_dtype(qconfig) dtypes = get_qconfig_dtypes(qconfig) if act_dtype == torch.float or \ not (self.binary_op in binary_op_supported_dtypes and dtypes in binary_op_supported_dtypes[self.binary_op]): if self.relu_node: op_out = quantized_graph.node_copy(self.binary_op_node, load_arg(quantized=torch.float)) relu_args = [op_out] relu_args.extend(load_arg(quantized=torch.float)(self.relu_node.args[1:])) relu_kwargs = load_arg(quantized=torch.float)(self.relu_node.kwargs) return create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.relu, tuple(relu_args), relu_kwargs), self.relu_node) else: return quantized_graph.node_copy(node, load_arg(quantized=torch.float)) else: if self.num_tensor_args == 2: # make sure both inputs are quantized to act_dtype load_arg(quantized={0: act_dtype, 1: act_dtype})(self.binary_op_node.args) args = load_arg(quantized=torch.float)(self.binary_op_node.args) kwargs = load_arg(quantized=torch.float)(self.binary_op_node.kwargs) op_out = quantized_graph.node_copy(self.binary_op_node, load_arg(quantized=torch.float)) def modified_load_arg(n: Node): if n.name == self.binary_op_node.name: return op_out else: return load_arg(quantized=torch.float)(n) if self.relu_node: op_out = quantized_graph.node_copy(self.relu_node, modified_load_arg) activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None return quantize_node( op_out, activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) @register_quant_pattern(torch.cat) class CatQuantizeHandler(QuantizeHandler): def is_general_tensor_value_op(self) -> bool: return True def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: if not self.all_node_args_are_tensors: return NotImplemented act_dtype = activation_dtype(qconfig) if act_dtype == torch.float: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return op_out else: activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None # make sure the first argument is quantized to act_dtype load_arg(quantized={0: act_dtype})(node.args) args = list(load_arg(quantized=torch.float)(node.args)) kwargs = load_arg(quantized=torch.float)(node.kwargs) op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return quantize_node( op_out, activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) # handle conv, maybe followed by relu # NB: matching order is reversed, that is we match from the bottom of this list to the beginning @register_quant_pattern(torch.nn.Conv1d) @register_quant_pattern(torch.nn.Conv2d) @register_quant_pattern(torch.nn.Conv3d) @register_quant_pattern(torch.nn.functional.conv1d) @register_quant_pattern(torch.nn.functional.conv2d) @register_quant_pattern(torch.nn.functional.conv3d) # TODO: add qat.Conv1d @register_quant_pattern(torch.nn.qat.Conv2d) @register_quant_pattern(torch.nn.qat.Conv3d) @register_quant_pattern(torch.nn.intrinsic.ConvReLU1d) @register_quant_pattern(torch.nn.intrinsic.ConvReLU2d) @register_quant_pattern(torch.nn.intrinsic.ConvReLU3d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvBn1d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvBn2d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvBn3d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvBnReLU1d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvBnReLU2d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvBnReLU3d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvReLU2d) @register_quant_pattern(torch.nn.intrinsic.qat.ConvReLU3d) @register_quant_pattern((torch.nn.functional.relu, torch.nn.functional.conv1d)) @register_quant_pattern((torch.nn.functional.relu, torch.nn.functional.conv2d)) @register_quant_pattern((torch.nn.functional.relu, torch.nn.functional.conv3d)) @register_quant_pattern((torch.nn.ReLU, torch.nn.functional.conv1d)) @register_quant_pattern((torch.nn.ReLU, torch.nn.functional.conv2d)) @register_quant_pattern((torch.nn.ReLU, torch.nn.functional.conv3d)) # just for error checks @register_quant_pattern((torch.nn.ReLU, torch.nn.Conv1d)) @register_quant_pattern((torch.nn.ReLU, torch.nn.Conv2d)) @register_quant_pattern((torch.nn.ReLU, torch.nn.Conv3d)) @register_quant_pattern((torch.nn.functional.relu, torch.nn.Conv2d)) @register_quant_pattern((torch.nn.functional.relu, torch.nn.Conv3d)) # TODO: rename Relu -> ReLU to be more consistent with other classes class ConvReluQuantizeHandler(QuantizeHandler): def __init__(self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) self.relu_node = None if (node.op == 'call_function' and node.target is torch.nn.functional.relu) or \ (node.op == 'call_module' and isinstance(modules[str(node.target)], torch.nn.ReLU)): self.relu_node = node node = node.args[0] # type: ignore[assignment] self.conv_node = node if node.op == "call_module": self.conv = modules[str(self.conv_node.target)] elif node.op == "call_function": self.conv = node.target # type: ignore[assignment] def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: # Supported combinations are: # quant_type | activation (compute_type) | weight # static quint8 qint8 # tuple (activation_dtype, weight_dtype, compute_dtype) supported_dtypes = [ (torch.quint8, torch.qint8, None), ] # TODO: is_reference option for conv module dtypes = get_qconfig_dtypes(qconfig) # leave the op unquantized if the dtype combination is not supported if not is_reference and dtypes not in supported_dtypes: warnings.warn( "dtype combination: {} is not " "supported by Conv " "supported dtype combinations are: {}".format(dtypes, supported_dtypes)) if self.relu_node: conv_out = quantized_graph.node_copy(self.conv_node, load_arg(quantized=torch.float)) relu_args = [conv_out] relu_args.extend(load_arg(quantized=torch.float)(self.relu_node.args[1:])) relu_kwargs = load_arg(quantized=torch.float)(self.relu_node.kwargs) return create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.relu, tuple(relu_args), relu_kwargs), self.relu_node) else: