import torch import operator import warnings from torch.fx import ( GraphModule, ) from torch.fx.graph import ( Graph, Node, ) from torch.fx.node import Argument from ..quantize import ( propagate_qconfig_, ) from ..observer import ( ObserverBase, ) from ..qconfig import QConfigAny, is_reuse_input_qconfig from ..qconfig_dict_utils import ( get_flattened_qconfig_dict, convert_dict_to_ordered_dict, update_qconfig_for_qat, ) from .qconfig_utils import ( generate_qconfig_map, update_qconfig_for_fusion, get_standalone_module_configs, ) from .quantization_patterns import ( QuantizeHandler, ) from torch.ao.quantization.quantization_types import ( Pattern, NodePattern ) from ._equalize import ( is_equalization_observer, node_supports_equalization, ) from .graph_module import ( ObservedGraphModule, ObservedStandaloneGraphModule, ) from .pattern_utils import ( MatchResult, sorted_patterns_dict, ) from .match_utils import ( find_matches, ) from ..utils import _parent_name from .utils import ( get_custom_module_class_keys, all_node_args_have_no_tensors, assert_and_get_unique_device, get_non_observable_arg_indexes_and_types, get_new_attr_name_with_prefix, NON_QUANTIZABLE_WEIGHT_OPS, WEIGHT_INDEX_DICT, BIAS_INDEX_DICT, ) from torch.ao.quantization.quantize import ( is_activation_post_process, convert ) from ..utils import ( get_qconfig_dtypes, get_swapped_custom_module_class, activation_is_statically_quantized, activation_is_int8_quantized, ) from ..backend_config.utils import ( get_pattern_to_dtype_configs, get_pattern_to_input_type_to_index, get_module_to_qat_module, get_fusion_pattern_to_root_node_getter, ) from ..backend_config import ( get_native_backend_config_dict, ) from .backend_config_utils import ( get_pattern_to_quantize_handlers, ) from typing import Any, Callable, Dict, List, Optional, Tuple, Union, Set from collections import defaultdict # list of dtypes to not add observers to DO_NOT_OBS_DTYPE_LIST = [int, float, torch.bool, None] def is_activation_post_process_node(node: Node, modules: Dict[str, torch.nn.Module]) -> bool: return isinstance(node, torch.fx.Node) and node.op == "call_module" and \ is_activation_post_process(modules[str(node.target)]) def node_arg_is_weight(node: Node, arg: Any) -> bool: if isinstance(node, Node) and node.op == 'call_function' and \ node.target in WEIGHT_INDEX_DICT: for i, node_arg in enumerate(node.args): if arg is node_arg and i in \ WEIGHT_INDEX_DICT[node.target]: # type: ignore[index] return True for kwarg_name, kwarg_value in node.kwargs.items(): if kwarg_name == 'weight' and arg is kwarg_value: return True return False def node_arg_is_bias(node: Node, arg: Any) -> bool: if not isinstance(node, Node) or node.op != 'call_function' or \ node.target not in BIAS_INDEX_DICT: return False for i, node_arg in enumerate(node.args): if arg is node_arg and i in \ BIAS_INDEX_DICT[node.target]: # type: ignore[index] return True return node.kwargs.get('bias', None) is arg def is_input_arg_dtype_supported_by_backend( arg: Argument, node: Node, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], dtype_config: Dict[str, torch.dtype], ) -> bool: """ Check if the configured qconfig for the argument is supported by the backend or not """ if isinstance(arg, (list, tuple)): return all(map(lambda a: is_input_arg_dtype_supported_by_backend(a, node, node_name_to_target_dtype, dtype_config), arg)) if not isinstance(arg, Node): return True # TODO: support check for standalone module is_weight = node_arg_is_weight(node, arg) is_bias = node_arg_is_bias(node, arg) is_activation = not is_weight and not is_bias if is_activation: is_dynamic = dtype_config.get("is_dynamic", False) if is_dynamic: input_activation_dtype = dtype_config.get("input_dtype", None) # TODO: change this after the is_dynamic refactor is landed compute_dtype = node_name_to_target_dtype[node.name].get("input_activation_compute_dtype", None) return input_activation_dtype is None or \ compute_dtype == input_activation_dtype else: input_activation_dtype = dtype_config.get("input_dtype", None) return input_activation_dtype is None or \ node_name_to_target_dtype[node.name]["input_activation_dtype"] == input_activation_dtype elif is_weight: weight_dtype = dtype_config.get("weight_dtype", None) return weight_dtype is None or node_name_to_target_dtype[node.name]["weight_dtype"] == weight_dtype else: # bias bias_dtype = dtype_config.get("bias_dtype", None) return bias_dtype is None or