from collections import namedtuple from typing import List, Dict, Any import operator import torch from torch.ao.quantization.backend_config.observation_type import ObservationType import torch.nn.functional as F import torch.nn as nn import torch.nn.intrinsic as nni import torch.nn.intrinsic.qat as nniqat import torch.nn.qat as nnqat import torch.nn.quantized._reference as nnqr from ..observer import ( default_fixed_qparams_range_0to1_observer, default_fixed_qparams_range_neg1to1_observer, ) from ..fake_quantize import FixedQParamsFakeQuantize from ..fuser_method_mappings import ( reverse_sequential_wrapper2, reverse2, reverse3, fuse_conv_bn, fuse_conv_bn_relu, fuse_linear_bn, fuse_convtranspose_bn, ) # TODO: rename to be more explict, e.g. qat_conv_relu _ConvMetadata = namedtuple( "_ConvMetadata", ["root", "transpose", "bn", "reference", "transpose_reference", "fused_conv_relu", "fused_conv_bn", "fused_conv_bn_relu", "qat", "relu_qat", "bn_qat", "bn_relu_qat", "func"]) _Conv1dMetadata = _ConvMetadata( nn.Conv1d, nn.ConvTranspose1d, nn.BatchNorm1d, nnqr.Conv1d, nnqr.ConvTranspose1d, nni.ConvReLU1d, nni.ConvBn1d, nni.ConvBnReLU1d, nnqat.Conv1d, nniqat.ConvReLU1d, nniqat.ConvBn1d, nniqat.ConvBnReLU1d, F.conv1d) _Conv2dMetadata = _ConvMetadata( nn.Conv2d, nn.ConvTranspose2d, nn.BatchNorm2d, nnqr.Conv2d, nnqr.ConvTranspose2d, nni.ConvReLU2d, nni.ConvBn2d, nni.ConvBnReLU2d, nnqat.Conv2d, nniqat.ConvReLU2d, nniqat.ConvBn2d, nniqat.ConvBnReLU2d, F.conv2d) _Conv3dMetadata = _ConvMetadata( nn.Conv3d, nn.ConvTranspose3d, nn.BatchNorm3d, nnqr.Conv3d, nnqr.ConvTranspose3d, nni.ConvReLU3d, nni.ConvBn3d, nni.ConvBnReLU3d, nnqat.Conv3d, nniqat.ConvReLU3d, nniqat.ConvBn3d, nniqat.ConvBnReLU3d, F.conv3d) # =================== # | DTYPE CONFIGS | # =================== # weighted op int8 dtype config # this is config for ops that has quantized weights, like linear, conv weighted_op_int8_dtype_config = { # optional, input activation dtype "input_dtype": torch.quint8, # optional, weight dtype "weight_dtype": torch.qint8, # optional, bias dtype "bias_dtype": torch.float, # optional, output activation dtype "output_dtype": torch.quint8 } default_op_quint8_dtype_config = { # optional, input activation dtype "input_dtype": torch.quint8, # optional, output activation dtype "output_dtype": torch.quint8, } default_op_fp16_dtype_config = { # optional, input activation dtype "input_dtype": torch.float16, # optional, weight dtype "weight_dtype": torch.float16, # optional, bias dtype "bias_dtype": torch.float16, # optional, output activation dtype "output_dtype": torch.float16, } default_dynamic_int8_dtype_config = { "input_dtype": torch.quint8, "weight_dtype": torch.qint8, "bias_dtype": torch.float, "output_dtype": torch.float, # currently the dtype check is not yet enabled, so we provided the dtype_configs but # it is not really used yet, # we will enable it a bit later after we moved everything to backend_config_dict "is_dynamic": True, } default_dynamic_float16_dtype_config = { "input_dtype": torch.float16, "weight_dtype": torch.float16, "bias_dtype": torch.float, "output_dtype": torch.float, # currently the dtype check is not yet enabled, so we provided the dtype_configs but # it is not really used yet, # we will enable it a bit later after we moved everything to backend_config_dict "is_dynamic": True, } weight_only_quint8_dtype_config = { "input_dtype": torch.float, "weight_dtype": torch.quint8, "output_dtype": torch.float, } weight_only_quint4x2_dtype_config = { "input_dtype": torch.float, "weight_dtype": torch.quint4x2, "output_dtype": torch.float, } # ====================== # | OPERATOR CONFIGS | # ====================== def _get_default_op_backend_config(op, dtype_configs): return { "pattern": op, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, "dtype_configs": dtype_configs, } _DEFAULT_OP_INT8_CONFIGS = [ _get_default_op_backend_config(op, [default_op_quint8_dtype_config]) for op in [ torch.nn.ELU, torch.nn.LeakyReLU, torch.nn.Hardswish, torch.nn.InstanceNorm1d, torch.nn.InstanceNorm2d, torch.nn.InstanceNorm3d, torch.nn.LayerNorm, torch.nn.Dropout, torch.nn.functional.elu, torch.nn.functional.hardswish, torch.nn.functional.instance_norm, torch.nn.functional.leaky_relu, torch.nn.functional.dropout, torch.nn.functional.layer_norm, ]] def _get_linear_configs(dtype_configs): """ Return all configs related to linear modules and ops. """ observation_type = ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT linear_configs = [] # (1) Single linear modules/functions # ------------------------------------- # linear module linear_configs.append({ # Please see README under this folder for pattern format "pattern": torch.nn.Linear, "observation_type": observation_type, "dtype_configs": dtype_configs, # the root module for the pattern, used to query the reference quantized module # e.g. for a (torch.nn.ReLU, torch.nn.Linear) pattern, the root will be torch.nn.Linear "root_module": torch.nn.Linear, # the corresponding reference quantized module for the root module "reference_quantized_module_for_root": nnqr.Linear, "qat_module": nnqat.Linear, }) # linear qat module linear_configs.append({ "pattern": nnqat.Linear, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": torch.nn.Linear, "reference_quantized_module_for_root": nnqr.Linear, }) # functional linear linear_configs.append({ "pattern": torch.nn.functional.linear, "observation_type": observation_type, "dtype_configs": dtype_configs, }) # (2) Linear + relu # ------------------- # 2.1 linear module + relu fusion config # linear relu, linear module + relu module linear_configs.append({ "pattern": (torch.nn.ReLU, torch.nn.Linear), "dtype_configs": dtype_configs, "fuser_method": reverse_sequential_wrapper2(nni.LinearReLU), "fused_module": nni.LinearReLU, }) # linear relu, linear module + functional relu linear_configs.append({ "pattern": (torch.nn.functional.relu, torch.nn.Linear), "dtype_configs": dtype_configs, "fuser_method": reverse_sequential_wrapper2(nni.LinearReLU), "fused_module": nni.LinearReLU, }) # 2.2 linear module + relu, fused module configs # linear relu, fused module linear_configs.append({ "pattern": nni.LinearReLU, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": torch.nn.Linear, "reference_quantized_module_for_root": nnqr.Linear, "qat_module": nniqat.LinearReLU, }) # linear relu, qat fused module linear_configs.append({ "pattern": nniqat.LinearReLU, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": torch.nn.Linear, "reference_quantized_module_for_root": nnqr.Linear, }) # 2.3 functional linear + relu configs # linear relu, functional linear + relu module linear_configs.append({ "pattern": (torch.nn.ReLU, F.linear), "observation_type": observation_type, "dtype_configs": dtype_configs, }) # linear relu, functional linear + functional relu linear_configs.append({ "pattern": (F.relu, F.linear), "observation_type": observation_type, "dtype_configs": dtype_configs, }) # (3) Linear + batchnorm # ------------------------ # 3.1 linear bn fusion linear_configs.append({ "pattern": (nn.BatchNorm1d, nn.Linear), "dtype_configs": dtype_configs, "fuser_method": reverse2(fuse_linear_bn), "fused_module": nni.LinearBn1d, }) # 3.2 linear bn fused # linear bn, fused module linear_configs.append({ "pattern": nni.LinearBn1d, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": torch.nn.Linear, "reference_quantized_module_for_root": nnqr.Linear, "qat_module": nniqat.LinearBn1d, }) # linear bn, qat fused module linear_configs.append({ "pattern": nniqat.LinearBn1d, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": torch.nn.Linear, "reference_quantized_module_for_root": nnqr.Linear, }) return linear_configs def _get_conv_configs(): """ Return all configs related to conv modules and ops. """ conv_configs = [] observation_type = ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT dtype_configs = [weighted_op_int8_dtype_config] for convs in [_Conv1dMetadata, _Conv2dMetadata, _Conv3dMetadata]: # (1) Single conv modules/functions # ----------------------------------- # conv module conv_configs.append({ "pattern": convs.root, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, "qat_module": convs.qat, }) # conv qat module conv_configs.append({ "pattern": convs.qat, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, }) # functional conv conv_configs.append({ "pattern": convs.func, "observation_type": observation_type, "dtype_configs": dtype_configs, }) # (2) Conv + relu # ----------------- # 2.1 conv module + relu fusion configs # conv relu fusion, conv module + relu module conv_configs.append({ "pattern": (torch.nn.ReLU, convs.root), "dtype_configs": dtype_configs, "fuser_method": reverse_sequential_wrapper2(convs.fused_conv_relu), "fused_module": convs.fused_conv_relu, }) # conv relu fusion, conv module + functional relu conv_configs.append({ "pattern": (F.relu, convs.root), "dtype_configs": dtype_configs, "fuser_method": reverse_sequential_wrapper2(convs.fused_conv_relu), "fused_module": convs.fused_conv_relu, }) # 2.2 conv module + relu fused module configs # conv relu, fused module conv_configs.append({ "pattern": convs.fused_conv_relu, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, "qat_module": convs.relu_qat, }) # conv relu, qat fused module conv_configs.append({ "pattern": convs.relu_qat, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, }) # 2.3 functional conv + relu configs # conv relu, functional conv + relu module conv_configs.append({ "pattern": (torch.nn.ReLU, convs.func), "observation_type": observation_type, "dtype_configs": dtype_configs, }) # conv relu, functional conv + functional relu conv_configs.append({ "pattern": (F.relu, convs.func), "observation_type": observation_type, "dtype_configs": dtype_configs, }) # fused conv relu conv_configs.append({ "pattern": convs.fused_conv_relu, "dtype_configs": dtype_configs, "qat_module": convs.relu_qat, }) conv_configs.append({ "pattern": convs.relu_qat, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, }) # (3) Conv + batchnorm (+ relu) # ------------------------------- # 3.1 conv bn fusion configs # conv + bn fusion conv_configs.append({ "pattern": (convs.bn, convs.root), "dtype_configs": dtype_configs, "fuser_method": reverse2(fuse_conv_bn), "fused_module": convs.fused_conv_bn, }) # conv + bn + relu module fusion conv_configs.append({ "pattern": (nn.ReLU, (convs.bn, convs.root)), "dtype_configs": dtype_configs, "fuser_method": reverse3(fuse_conv_bn_relu), "fused_module": convs.fused_conv_bn_relu, }) # conv + bn + relu functional fusion conv_configs.append({ "pattern": (F.relu, (convs.bn, convs.root)), "dtype_configs": dtype_configs, "root_module": convs.root, "fuser_method": reverse3(fuse_conv_bn_relu), "fused_module": convs.fused_conv_bn_relu, }) # TODO: we can add fusion for torch.relu as well # 3.2 conv + bn (+ relu) fused module configs # fused conv bn conv_configs.append({ "pattern": convs.fused_conv_bn, "dtype_configs": dtype_configs, "qat_module": convs.bn_qat, }) # fused conv bn relu conv_configs.append({ "pattern": convs.fused_conv_bn_relu, "dtype_configs": dtype_configs, "qat_module": convs.bn_relu_qat, }) # conv bn, qat fused module conv_configs.append({ "pattern": convs.bn_qat, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, }) # conv bn relu, qat fused module conv_configs.append({ "pattern": convs.bn_relu_qat, "observation_type": observation_type, "dtype_configs": dtype_configs, "root_module": convs.root, "reference_quantized_module_for_root": convs.reference, }) # (4) conv transpose and its fusion # 4.1 conv transpose config conv_configs.append({ "pattern": convs.transpose, "dtype_configs": dtype_configs, "root_module": convs.transpose, "reference_quantized_module_for_root": convs.transpose_reference, }) # 4.2 conv transpose + bn fusion conv_configs.append({ "pattern": (convs.bn, convs.transpose), "dtype_configs": dtype_configs, "fuser_method": reverse2(fuse_convtranspose_bn), "root_module": convs.transpose, "reference_quantized_module_for_root": convs.transpose_reference, }) return conv_configs def _get_binary_op_configs(dtype_configs): binary_op_configs: List[Dict[str, Any]] = [] num_tensor_args_to_observation_type_mapping = { # TODO: this is not used right now since we have extra check in prepare # will need to change this to NO_OBSERVER later after we implemented # Tensor dtype inference properly 0: ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, 