return quantized_graph.node_copy(node, load_arg(quantized=torch.float)) activation_int8_quantized = activation_is_int8_quantized(qconfig) if self.conv_node.op == 'call_module': # note that relu should already be fused into conv module in the fusion step assert self.relu_node is None, 'conv module and relu fusion is not executed, ' \ 'please make sure to run fusion before prepare' output_activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert output_activation_post_process is not None module_types_supports_reference_pattern = [ torch.nn.Conv1d, torch.nn.Conv2d, torch.nn.Conv3d, torch.nn.intrinsic.ConvReLU1d, torch.nn.intrinsic.ConvReLU2d, torch.nn.intrinsic.ConvReLU3d, ] module_types_supports_reference_pattern.extend(list(QAT_CONV_MODULE_CLASSES)) # We'll always produce reference pattern for torch.nn.Conv*d, # will remove the else branch after we migrated all use cases if is_reference or \ type(self.conv) in module_types_supports_reference_pattern and \ dtypes in [(torch.quint8, torch.qint8, None)]: # produce dequant - float_op - quant pattern dtype = torch.float if activation_int8_quantized: dtype = activation_dtype(qconfig) activation = load_arg(quantized=dtype)(self.conv_node.args[0]) args = load_arg(quantized=torch.float)(self.conv_node.args) # Get the float conv and attach quantization scheme and quantization # parameters of weight to the module # and qparam is a dictionary of # {"qscheme": ..., "scale": ..., "zero_point": ...} for per tensor quantization or # {"qscheme": ..., "scale": ..., "zero_point": ..., "axis": ...} for per channel quantization float_conv = self.conv fused_conv = None if isinstance( float_conv, QAT_CONV_MODULE_CLASSES): # case 1. converting qat conv module to # a float conv module, we need to attch # weight fake_quant to the conv module, # weight fake_quant is assumed to be run during # QAT so we don't need to run it again here float_conv = float_conv.to_float() # type: ignore[operator] # change qat conv to conv parent_name, name = _parent_name(self.conv_node.target) setattr(modules[parent_name], name, float_conv) if isinstance(float_conv, torch.nn.intrinsic._FusedModule): fused_conv = float_conv float_conv = fused_conv[0] weight_post_process = self.conv.weight_fake_quant else: # case 2. converting a conv module/fused conv module # to float conv module, we need to attach # weight observer to the conv module and run it # with conv weight if isinstance(float_conv, torch.nn.intrinsic._FusedModule): fused_conv = float_conv float_conv = fused_conv[0] # type: ignore[index] assert qconfig is not None weight_post_process = qconfig.weight() # return early when we don't have a valid match # this typically happens when we called the same conv multiple times in the # same graph, and it is transformed in previous steps into a reference conv already if type(float_conv) not in [torch.nn.Conv1d, torch.nn.Conv2d, torch.nn.Conv3d]: op_out = create_node_from_old_node_preserve_meta( quantized_graph, ('call_module', self.conv_node.target, args, {}), self.conv_node) return op_out qconv_cls = get_static_quant_module_class( type(float_conv), is_reference=True) # run weight observer # TODO: This is currently a hack for QAT to get the right shapes for scale and zero point. # In the future, we should require the user to calibrate the model after calling prepare weight_post_process(float_conv.weight) # type: ignore[operator] weight_qparams = get_qparam_dict(weight_post_process) # hardcoded for now, TODO: expose the api to user, # we can have a map from module to reference module # and allow user to register new ones ref_conv = qconv_cls.from_float(float_conv, weight_qparams) # type: ignore[attr-defined] # if the parent is a fused conv (Sequential), we can replace the first # item to ref conv, otherwise we can update # the conv instance in the module tree if fused_conv is not None: fused_conv[0] = ref_conv parent_name, name = _parent_name(self.conv_node.target) setattr(modules[parent_name], name, fused_conv) else: parent_name, name = _parent_name(self.conv_node.target) setattr(modules[parent_name], name, ref_conv) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ('call_module', self.conv_node.target, args, {}), self.conv_node) if output_activation_post_process: op_out = quantize_node( op_out, output_activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) return op_out else: if convert_custom_config_dict is None: convert_custom_config_dict = {} additional_static_quant_mapping = convert_custom_config_dict.get("static", {}) # 1. attach activation post process to module self.conv.activation_post_process = output_activation_post_process # 2. select quantized class qconv_cls = get_static_quant_module_class( type(self.conv), additional_static_quant_mapping, is_reference=is_reference) quantized = qconv_cls.from_float(self.conv) parent_name, name = _parent_name(self.conv_node.target) setattr(modules[parent_name], name, quantized) return create_node_from_old_node_preserve_meta( quantized_graph, ( 'call_module', self.conv_node.target, (load_arg(quantized=torch.quint8)(self.conv_node.args[0]),), {}, ), self.conv_node) else: # call_function assert self.conv_node.op == "call_function" conv_functional_ops = { torch.nn.functional.conv1d, torch.nn.functional.conv2d, torch.nn.functional.conv3d, } if is_reference or self.conv_node.target in conv_functional_ops and\ dtypes in [(torch.quint8, torch.qint8, None)]: # make sure the input and weight are quantized to torch.quint8, torch.qint8, respectively load_arg(quantized={0: torch.quint8, 1: torch.qint8})(self.conv_node.args) args = load_arg(quantized=torch.float)(self.conv_node.args) kwargs = load_arg(quantized=torch.float)(self.conv_node.kwargs) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", self.conv, args, kwargs), self.conv_node) if self.relu_node: relu_args = [op_out] relu_args.extend(load_arg(quantized=torch.float)(self.relu_node.args[1:])) relu_kwargs = load_arg(quantized=torch.float)(self.relu_node.kwargs) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.relu, tuple(relu_args), relu_kwargs), self.relu_node) if activation_int8_quantized: root_module = modules[''] act_post_process_name = self.relu_node.name if self.relu_node else self.conv_node.name act_post_process_node = self.relu_node if self.relu_node else self.conv_node activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None return quantize_node( op_out, activation_post_process, act_post_process_node, modules, quantized_graph, node_name_to_scope, is_input=False) else: # output for dynamically quantized conv op is not quantized return op_out else: assert len(self.conv_node.args) >= 7, \ "only conv2d calls with all arguments specified is supported right now in is_reference=False option" # make sure the input and weight are quantized to torch.quint8, torch.qint8, respectively args = load_arg(quantized={0: torch.quint8, 1: torch.qint8})(self.conv_node.args) # pack weight weight = load_arg(quantized=torch.qint8)(self.conv_node.args[1]) other_args = load_arg(quantized=torch.float)(self.conv_node.args[2:]) bias, stride, padding, dilation, groups = other_args if self.conv == torch.nn.functional.conv1d: # F.conv1d can take `int` as well as `list[int]` for stride, # padding, dilation, but the prepack op cannot. Convert # these to lists if needed. stride = [stride] if isinstance(stride, int) else stride padding = [padding] if isinstance(padding, int) else padding dilation = [dilation] if isinstance(dilation, int) else dilation prepack_args = (weight, bias, stride, padding, dilation, groups) prepack_op = get_qconv_prepack_op(self.conv) packed_weight = quantized_graph.create_node( "call_function", prepack_op, prepack_args, {}) assert activation_int8_quantized, \ "currently only static quantization is supported for conv" # construct conv input if activation_int8_quantized: qconv_op = get_qconv_op(self.conv, self.relu_node is not None) conv_input = load_arg(quantized=torch.quint8)(self.conv_node.args[0]) activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None scale, zero_point, _ = get_per_tensor_qparams(activation_post_process) scale_node, zero_point_node = \ create_qparam_nodes( self.conv_node.name, scale, zero_point, modules, quantized_graph, node_name_to_scope) qconv_args = (conv_input, packed_weight, scale_node, zero_point_node) kwargs = load_arg(quantized=torch.float)(self.conv_node.kwargs) op = create_node_from_old_node_preserve_meta( quantized_graph, ('call_function', qconv_op, qconv_args, kwargs), self.conv_node) # Store the name of the fused op to get the path of node after fusion as well. # TODO: may need to change the key to Node regenerate the map in each transformation, # since we might not be able to rely on the name node_name_to_scope[op.name] = node_name_to_scope[self.conv_node.name] return op else: # conv2d_dyanmic branch raise Exception("Only static quant is supported for conv") @register_quant_pattern(torch.nn.Linear) @register_quant_pattern(torch.nn.functional.linear) @register_quant_pattern(torch.nn.qat.Linear) @register_quant_pattern(torch.nn.intrinsic.LinearReLU) @register_quant_pattern(torch.nn.intrinsic.qat.LinearReLU) @register_quant_pattern((torch.nn.functional.relu, torch.nn.functional.linear)) @register_quant_pattern((torch.nn.ReLU, torch.nn.functional.linear)) @register_quant_pattern(torch.nn.intrinsic.LinearBn1d) @register_quant_pattern(torch.nn.intrinsic.qat.LinearBn1d) # for error checks @register_quant_pattern((torch.nn.ReLU, torch.nn.Linear)) @register_quant_pattern((torch.nn.functional.relu, torch.nn.Linear)) class LinearReLUQuantizeHandler(QuantizeHandler): def __init__( self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) self.relu_node = None if (node.op == 'call_function' and node.target is torch.nn.functional.relu) or \ (node.op == 'call_module' and isinstance(modules[str(node.target)], torch.nn.ReLU)): self.relu_node = node node = node.args[0] # type: ignore[assignment] self.linear_node = node if node.op == 'call_module': self.linear = modules[str(self.linear_node.target)] def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: if convert_custom_config_dict is None: convert_custom_config_dict = {} # Supported combinations are: # quant_type | activation (compute_type) | weight # static quint8 qint8 # dynamic float32 (quint8) qint8 # weight_only float32 float16 # tuple (activation_dtype, weight_dtype, compute_dtype) supported_dtypes = [ (torch.quint8, torch.qint8, None), (torch.float32, torch.qint8, torch.quint8), (torch.float32, torch.float16, None), # static float16 quantization (torch.float16, torch.float16, None), ] dtypes = get_qconfig_dtypes(qconfig) # leave the op unquantized if the dtype combination is not supported if not is_reference and dtypes not in supported_dtypes: warnings.warn( "dtype combination: {} is not " "supported by Linear " "supported dtype combinations are: {}".format(dtypes, supported_dtypes)) if self.relu_node: op_out = quantized_graph.node_copy(self.linear_node, load_arg(quantized=torch.float)) relu_args = [op_out] relu_args.extend(load_arg(quantized=torch.float)(self.relu_node.args[1:])) relu_kwargs = load_arg(quantized=torch.float)(self.relu_node.kwargs) return create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.relu, tuple(relu_args), relu_kwargs), self.relu_node) else: return quantized_graph.node_copy(node, load_arg(quantized=None)) activation_int8_quantized = activation_is_int8_quantized(qconfig) activation_statically_quantized = activation_is_statically_quantized(qconfig) weight_dtype = dtypes[1] if self.linear_node.op == 'call_module': output_activation_post_process = \ self._maybe_get_last_node_only_observer(modules) # note that relu should already be fused