node_name_to_target_dtype[node.name]["bias_dtype"] == bias_dtype def is_output_dtype_supported_by_backend( node: Node, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], dtype_config: Dict[str, torch.dtype], ) -> bool: """ Check if the configured qconfig for the output is supported by the backend or not """ output_dtype = dtype_config.get("output_dtype", None) return output_dtype is None or \ output_dtype == node_name_to_target_dtype[node.name]["output_activation_dtype"] def is_observer_in_same_graph(node, modules, node_name_to_target_dtype): """ Check if observer in same graph when the node output is not fp32 and input is 'placeholder' the input is assumed to be quantized, so it is observed in a different place rather than not observed. """ node_output_dtype = get_arg_target_dtype_as_output(node, modules, node_name_to_target_dtype) if len(node.args) > 0 and isinstance(node.args[0], Node): if node_output_dtype == torch.quint8 and node.args[0].op == 'placeholder': return False return True def is_pattern_dtype_config_supported_by_backend( pattern: Optional[Pattern], matched_node_pattern: Optional[NodePattern], node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], backend_config_dict: Optional[Dict[str, Any]] ) -> bool: """ Check is the dtype configuration of a pattern is supported by the backend or not """ if backend_config_dict is None or pattern is None: return True assert matched_node_pattern is not None and len(matched_node_pattern) >= 1 pattern_to_dtype_configs = get_pattern_to_dtype_configs(backend_config_dict) dtype_configs: List[Dict[str, Any]] = pattern_to_dtype_configs.get(pattern, []) # TODO: this only works for one input and one output patterns, need to generalize to multiple # inputs/output root_node = _default_root_node_getter(matched_node_pattern) input_node = root_node output_node = matched_node_pattern[0] for dtype_config in dtype_configs: # check if arg dtype are supported supported = True for arg in input_node.args: supported = supported and \ is_input_arg_dtype_supported_by_backend( arg, input_node, node_name_to_target_dtype, dtype_config) for k, arg in input_node.kwargs.items(): supported = supported and \ is_input_arg_dtype_supported_by_backend( arg, input_node, node_name_to_target_dtype, dtype_config) # check if output dtype is supported supported = supported and is_output_dtype_supported_by_backend( output_node, node_name_to_target_dtype, dtype_config) if supported: return True return False def prepare_get_standalone_module_configs( node: Node, modules: Dict[str, torch.nn.Module], prepare_custom_config_dict: Dict[str, Any], parent_qconfig: QConfigAny, parent_backend_config_dict: Optional[Dict[str, Any]], ) -> Tuple[Dict[str, Any], Dict[str, Any], Dict[str, Any]]: """ Returns the standalone module qconfig_dict and prepare_config_dict for `node`, assuming that the module pointed to by `node` is a standalone modules. """ standalone_module_name = str(node.target) standalone_module_type = type(modules[standalone_module_name]) # type: ignore[index] sm_qconfig_dict, sm_prepare_config_dict, sm_backend_config_dict = \ get_standalone_module_configs(standalone_module_name, standalone_module_type, prepare_custom_config_dict) # fallback to use parent module's qconfig if user didn't specify qconfig dict if sm_qconfig_dict is None: sm_qconfig_dict = {"": parent_qconfig} if sm_prepare_config_dict is None: sm_prepare_config_dict = {} # TODO: sm_backend_config_dict can fallback to use parent's backend_config_dict # as well, this can be added later if sm_backend_config_dict is None: sm_backend_config_dict = parent_backend_config_dict return sm_qconfig_dict, sm_prepare_config_dict, sm_backend_config_dict def qat_swap_modules( root: torch.nn.Module, module_to_qat_module: Dict[Callable, Callable]) -> None: convert(root, mapping=module_to_qat_module, inplace=True, remove_qconfig=False) def add_matched_node_name_to_set(matched_node_pattern: NodePattern, s: Set[str]): if isinstance(matched_node_pattern, Node): s.add(matched_node_pattern.name) elif isinstance(matched_node_pattern, (list, tuple)): for maybe_node in matched_node_pattern: add_matched_node_name_to_set(maybe_node, s) # this is temporary, will be removed soon def _default_root_node_getter(node_pattern): while not isinstance(node_pattern, Node): node_pattern = node_pattern[-1] return node_pattern def insert_observer( node: Node, observer: ObserverBase, model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, ) -> Node: """ Attaches `observer` to `model`, and creates a node which calls `observer` on the output of `node`. """ model_device = assert_and_get_unique_device(model) if model_device: observer.to(model_device) # add observer module as attribute if is_equalization_observer(observer): prefix = node.name + '_equalization_process_' else: prefix = 'activation_post_process_' get_new_observer_name = get_new_attr_name_with_prefix(prefix) observer_name = get_new_observer_name(model) setattr(model, observer_name, observer) modules[observer_name] = observer with graph.inserting_after(node): new_obs = graph.create_node( 'call_module', observer_name, (node,), {}) return new_obs def get_target_activation_dtype_for_node( node: Node, qconfig: QConfigAny, inputs_seen_counter: int, outputs_seen_counter: int, input_quantized_idxs: List[int], output_quantized_idxs: List[int], qhandler: Optional[QuantizeHandler], modules: Dict[str, torch.nn.Module], cache_for_no_tensor_check: Dict[Node, bool], ) -> Dict[str, Optional[Union[torch.dtype, type]]]: """ Returns the expected dtype of the input and output of this node after convert. If the value is not None, it represents the dtype of the Tensor. If the value is None, it means the value is not a Tensor. Note: this is for activations only, weight dtypes are not handled here. TODO(future PR, if needed): explicitly spell out the non-Tensor dtypes. """ if node.op == 'placeholder': if inputs_seen_counter in input_quantized_idxs: return { "input_activation_dtype": torch.quint8, "output_activation_dtype": torch.quint8, } else: # if dtype is fp32 (default), do nothing # note: other dtypes are not supported return { "input_activation_dtype": torch.float, "output_activation_dtype": torch.float, } elif node.op in ('call_module', 'call_method', 'call_function'): args_have_no_tensors = \ all_node_args_have_no_tensors( node, modules, cache_for_no_tensor_check) if args_have_no_tensors: return { "input_activation_dtype": None, "output_activation_dtype": None, } # TODO(future PR): consider stopping matching getitem is_getitem = node.op == 'call_function' and \ node.target == operator.getitem if is_getitem: return { "input_activation_dtype": torch.float, "output_activation_dtype": torch.float, } # get qconfig to determine the eventual dtype of this node if qconfig is not None: if qhandler is not None and qhandler.input_output_observed(): act_dtype, weight_dtype, act_compute_dtype = \ get_qconfig_dtypes(qconfig) bias_dtype = torch.float16 \ if act_dtype == torch.float16 and weight_dtype == torch.float16 \ else torch.float return { "input_activation_dtype": act_dtype, "input_activation_compute_dtype": act_compute_dtype, "weight_dtype": weight_dtype, "bias_dtype": bias_dtype, "output_activation_dtype": act_dtype, } return { "input_activation_dtype": torch.float, "output_activation_dtype": torch.float, } elif node.op == 'get_attr': return { "input_activation_dtype": torch.float, "output_activation_dtype": torch.float, } elif node.op == 'output': if outputs_seen_counter in output_quantized_idxs: return { "input_activation_dtype": torch.quint8, "output_activation_dtype": torch.quint8 } else: # if dtype is fp32 (default), do nothing # note: other dtypes are not supported return { "input_activation_dtype": torch.float, "output_activation_dtype": torch.float, } else: raise AssertionError(f'need to handle {node.format_node()}') def get_arg_target_dtype_as_output( arg: Node, modules: Dict[str, torch.nn.Module], node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], ) -> Optional[Union[torch.dtype, type]]: """ Get the target output activation dtype for the argumnet in the original graph, skipping inserted observers We are assuming that the observers are inserted correctly, and the dtype for argument in quantized graph will match what is specified by the qconfig """ assert isinstance(arg, Node) if is_activation_post_process_node(arg, modules): observed_arg = arg.args[0] assert isinstance(observed_arg, Node), "Currently we only support observing Node" return node_name_to_target_dtype[observed_arg.name]["output_activation_dtype"] else: return node_name_to_target_dtype[arg.name]["output_activation_dtype"] def get_arg_target_dtype_as_input_to_node( arg: Node, node: Node, modules: Dict[str, torch.nn.Module], node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], ) -> Optional[Union[torch.dtype, type]]: """ Get the target argument dtype for the argument `arg`, as input to node `node` """ assert isinstance(arg, Node) is_weight = node_arg_is_weight(node, arg) is_bias = node_arg_is_bias(node, arg) is_activation = not is_weight and not is_bias if is_activation: return node_name_to_target_dtype[node.name]["input_activation_dtype"] elif is_weight: if node.target in NON_QUANTIZABLE_WEIGHT_OPS: return None else: return node_name_to_target_dtype[node.name]["weight_dtype"] else: return node_name_to_target_dtype[node.name]["bias_dtype"] def get_arg_target_compute_dtype_as_input_to_node( arg: Node, node: Node, modules: Dict[str, torch.nn.Module], node_name_to_target_dtype: Dict[str, Dict[str, Union[torch.dtype, type, None]]], ) -> Union[torch.dtype, type, None]: """ Get the target argument dtype for the argument `arg`, as input to node `node` """ assert isinstance(arg, Node) is_weight = node_arg_is_weight(node, arg) is_bias = node_arg_is_bias(node, arg) is_activation = not is_weight and not is_bias if is_activation and \ "input_activation_compute_dtype" in node_name_to_target_dtype[node.name]: return node_name_to_target_dtype[node.name]["input_activation_compute_dtype"] else: return None def maybe_insert_input_observer_for_arg_or_kwarg( node: Union[Node, Any], arg: Argument, qconfig: QConfigAny, model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], qhandler: Optional[QuantizeHandler], prepare_custom_config_dict: Dict[str, Any], backend_config_dict: Optional[Dict[str, Any]], ) -> Argument: """ Given a `node` and an `arg`, inserts an input observer between `node` and `arg` if necessary. """ # for ops such as torch.cat([x0, x1]), # traverse through the list if isinstance(arg, (list, tuple)): new_arg_to_return = [] for inner_arg in arg: new_inner_arg = maybe_insert_input_observer_for_arg_or_kwarg( node, inner_arg, qconfig, model, modules, graph, node_name_to_target_dtype, qhandler, prepare_custom_config_dict, backend_config_dict) new_arg_to_return.append(new_inner_arg) return type(arg)(new_arg_to_return) if not isinstance(arg, Node): return arg assert isinstance(arg, Node) # default (no observer) new_arg = arg is_standalone_module = qhandler is not None and qhandler.is_standalone_module() assert qconfig is not None if not is_standalone_module: # regular flow for most nodes, except standalone modules is_weight = node_arg_is_weight(node, arg) is_reuse_input_qconfig_ = is_reuse_input_qconfig(qconfig) act_post_process_ctr = qconfig.weight if is_weight else \ qconfig.activation arg_as_output_target_dtype = get_arg_target_dtype_as_output(arg, modules, node_name_to_target_dtype) arg_as_input_target_dtype = get_arg_target_dtype_as_input_to_node(arg, node, modules, node_name_to_target_dtype) arg_as_input_target_compute_dtype = \ get_arg_target_compute_dtype_as_input_to_node( arg, node, modules, node_name_to_target_dtype) needs_obs = ( # if the dtypes are different, we need an observer (arg_as_output_target_dtype != arg_as_input_target_dtype) and # except if the second dtype is float, a dequant will be inserted # without an observer in convert # TODO(future PR): change this so a placeholder is inserted for # future dequants, to make the logic easier to understand (arg_as_input_target_dtype != torch.float) and # if arg output dtype is in DO_NOT_OBS_DTYPE_LIST do not insert observer (arg_as_output_target_dtype not in DO_NOT_OBS_DTYPE_LIST) and # if qconfig is reuse_input qconfig, we won't insert extra observer for input not is_reuse_input_qconfig_ or # need to add input observer for dynamic quantization # only add observer for first input for now, we may need to extend # qconfig_dict and backend_config_dict to support more general configurations # of dynamic quantization, e.g. dynamically quantizing second input, third # input etc. (arg_as_input_target_compute_dtype in [torch.quint8, torch.int8, torch.float16]) and arg is node.args[0] ) else: # custom flow for standalone modules _sm_qconfig_dict, sm_prepare_config_dict, _sm_backend_config_dict = \ prepare_get_standalone_module_configs( node, modules, prepare_custom_config_dict, qconfig, backend_config_dict) sm_input_quantized_idxs = \ sm_prepare_config_dict.get('input_quantized_idxs', []) # for args, this is set to the index of the current arg # for kwargs, this is left at None cur_input_idx = None for arg_idx, arg_to_check in enumerate(node.args): if arg_to_check is arg: cur_input_idx = arg_idx break if cur_input_idx is None: needs_obs = False else: arg_as_output_target_dtype = get_arg_target_dtype_as_output(arg, modules, node_name_to_target_dtype) arg_as_input_target_dtype = torch.quint8 if cur_input_idx in sm_input_quantized_idxs \ else torch.float needs_obs = ( (arg_as_output_target_dtype != arg_as_input_target_dtype) and (arg_as_input_target_dtype != torch.float) ) act_post_process_ctr = qconfig.activation if needs_obs: new_obs_mod = act_post_process_ctr() existing_obs_node = None # Before using the new observer, check if an observer # of the correct type already exists. If it does, use it. # This prevents duplicate observer insertions if a node is # used by multiple nodes. # TODO: this is looking into how the value is used in the future # we should remove this # removing this means we insert one observer for each use, even if they # have the same dtype, we can have an extra pass that removes the extra observers for maybe_obs_node, _ in arg.users.items(): if maybe_obs_node.op == 'call_module': maybe_obs_mod = modules[maybe_obs_node.target] # type: ignore[index] if ( type(maybe_obs_mod) == type(new_obs_mod) and maybe_obs_mod.dtype == arg_as_input_target_dtype ): existing_obs_node = maybe_obs_node break if existing_obs_node is None: new_obs_node = insert_observer( arg, new_obs_mod, model, modules, graph) # override this arg to be the observed arg new_arg = new_obs_node else: new_arg = existing_obs_node return new_arg def maybe_insert_input_observers_for_node( node: Node, qconfig: QConfigAny, model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], qhandler: Optional[QuantizeHandler], prepare_custom_config_dict: Dict[str, Any], backend_config_dict: Optional[Dict[str, Any]], ) -> None: """ If needed, inserts observers to the input args and kwargs of `node`. Note: modifies `node` inplace. For example, if cur_node needs an observer after prev_node, we change from prev_node -> cur_node To prev_node -> obs -> cur_node """ if qconfig is None: # if quantization is turned off for this node, we do not need # to insert input observers return assert qconfig is not None # Look through every input arg. If that arg's target dtype does not # match the current node's target dtype, insert an observer. new_args = [] for arg in node.args: new_arg = maybe_insert_input_observer_for_arg_or_kwarg( node, arg, qconfig, model, modules, graph, node_name_to_target_dtype, qhandler, prepare_custom_config_dict, backend_config_dict) new_args.append(new_arg) new_kwargs = {} for k, kwarg in node.kwargs.items(): new_kwarg = maybe_insert_input_observer_for_arg_or_kwarg( node, kwarg, qconfig, model, modules, graph, node_name_to_target_dtype, qhandler, prepare_custom_config_dict, backend_config_dict) new_kwargs[k] = new_kwarg # assign the new args and kwargs to the node, inplace node.args = tuple(new_args) node.kwargs = new_kwargs def maybe_insert_input_equalization_observers_for_node( node: Node, equalization_qconfig: Any, model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], is_branch: bool, ) -> None: """ If `node` needs to be equalized, find the input/weight observers it needs in `equalization_qconfig`, creates them, and inserts it into `graph`. If `node` does not need an equalization observer, returns None. """ if equalization_qconfig is None or not node_supports_equalization(node, modules): return if is_branch: warnings.warn( f"Cannot equalize {node} because it is part of a branch." ) return new_args = [] for arg in node.args: if not isinstance(arg, Node) or node_arg_is_bias(node, arg): new_args.append(arg) continue is_weight = node_arg_is_weight(node, arg) act_eq_process_ctr = equalization_qconfig.weight if is_weight else \ equalization_qconfig.input_activation new_eq_obs_mod = act_eq_process_ctr() new_eq_obs_node = insert_observer( arg, new_eq_obs_mod, model, modules, graph) new_args.append(new_eq_obs_node) # assign the new args and kwargs to the node, inplace node.args = tuple(new_args) def maybe_insert_output_observer_for_node( node: Node, model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, matches: Dict[str, MatchResult], node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], matched_pattern: Any, qhandler: Optional[QuantizeHandler], is_qat: bool, ) -> Optional[Node]: """ If `node` needs an output observer, creates it, inserts it into `graph` and returns it. If `node` does not need an output observer, returns None. """ root_node, _, pattern, qhandler, qconfig = matches.get( node.name, (None, None, None, None, None)) if qhandler is None: return None assert qconfig is not None