1: ObservationType.OUTPUT_SHARE_OBSERVER_WITH_INPUT, 2: ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, } for op_with_quantized_bop_scalar_variant in [ operator.add, torch.add, operator.mul, torch.mul]: binary_op_configs.append({ "pattern": (torch.nn.ReLU, op_with_quantized_bop_scalar_variant), "num_tensor_args_to_observation_type": num_tensor_args_to_observation_type_mapping, "dtype_configs": dtype_configs, }) binary_op_configs.append({ "pattern": (torch.nn.functional.relu, op_with_quantized_bop_scalar_variant), "num_tensor_args_to_observation_type": num_tensor_args_to_observation_type_mapping, "dtype_configs": dtype_configs, }) binary_op_configs.append({ "pattern": (torch.relu, op_with_quantized_bop_scalar_variant), "num_tensor_args_to_observation_type": num_tensor_args_to_observation_type_mapping, "dtype_configs": dtype_configs, }) binary_op_configs.append({ "pattern": op_with_quantized_bop_scalar_variant, "num_tensor_args_to_observation_type": num_tensor_args_to_observation_type_mapping, "dtype_configs": dtype_configs, }) return binary_op_configs def _get_fixed_qparams_op_configs(): fixed_qparams_op_configs = [] for fixed_qparam_op, output_observer in [ (torch.nn.Hardsigmoid, default_fixed_qparams_range_0to1_observer), (torch.nn.functional.hardsigmoid, default_fixed_qparams_range_0to1_observer), ("hardsigmoid", default_fixed_qparams_range_0to1_observer), ("hardsigmoid_", default_fixed_qparams_range_0to1_observer), (torch.nn.Sigmoid, default_fixed_qparams_range_0to1_observer), (torch.sigmoid, default_fixed_qparams_range_0to1_observer), ("sigmoid", default_fixed_qparams_range_0to1_observer), ("sigmoid_", default_fixed_qparams_range_0to1_observer), (torch.nn.Tanh, default_fixed_qparams_range_neg1to1_observer), (torch.tanh, default_fixed_qparams_range_neg1to1_observer), ("tanh", default_fixed_qparams_range_neg1to1_observer), ("tanh_", default_fixed_qparams_range_neg1to1_observer), (torch.nn.Softmax, default_fixed_qparams_range_0to1_observer), ]: fixed_qparams_op_configs.append({ "pattern": fixed_qparam_op, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, # TODO: The following two keys are temporary, since we don't want to put observer in the configs # we expect that it's provided by user # What we want to put here is the requirement on observers, in this case dtype, # quant_min, quant_max etc., but we need to first move all configs to # backend_config_dict to do that, we'll remove these keys after we fully migrated # everything to use backend_config_dict "_overwrite_output_fake_quantizer": FixedQParamsFakeQuantize.with_args(observer=output_observer), "_overwrite_output_observer": output_observer, "dtype_configs": [ weighted_op_int8_dtype_config, default_op_fp16_dtype_config, ], }) return fixed_qparams_op_configs _CAT_CONFIG = { "pattern": torch.cat, "observation_type": ObservationType.OUTPUT_SHARE_OBSERVER_WITH_INPUT, "dtype_configs": [ default_op_quint8_dtype_config, ] } def _get_bn_configs(): """ Get configs related to batchnorm """ bn_configs = [] bn_to_fused_bn = { torch.nn.BatchNorm2d: nni.BNReLU2d, torch.nn.BatchNorm3d: nni.BNReLU3d, } for bn in bn_to_fused_bn.keys(): fused_bn = bn_to_fused_bn[bn] # bn module + relu module fusion config bn_configs.append({ "pattern": (torch.nn.ReLU, bn), "dtype_configs": [default_op_quint8_dtype_config], "fuser_method": reverse_sequential_wrapper2(fused_bn), "fused_module": fused_bn, }) # bn module + F.relu fusion config bn_configs.append({ "pattern": (torch.nn.functional.relu, bn), "dtype_configs": [default_op_quint8_dtype_config], "fuser_method": reverse_sequential_wrapper2(bn_to_fused_bn[bn]), "fused_module": fused_bn, }) bn_configs.append({ "pattern": bn, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, "dtype_configs": [default_op_quint8_dtype_config], }) # fused bn configs for fused_bn in bn_to_fused_bn.values(): bn_configs.append({ "pattern": fused_bn, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, "dtype_configs": [default_op_quint8_dtype_config], }) return bn_configs def _get_share_qparams_op_configs(dtype_configs): """ Get the operator config for the operators