into linear modul in the fusion step assert self.relu_node is None, 'linear module and relu fusion is not executed, ' \ 'please make sure to run fusion before prepare' # we'll always produce reference pattern for the following modules # will remove the else branch after we migrated all use cases module_allowlist = [ torch.nn.Linear, torch.nn.qat.Linear, torch.nn.intrinsic.LinearReLU, torch.nn.intrinsic.qat.LinearReLU, torch.nn.intrinsic.LinearBn1d, torch.nn.intrinsic.qat.LinearBn1d, ] if is_reference or type(self.linear) in module_allowlist and dtypes in [(torch.quint8, torch.qint8, None)]: # produce dequant - float_op - quant pattern dtype = torch.float if activation_int8_quantized: dtype = activation_dtype(qconfig) activation = load_arg(quantized=dtype)(self.linear_node.args[0]) args = load_arg(quantized=torch.float)(self.linear_node.args) # Get the float linear and attach qscheme and qparams the the module float_linear = self.linear fused_linear = None qat_modules = ( torch.nn.qat.Linear, torch.nn.intrinsic.qat.LinearReLU, torch.nn.intrinsic.qat.LinearBn1d, ) static_fused_modules = ( torch.nn.intrinsic.LinearReLU, torch.nn.intrinsic.LinearBn1d, ) if isinstance(float_linear, qat_modules): float_linear = float_linear.to_float() # change qat linear to linear parent_name, name = _parent_name(self.linear_node.target) setattr(modules[parent_name], name, float_linear) # Attach weight fake quant to the linear module if isinstance(float_linear, static_fused_modules): fused_linear = float_linear float_linear = float_linear[0] weight_post_process = self.linear.weight_fake_quant else: if isinstance(float_linear, static_fused_modules): fused_linear = float_linear float_linear = self.linear[0] # type: ignore[index] # Attach the weight observer to the module weight_post_process = qconfig.weight() # type: ignore[union-attr] # Run weight observer # TODO: This is currently a hack for QAT to get the right shapes for scale and zero point. # In the future, we should require the user to calibrate the model after calling prepare weight_post_process(float_linear.weight) # type: ignore[operator] weight_qparams = get_qparam_dict(weight_post_process) # TODO: include the configuration in backend_config_dict # we can have a map from module to reference module # and allow user to register new ones qlinear_cls = get_static_quant_module_class( type(float_linear), is_reference=True) ref_linear = qlinear_cls.from_float(float_linear, weight_qparams) # if the parent is a fused linear (Sequential), we can replace the first # item to ref linear, otherwise we can update # the linear instance in the module tree if fused_linear is not None: fused_linear[0] = ref_linear else: parent_name, name = _parent_name(self.linear_node.target) setattr(modules[parent_name], name, ref_linear) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ('call_module', self.linear_node.target, args, {}), self.linear_node) if output_activation_post_process: op_out = quantize_node( op_out, output_activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) return op_out # non-reference option else: # 1. attach output activation post process to linear module if output_activation_post_process: self.linear.activation_post_process = output_activation_post_process # 2. select corresponding quantized linear class for the float linear class if activation_int8_quantized: additional_static_quant_mapping = convert_custom_config_dict.get("static", {}) qlinear = get_static_quant_module_class( type(self.linear), additional_static_quant_mapping) else: assert dtypes in [ (torch.float32, torch.qint8, torch.quint8), (torch.float32, torch.float16, None), ], f"dtype {dtypes} not supported yet" additional_dynamic_quant_mapping = convert_custom_config_dict.get("dynamic", {}) qlinear = get_dynamic_quant_module_class(type(self.linear), additional_dynamic_quant_mapping) quantized = qlinear.from_float(self.linear) parent_name, name = _parent_name(self.linear_node.target) setattr(modules[parent_name], name, quantized) # activation needs to be quantized for static quantization dtype = torch.float if activation_int8_quantized: dtype = activation_dtype(qconfig) return create_node_from_old_node_preserve_meta( quantized_graph, ( 'call_module', self.linear_node.target, (load_arg(quantized=dtype)(self.linear_node.args[0]),), {}, ), self.linear_node) else: # call_function assert self.linear_node.op == 'call_function' if is_reference or self.linear_node.target == torch.nn.functional.linear and\ dtypes in [(torch.quint8, torch.qint8, None)]: quantized_input_dtypes = [torch.float, torch.float] if activation_int8_quantized: quantized_input_dtypes[0] = torch.quint8 if weight_is_statically_quantized(qconfig): quantized_input_dtypes[1] = torch.qint8 args = load_arg(quantized=quantized_input_dtypes)(self.linear_node.args) args = load_arg(quantized=torch.float)(self.linear_node.args) kwargs = load_arg(quantized=torch.float)(self.linear_node.kwargs) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.linear, args, kwargs), self.linear_node) if self.relu_node: relu_args = [op_out] relu_args.extend(load_arg(quantized=torch.float)(self.relu_node.args[1:])) relu_kwargs = load_arg(quantized=torch.float)(self.relu_node.kwargs) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.relu, tuple(relu_args), relu_kwargs), self.relu_node) if activation_statically_quantized: # quantize output for statically quantized linear op root_module = modules[''] act_post_process_name = self.relu_node.name if self.relu_node else self.linear_node.name act_post_process_node = self.relu_node if self.relu_node else self.linear_node activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None return quantize_node( op_out, activation_post_process, act_post_process_node, modules, quantized_graph, node_name_to_scope, is_input=False, output_prefix="") else: # output for