that works for both float and quantized input if input is quantized, the output Tensor shares the same quantization parameter with input. Example operator: avgpool2d, reshape, transpose, maxpool2d Example observed operator: observer_0 - avgpool2d - observer_0 (same observer instance as input) """ def _get_share_qprams_op_backend_config(op): return { "pattern": op, "observation_type": ObservationType.OUTPUT_SHARE_OBSERVER_WITH_INPUT, "dtype_configs": dtype_configs, } share_qparams_ops = [ torch.nn.AdaptiveAvgPool1d, torch.nn.AdaptiveAvgPool2d, torch.nn.AdaptiveAvgPool3d, torch.nn.AvgPool1d, torch.nn.AvgPool2d, torch.nn.AvgPool3d, torch.nn.Hardtanh, torch.nn.Identity, torch.nn.MaxPool1d, torch.nn.MaxPool2d, torch.nn.MaxPool3d, torch.nn.ReLU, torch.nn.ReLU6, torch.adaptive_avg_pool1d, torch.nn.functional.adaptive_avg_pool2d, torch.nn.functional.adaptive_avg_pool3d, torch.nn.functional.hardtanh, torch.nn.functional.hardtanh_, torch.nn.functional.interpolate, torch.nn.functional.max_pool1d, torch.nn.functional.max_pool2d, torch.nn.functional.max_pool3d, torch.nn.functional.relu, torch.nn.functional.relu6, torch.avg_pool1d, torch._C._nn.avg_pool2d, torch._C._nn.avg_pool3d, torch.clamp, torch.flatten, torch.mean, torch.repeat_interleave, torch.transpose, torch.squeeze, torch.stack, torch.unsqueeze, operator.floordiv, "contiguous", "clamp", "detach", "detach_", "mean", "permute", "repeat", "repeat_interleave", "reshape", "resize_", "relu", "relu_", "shape", "size", "squeeze", "squeeze_", "transpose", "unsqueeze", "unsqueeze_", "view" ] return [_get_share_qprams_op_backend_config(op) for op in share_qparams_ops] def _get_rnn_op_configs(): rnn_op_configs = [] for rnn_op, ref_rnn_op in [ (nn.GRUCell, nnqr.GRUCell), (nn.LSTMCell, nnqr.LSTMCell), (nn.RNNCell, nnqr.RNNCell), (nn.LSTM, nnqr.LSTM) ]: rnn_op_configs.append({ "pattern": rnn_op, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, "dtype_configs": [default_dynamic_int8_dtype_config, default_dynamic_float16_dtype_config], "root_module": rnn_op, "reference_quantized_module_for_root": ref_rnn_op, }) return rnn_op_configs def _get_embedding_op_configs(): embedding_op_configs = [] for embedding_op, qat_embedding_op, ref_embedding_op in [ (nn.Embedding, nnqat.Embedding, nnqr.Embedding), (nn.EmbeddingBag, nnqat.EmbeddingBag, nnqr.EmbeddingBag), ]: embedding_op_configs.append({ "pattern": embedding_op, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, "dtype_configs": [ weight_only_quint8_dtype_config, weight_only_quint4x2_dtype_config ], "qat_module": qat_embedding_op, "root_module": embedding_op, "reference_quantized_module_for_root": ref_embedding_op, # This is temporary, and will be removed soon "_input_output_observed": False }) # config for qat op embedding_op_configs.append({ "pattern": qat_embedding_op, "observation_type": ObservationType.OUTPUT_USE_DIFFERENT_OBSERVER_AS_INPUT, "dtype_configs": [ weight_only_quint8_dtype_config, weight_only_quint4x2_dtype_config ], "root_module": embedding_op, "reference_quantized_module_for_root": ref_embedding_op, # This is temporary, and will be removed soon "_input_output_observed": False }) return embedding_op_configs def get_native_backend_config_dict(): """ Get backend_config_dict for PyTorch Native backend (fbgemm/qnnpack). """ linear_dtype_configs = [ weighted_op_int8_dtype_config, default_dynamic_int8_dtype_config, default_dynamic_float16_dtype_config, # TODO: maybe remove this since fbgemm/qnnpack doesn't have kernels for it default_op_fp16_dtype_config, ] binary_op_dtype_configs = [ weighted_op_int8_dtype_config, default_op_fp16_dtype_config, ] share_qparams_op_dtype_configs = [ default_op_quint8_dtype_config, default_op_fp16_dtype_config ] return { # optional "name": "native", "configs": [ *_DEFAULT_OP_INT8_CONFIGS, *_get_linear_configs(linear_dtype_configs), *_get_conv_configs(), *_get_binary_op_configs(binary_op_dtype_configs), *_get_fixed_qparams_op_configs(), _CAT_CONFIG, *_get_bn_configs(), *_get_share_qparams_op_configs(share_qparams_op_dtype_configs), *_get_rnn_op_configs(), *_get_embedding_op_configs(), ], } __all__ = [ "get_native_backend_config_dict", ]