dynamically quantized linear op is not quantized return op_out else: # non-reference option # prepacking weights for static int8 quant and dynamic quant if dtypes != (torch.float16, torch.float16, None): # linear args # (x, weight, bias, ...) # TODO: the name should be weight is int8 quantized weight_quantized = weight_is_statically_quantized(qconfig) dtype = weight_dtype if weight_quantized else torch.float linear_weight = load_arg(quantized=dtype)(self.linear_node.args[1]) # get other arguments kwargs = {**load_arg(quantized=torch.float)(self.linear_node.kwargs)} # all args after bias, including bias other_args = load_arg(quantized=torch.float)(self.linear_node.args[2:]) # bias might be either positional, or a keyword argument if len(self.linear_node.args) > 2: bias = load_arg(quantized=torch.float)(self.linear_node.args[2]) other_args = other_args[1:] # remove the bias argument else: bias = kwargs.pop('bias', None) prepack_args = (linear_weight, bias) prepack_op = get_linear_prepack_op_for_dtype(weight_dtype) packed_weight = quantized_graph.create_node( 'call_function', prepack_op, prepack_args, {}) # construct linear input if activation_int8_quantized: qlinear_op = torch.ops.quantized.linear_relu if self.relu_node else torch.ops.quantized.linear linear_input = load_arg(quantized=torch.quint8)(self.linear_node.args[0]) activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None scale, zero_point, _ = get_per_tensor_qparams(activation_post_process) scale_node, zero_point_node = \ create_qparam_nodes( self.linear_node.name, scale, zero_point, modules, quantized_graph, node_name_to_scope) qlinear_args = (linear_input, packed_weight, scale_node, zero_point_node) op = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", qlinear_op, qlinear_args, kwargs), self.linear_node) # Store the name of the fused op to get the path of node after fusion as well. # TODO: may need to change the key to Node regenerate the map in each transformation, # since we might not be able to rely on the name node_name_to_scope[op.name] = node_name_to_scope[self.linear_node.name] return op elif dtypes in [(torch.float32, torch.qint8, torch.quint8), (torch.float32, torch.float16, None)]: # choose linear dynamic or linear dynamic fp16 op based on weight dtype if weight_dtype == torch.qint8: if self.relu_node: qlinear_op = torch.ops.quantized.linear_relu_dynamic else: qlinear_op = torch.ops.quantized.linear_dynamic else: if self.relu_node: qlinear_op = torch.ops.quantized.linear_relu_dynamic_fp16 else: qlinear_op = torch.ops.quantized.linear_dynamic_fp16 linear_input = load_arg(quantized=torch.float)(self.linear_node.args[0]) qlinear_args = (linear_input, packed_weight) # type: ignore[assignment] op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", qlinear_op, qlinear_args, kwargs), self.linear_node) # Store the name of the dynamic op to get the path of node after replacement as well. # TODO: may need to change the key to Node regenerate the map in each transformation, # since we might not be able to rely on the name node_name_to_scope[op_out.name] = node_name_to_scope[self.linear_node.name] return op_out else: assert dtypes == (torch.float16, torch.float16, None) # TODO (refactor) this is duplicated, maybe have a helper function if self.relu_node: op_out = quantized_graph.node_copy(self.linear_node, load_arg(quantized=torch.float)) relu_args = [op_out] relu_args.extend(load_arg(quantized=torch.float)(self.relu_node.args[1:])) relu_kwargs = load_arg(quantized=torch.float)(self.relu_node.kwargs) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_function", torch.nn.functional.relu, tuple(relu_args), relu_kwargs), self.relu_node) else: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return quantized_graph.create_node( "call_method", "to", (op_out, torch.float16), {}) @register_quant_pattern(torch.nn.BatchNorm2d) @register_quant_pattern(torch.nn.BatchNorm3d) @register_quant_pattern(torch.nn.intrinsic.BNReLU2d) @register_quant_pattern(torch.nn.intrinsic.BNReLU3d) class BatchNormQuantizeHandler(QuantizeHandler): def __init__( self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) assert node.op == 'call_module' self.bn_node = node self.bn = modules[str(self.bn_node.target)] def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: if convert_custom_config_dict is None: convert_custom_config_dict = {} additional_static_quant_mapping = convert_custom_config_dict.get("static", {}) # 1. attach activation post process to module output_activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert output_activation_post_process is not None if is_reference: # produce dequant - float_op - quant pattern dtype = activation_dtype(qconfig) activation = load_arg(quantized=dtype)(self.bn_node.args[0]) args = load_arg(quantized=torch.float)(self.bn_node.args) op_out = create_node_from_old_node_preserve_meta( quantized_graph, ("call_module", self.bn_node.target, args, {}), self.bn_node) if output_activation_post_process: op_out = quantize_node( op_out, output_activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) return op_out else: self.bn.activation_post_process = output_activation_post_process qbn_cls = get_static_quant_module_class(type(self.bn), additional_static_quant_mapping) quantized = qbn_cls.from_float(self.bn) parent_name, name = _parent_name(self.bn_node.target) setattr(modules[parent_name], name, quantized) return create_node_from_old_node_preserve_meta( quantized_graph, ( 'call_module', self.bn_node.target, load_arg(quantized=[0])(self.bn_node.args), load_arg(quantized=torch.float)(self.bn_node.kwargs), ), self.bn_node) @register_quant_pattern(torch.nn.qat.Embedding) @register_quant_pattern(torch.nn.qat.EmbeddingBag) @register_quant_pattern(torch.nn.Embedding) @register_quant_pattern(torch.nn.EmbeddingBag) class EmbeddingQuantizeHandler(QuantizeHandler): def __init__( self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) def input_output_observed(self) -> bool: return False def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: # Supported combinations are: # quant_type | activation | weight | activation_compute_type # weight_only | float32 | quint8 | None # weight_only | float32 | quint4x2 | None # tuple (activation_dtype, weight_dtype, compute_dtype) supported_dtypes = [ (torch.float32, torch.quint8, None), (torch.float32, torch.quint4x2, None), ] assert node.op == 'call_module' emb_node = node dtypes = get_qconfig_dtypes(qconfig) # leave the op unquantized if the dtype combination is not supported if dtypes not in supported_dtypes: warnings.warn( "dtype combination: {} is not " "supported by Embedding/EmbeddingBag, " "supported dtype combinations are: {}".format(dtypes, supported_dtypes)) return quantized_graph.node_copy(node, load_arg(quantized=None)) emb = modules[str(emb_node.target)] qemb = get_static_quant_module_class(type(emb)) quantized = qemb.from_float(emb) parent_name, name = _parent_name(emb_node.target) setattr(modules[parent_name], name, quantized) return create_node_from_old_node_preserve_meta( quantized_graph, ( 'call_module', emb_node.target, load_arg(quantized=torch.float)(emb_node.args), load_arg(quantized=torch.float)(emb_node.kwargs), ), emb_node) # TODO (maybe): merge with embedding quantize handler @register_quant_pattern(torch.nn.GRUCell) @register_quant_pattern(torch.nn.LSTMCell) @register_quant_pattern(torch.nn.RNNCell) @register_quant_pattern(torch.nn.LSTM) class RNNDynamicQuantizeHandler(QuantizeHandler): def __init__( self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: # Supported combinations are: # quant_type | activation | weight | activation_compute_type # dynamic | float32 | qint8 | quint8 # dynamic | float32 | float16 | None # tuple (activation_dtype, weight_dtype, compute_dtype) supported_dtypes = [ (torch.float32, torch.qint8, torch.quint8), (torch.float32, torch.float16, None), ] assert node.op == 'call_module' dtypes = get_qconfig_dtypes(qconfig) # leave the op unquantized if the dtype combination is not supported if dtypes not in supported_dtypes: warnings.warn( "dtype combination: {} is not " "supported by Embedding/EmbeddingBag, " "supported dtype combinations are: {}".format(dtypes, supported_dtypes)) return quantized_graph.node_copy(node, load_arg(quantized=None)) act_dtype, weight_dtype, compute_dtype = dtypes activation = load_arg(quantized=act_dtype)(node.args[0]) module = modules[str(node.target)] qmodule_cls = get_dynamic_quant_module_class(type(module)) qmodule = qmodule_cls.from_float(module) parent_name, name = _parent_name(node.target) setattr(modules[parent_name], name, qmodule) return create_node_from_old_node_preserve_meta( quantized_graph, ( 'call_module', node.target, load_arg(quantized=torch.float)(node.args), load_arg(quantized=torch.float)(node.kwargs), ), node) ARGS_TO_SKIP = { torch._ops.ops.quantized.hardswish: ['inplace'], torch._ops.ops.quantized.elu: ['inplace'], torch._ops.ops.quantized.dropout: ['inplace'], torch._ops.ops.quantized.instance_norm: ['running_mean', 'running_var', 'use_input_stats', 'momentum'], } @register_quant_pattern(torch.nn.ConvTranspose1d) @register_quant_pattern(torch.nn.ConvTranspose2d) @register_quant_pattern(torch.nn.ELU) @register_quant_pattern(torch.nn.LeakyReLU) @register_quant_pattern(torch.nn.Hardswish) @register_quant_pattern(torch.nn.InstanceNorm1d) @register_quant_pattern(torch.nn.InstanceNorm2d) @register_quant_pattern(torch.nn.InstanceNorm3d) @register_quant_pattern(torch.nn.LayerNorm) @register_quant_pattern(torch.nn.SiLU) @register_quant_pattern(torch.nn.Mish) @register_quant_pattern(torch.nn.Dropout) # we currently only support reference patterns for these ops so they have been removed # until they receive a proper fp16 kernel. To use the reference pattern, use a custom qconfig # @register_quant_pattern(torch.nn.GELU) # @register_quant_pattern(torch.nn.Softmax) @register_quant_pattern(torch.nn.functional.elu) @register_quant_pattern(torch.nn.functional.hardswish) @register_quant_pattern(torch.nn.functional.instance_norm) @register_quant_pattern(torch.nn.functional.layer_norm) @register_quant_pattern(torch.nn.functional.leaky_relu) @register_quant_pattern(torch.nn.functional.silu) @register_quant_pattern(torch.nn.functional.mish) @register_quant_pattern(torch.nn.functional.dropout) # we currently only support reference patterns for these ops so they have been removed # until they receive a proper fp16 kernel. To use the reference pattern, use a custom qconfig # @register_quant_pattern(torch.nn.functional.gelu) # @register_quant_pattern(torch.nn.functional.softmax) @register_quant_pattern(torch.sum) class DefaultNodeQuantizeHandler(QuantizeHandler): """ Common quantized op, first input and first output will be quantized """ def __init__( self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) if node.op == "call_function" or node.op == "call_method": self.op = node.target elif node.op == "call_module": self.op = type(modules[str(node.target)]) def is_output_quantized(self, qconfig): dtypes = get_qconfig_dtypes(qconfig) return self.op in default_op_supported_dtypes and \ dtypes in default_op_supported_dtypes[self.op] def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: if not self.all_node_args_are_tensors: return NotImplemented assert node.op in ['call_module', 'call_function'], 'Only call_module and ' + \ 'call_function are handled in DefaultNode' if convert_custom_config_dict is None: convert_custom_config_dict = {} additional_static_quant_mapping = convert_custom_config_dict.get("static", {}) dtypes = get_qconfig_dtypes(qconfig) if not is_reference and dtypes not in default_op_supported_dtypes[self.op]: warnings.warn( "dtype combination: {} is not " "supported by {} " "supported dtype combinations are: {}".format(dtypes, self.op, default_op_supported_dtypes[self.op])) return quantized_graph.node_copy(node, load_arg(quantized=torch.float)) # We can produce reference for a dtypes including # (torch.quint8, torch.qint8, torch.qint32, torch.float16) act_dtype = activation_dtype(qconfig) if act_dtype == torch.float: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return op_out else: activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None # make sure the input is quantized to act_dtype load_arg(quantized={0: act_dtype})(node.args) args = load_arg(quantized=torch.float)(node.args) kwargs = load_arg(quantized=torch.float)(node.kwargs) # swap float module to reference module (ConvTranspose) float_module = modules[str(node.target)] if node.op == "call_module" else None if type(float_module) in [torch.nn.ConvTranspose1d, torch.nn.ConvTranspose2d]: ref_module_cls = get_static_quant_module_class(type(float_module), is_reference=True) weight_post_process = qconfig.weight() # type: ignore[union-attr] weight_post_process(float_module.weight) # type: ignore[union-attr] weight_qparams = get_qparam_dict(weight_post_process) ref_module = ref_module_cls.from_float(float_module, weight_qparams) # type: ignore[attr-defined] parent_name, name = _parent_name(node.target) setattr(modules[parent_name], name, ref_module) op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return quantize_node( op_out, activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) @register_quant_pattern(torch.nn.Hardsigmoid, default_affine_fixed_qparams_observer) @register_quant_pattern(torch.nn.functional.hardsigmoid, default_affine_fixed_qparams_observer) @register_quant_pattern('hardsigmoid', default_affine_fixed_qparams_observer) @register_quant_pattern('hardsigmoid_', default_affine_fixed_qparams_observer) @register_quant_pattern(torch.nn.Sigmoid, default_affine_fixed_qparams_observer) @register_quant_pattern(torch.sigmoid, default_affine_fixed_qparams_observer) @register_quant_pattern('sigmoid', default_affine_fixed_qparams_observer) @register_quant_pattern('sigmoid_', default_affine_fixed_qparams_observer) @register_quant_pattern(torch.nn.Tanh, default_symmetric_fixed_qparams_observer) @register_quant_pattern(torch.tanh, default_symmetric_fixed_qparams_observer) @register_quant_pattern('tanh', default_symmetric_fixed_qparams_observer) @register_quant_pattern('tanh_', default_symmetric_fixed_qparams_observer) class FixedQParamsOpQuantizeHandler(QuantizeHandler): def __init__(self, node: Node, modules: Dict[str, torch.nn.Module]): super().__init__(node, modules) self.node = node def should_mark_output_quantized_from_input_quantized_status( self, qconfig: QConfigAny ) -> bool: # FixQParamOps are the same as CopyNode in int8 quantization return activation_dtype(qconfig) in [torch.quint8, torch.qint8] # some qhandlers override the activations constructor def get_activation_ctr(self, qconfig, pattern, is_training) -> Optional[Callable]: act_dtype = activation_dtype(qconfig) if act_dtype == torch.quint8: return get_default_output_activation_post_process_map(is_training).get( pattern, qconfig.activation) else: return qconfig.activation def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: act_dtype = activation_dtype(qconfig) if act_dtype == torch.float: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return op_out else: activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None # make sure the input is quantized to act_dtype load_arg(quantized={0: act_dtype})(node.args) args = load_arg(quantized=torch.float)(node.args) kwargs = load_arg(quantized=torch.float)(node.kwargs) op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return quantize_node( op_out, activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) @register_quant_pattern(torch.nn.AdaptiveAvgPool1d) @register_quant_pattern(torch.nn.AdaptiveAvgPool2d) @register_quant_pattern(torch.nn.AdaptiveAvgPool3d) @register_quant_pattern(torch.nn.AvgPool1d) @register_quant_pattern(torch.nn.AvgPool2d) @register_quant_pattern(torch.nn.AvgPool3d) @register_quant_pattern(torch.nn.Hardtanh) @register_quant_pattern(torch.nn.MaxPool1d) @register_quant_pattern(torch.nn.MaxPool2d) @register_quant_pattern(torch.nn.MaxPool3d) @register_quant_pattern(torch.nn.ReLU) @register_quant_pattern(torch.nn.ReLU6) @register_quant_pattern(torch.adaptive_avg_pool1d) @register_quant_pattern(torch.nn.functional.adaptive_avg_pool2d) @register_quant_pattern(torch.nn.functional.adaptive_avg_pool3d) @register_quant_pattern(torch.nn.functional.hardtanh) @register_quant_pattern(torch.nn.functional.hardtanh_) @register_quant_pattern(torch.nn.functional.interpolate) @register_quant_pattern(torch.nn.functional.max_pool1d) @register_quant_pattern(torch.nn.functional.max_pool2d) @register_quant_pattern(torch.nn.functional.max_pool3d) @register_quant_pattern(torch.nn.functional.relu) @register_quant_pattern(torch.nn.functional.relu6) @register_quant_pattern(torch.avg_pool1d) @register_quant_pattern(torch._C._nn.avg_pool2d) @register_quant_pattern(torch._C._nn.avg_pool3d) @register_quant_pattern(torch.clamp) @register_quant_pattern(torch.flatten) @register_quant_pattern(torch.mean) @register_quant_pattern(operator.floordiv) @register_quant_pattern('clamp') @register_quant_pattern('mean') @register_quant_pattern('relu') @register_quant_pattern('relu_') class CopyNodeQuantizeHandler(QuantizeHandler): """ Operators that works on both float and quantized input if input is quantized, the output Tensor shares the same quantization parameter with input. These ops will do computation on the input Tensor, e.g. average pool, so we will insert extra observer/fake_quant for the output of these operators. TODO: maybe rename this to TensorValueOpQuantizeHandler """ def should_mark_output_quantized_from_input_quantized_status( self, qconfig: QConfigAny ) -> bool: return True def is_general_tensor_value_op(self) -> bool: return True def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: # when activation dtype is torch.float, the node does not require # observation # e.g. dynamic quantization or weight_only quantization act_dtype = activation_dtype(qconfig) if act_dtype == torch.float: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return op_out else: activation_post_process = \ self._maybe_get_last_node_only_observer(modules) if activation_post_process is not None: # make sure the input is quantized to act_dtype load_arg(quantized={0: act_dtype})(node.args) args = list(load_arg(quantized=torch.float)(node.args)) kwargs = load_arg(quantized=torch.float)(node.kwargs) op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return quantize_node( op_out, activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) else: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return op_out class CustomModuleQuantizeHandler(QuantizeHandler): def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: """ Convert a float custom module to quantized custom module """ assert node.op == 'call_module' assert convert_custom_config_dict is not None custom_module_class_mapping = convert_custom_config_dict.get("observed_to_quantized_custom_module_class", None) assert custom_module_class_mapping is not None observed_custom_module = modules[str(node.target)] if activation_is_statically_quantized(qconfig): activation_post_process = \ self._maybe_get_last_node_only_observer(modules) assert activation_post_process is not None observed_custom_module.activation_post_process = activation_post_process quantized_custom_module_class = get_swapped_custom_module_class( observed_custom_module, custom_module_class_mapping, qconfig) quantized_custom_module = \ quantized_custom_module_class.from_observed(observed_custom_module) parent_name, name = _parent_name(node.target) setattr(modules[parent_name], name, quantized_custom_module) # hardcoded the quntized input to be None (take whatever is in the environemnt), # we can extend this # if there is a need, e.g. get the indexes of quantized inputs from some # module attribute like module._QUANTIZED_INPUT_INDEXES return quantized_graph.node_copy(node, load_arg(quantized=None)) @register_quant_pattern(torch.nn.Identity) @register_quant_pattern(torch.transpose) @register_quant_pattern(torch.repeat_interleave) @register_quant_pattern(torch.squeeze) @register_quant_pattern(torch.stack) @register_quant_pattern(torch.unsqueeze) @register_quant_pattern('contiguous') @register_quant_pattern('detach') @register_quant_pattern('detach_') @register_quant_pattern('permute') @register_quant_pattern('repeat') @register_quant_pattern('repeat_interleave') @register_quant_pattern('reshape') @register_quant_pattern('resize_') @register_quant_pattern('shape') @register_quant_pattern('size') @register_quant_pattern('squeeze') @register_quant_pattern('squeeze_') @register_quant_pattern('transpose') @register_quant_pattern('unsqueeze') @register_quant_pattern('unsqueeze_') @register_quant_pattern('view') class GeneralTensorShapeOpQuantizeHandler(QuantizeHandler): """ Operators that works on both float and quantized input if input is quantized, the output Tensor shares the same quantization parameter with input. These ops only do rearrangement of Tensor values, for example reshape, or just query the information about Tensor e.g. size, and we do not insert extra observer/fake_quant for the output of the operator. """ def is_general_tensor_shape_op(self) -> bool: return True def should_mark_output_quantized_from_input_quantized_status( self, qconfig: QConfigAny ) -> bool: return True def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: # when activation dtype is torch.float, the node does not require # observation # e.g. dynamic quantization or weight_only quantization act_dtype = activation_dtype(qconfig) if act_dtype == torch.float: op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return op_out else: activation_post_process = \ self._maybe_get_last_node_only_observer(modules) if activation_post_process is not None: args = list(load_arg(quantized=torch.float)(node.args)) kwargs = load_arg(quantized=torch.float)(node.kwargs) op_out = quantized_graph.node_copy(node, load_arg(quantized=torch.float)) return quantize_node( op_out, activation_post_process, node, modules, quantized_graph, node_name_to_scope, is_input=False) else: return quantized_graph.node_copy(node, load_arg(quantized=torch.float)) class StandaloneModuleQuantizeHandler(QuantizeHandler): """ Converts an observed standalone module to quantized standalone module by calling convert_fx on the observed standalone module. """ def convert(self, node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], quantized_graph: Graph, node_name_to_scope: Dict[str, Tuple[str, type]], load_arg: Callable, is_reference: bool = False, convert_custom_config_dict: Dict[str, Any] = None) -> Node: assert node.op == 'call_module' convert = torch.ao.quantization.quantize_fx._convert_standalone_module_fx # type: ignore[attr-defined] # We know that observed standalone module is a GraphModule since # it's produced by us observed_standalone_module : GraphModule = modules[str(node.target)] # type: ignore[assignment] input_quantized_idxs = observed_standalone_module._standalone_module_input_quantized_idxs.tolist() # type: ignore[operator] quantized_standalone_module = convert(observed_standalone_module, is_reference=is_reference) parent_name, name = _parent_name(node.target) # update the modules dict setattr(modules[parent_name], name, quantized_standalone_module) modules[str(node.target)] = quantized_standalone_module return quantized_graph.node_copy(node, load_arg(quantized=input_quantized_idxs))