assert node.op != 'output', 'observer insertion for outputs is handled elsewhere' is_standalone_module = qhandler is not None and qhandler.is_standalone_module() dtype = node_name_to_target_dtype[node.name]["output_activation_dtype"] should_insert_observer = dtype not in DO_NOT_OBS_DTYPE_LIST + [torch.float] # TODO(future PR): move the following logic to # should_insert_observer_for_output should_insert_observer = should_insert_observer and \ activation_is_statically_quantized(qconfig) # we never insert observers to output of standalone module, we assume # if needed, they are inserted inside the standalone module should_insert_observer = should_insert_observer and \ (not is_standalone_module) if should_insert_observer: act_post_process_ctr = qconfig.activation if activation_is_int8_quantized(qconfig): act_post_process_ctr = qhandler.get_activation_ctr( qconfig, matched_pattern, is_qat) observer = act_post_process_ctr() new_obs = insert_observer(node, observer, model, modules, graph) return new_obs else: return None def maybe_insert_observers_before_graph_output( graph_output_node: Node, output_quantized_idxs: List[int], node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], qconfig_map: Dict[str, QConfigAny], model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, ) -> None: """ If the output needs to be quantized and there are any nodes in the output which are not already observed, inserts observers for those nodes. """ # TODO(future PR): update the output_quantized_idxs API to match # arbitrary data structures. There is always a single output, and # that output can have arbitrary nesting of values. List[int] is # not the right data type for this. assert output_quantized_idxs == [0] or output_quantized_idxs == [], \ 'unrecognized format of output_quantized_idxs' # Currently dequants are inserted in the convert step. So, we only # have to do anything if the output is hardcoded to be quantized if output_quantized_idxs == []: return # TODO(future PR): support more dtypes in model outputs, if necessary output_target_dtype = torch.quint8 def _recursive_maybe_replace_node_with_obs( maybe_node: Argument, target_dtype: torch.dtype, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], qconfig_map: Dict[str, QConfigAny], model: torch.nn.Module, modules: Dict[str, torch.nn.Module], graph: Graph, ) -> Argument: """ Navigate an arbitrary data structure of lists, tuples, dicts. For each container type, recurse on all inputs. Once any Node is found, insert an observer if needed and do not recurse further. For example, given a structure of {'foo1': [[bar1]], 'foo2': {'foo3': [[[bar3]]]}} we recurse down to bar1 and bar3, observe them if necessary, and if we inserted an observer then replace the original node with its observer. Returns the data structure with all nodes needing observation being replaced by their observers. """ if isinstance(maybe_node, Node): # check dtype of this node this_node_dtype = get_arg_target_dtype_as_output( maybe_node, modules, node_name_to_target_dtype) if this_node_dtype != target_dtype: # insert observer qconfig = qconfig_map.get(maybe_node.name) # TODO(future PR): see if we need to allow specifying qconfig # on output nodes, to remove the restriction below. assert qconfig is not None, \ 'Quantizing the output node without a qconfig is not supported' observer_mod = qconfig.activation() observer_node = insert_observer( maybe_node, observer_mod, model, modules, graph) return observer_node else: return maybe_node elif isinstance(maybe_node, (list, tuple)): results = [] for inner_node in maybe_node: results.append(_recursive_maybe_replace_node_with_obs( inner_node, target_dtype, node_name_to_target_dtype, qconfig_map, model, modules, graph)) if isinstance(maybe_node, list): return results else: return tuple(results) elif isinstance(maybe_node, dict): results_dict = {} for k, inner_v in maybe_node.items(): results_dict[k] = _recursive_maybe_replace_node_with_obs( inner_v, target_dtype, node_name_to_target_dtype, qconfig_map, model, modules, graph) return results_dict else: return results new_args = [] for old_arg in graph_output_node.args: new_args.append( _recursive_maybe_replace_node_with_obs( old_arg, output_target_dtype, node_name_to_target_dtype, qconfig_map, model, modules, graph)) graph_output_node.args = tuple(new_args) # type: ignore[assignment] def maybe_propagate_dtype_for_node( node: Node, target_dtype: Union[torch.dtype, type], node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], matches: Dict[str, MatchResult], ) -> None: """ Assigns `target_dtype` to `node`. If `node` is a general tensor shape op (see GeneralTensorShapeOpQuantizeHandler in quantization_patterns.py for more details) also call this function recursively on the first argument, to propagate the dtype to the caller. """ node_name_to_target_dtype[node.name]["input_activation_dtype"] = target_dtype node_name_to_target_dtype[node.name]["output_activation_dtype"] = target_dtype # if this is a copy node, propagate to first arg root_node, _, pattern, qhandler, qconfig = matches.get( node.name, (None, None, None, None, None)) if qhandler is not None and qhandler.is_general_tensor_value_op(): prev_node = node.args[0] if isinstance(prev_node, Node): maybe_propagate_dtype_for_node( prev_node, target_dtype, node_name_to_target_dtype, matches) def propagate_dtypes_for_known_nodes( graph: Graph, node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]], matches: Dict[str, MatchResult], ) -> None: """ Currently we assume that inputs to the graph are either `torch.float` or `torch.quint8`, which is not always correct. For ops such as `x.masked_fill(mask, value)`, we know that the dtype of `mask` is a `BoolTensor`. Propagate this information throughout the graph. Note: not all dtypes in the graph will be correct after this pass, but a higher percentage of them will be correct. Hopefully in the future we can replace this with a better way to reason about dtypes of tensors. """ for node in graph.nodes: non_observable_arg_dict = get_non_observable_arg_indexes_and_types(node) for arg_type in non_observable_arg_dict: non_observable_indices = non_observable_arg_dict[arg_type](node) for index in non_observable_indices: arg = node.args[index] # when an argument is a tuple, it does not show up as another node so we need to go through # all elements of the tuple manually if isinstance(arg, tuple) or isinstance(arg, list): arg_list = list(arg) else: arg_list = [arg] for cur_arg in arg_list: # hard coded arguments show up but aren't `Node` typed and do not need dtype propgated if isinstance(cur_arg, torch.fx.node.Node): maybe_propagate_dtype_for_node( cur_arg, arg_type, node_name_to_target_dtype, matches) def maybe_make_input_output_share_observers( node: Node, model: torch.nn.Module, modules: Dict[str, torch.nn.Module], ) -> bool: """ Ensures that we share an observer for all input arguments as well as the output argument. In detail, given a graph of x0 -> obs0 -> op -> x2 / x1 -> obs1 / where node obs0 points to observer instance observer0, obs1 points to observer1 and obs2 points to observer2, we make nodes obs1 and ob2 point to observer0. Returns: whether the operation succeeded or not """ first_arg = None # find the first non-Tensor arg for i in range(len(node.args)): if isinstance(node.args[i], (Node, list, tuple)): first_arg = node.args[i] break # if there is no non-Tensor arg, return directly if first_arg is None: return False if isinstance(first_arg, (list, tuple)): first_arg_arg = first_arg[0] elif isinstance(first_arg, Node): first_arg_arg = first_arg else: return False # if we have a graph such as # observed_node -> non_observed_node -> cat # we need to navigate up to the first observer iteration_guard = 0 while not is_activation_post_process_node(first_arg_arg, modules): if not isinstance(first_arg_arg, Node): return False # did not find an activation_post_process for the op if first_arg_arg.op == "placeholder": return False # trace back the args until we found the first Tensor/Node trace_back_node = None for i in range(len(first_arg_arg.args)): trace_back_node = first_arg_arg.args[i] if isinstance(trace_back_node, Node): break if trace_back_node is None: return False first_arg_arg = trace_back_node iteration_guard += 1 if iteration_guard > 10000: raise AssertionError('Unable to find observer of previous node') assert isinstance(first_arg_arg, Node) target_to_use = first_arg_arg.target assert isinstance(target_to_use, str) obs_mod_to_use = modules[target_to_use] if isinstance(first_arg, (list, tuple)): # set all other input observer nodes to use that module for input_idx, input_arg in enumerate(first_arg): if input_idx == 0: continue iteration_guard = 0 while not is_activation_post_process_node(input_arg, modules): # failed to trace back since no input arg for the current node if len(input_arg.args) < 1: return False input_arg = input_arg.args[0] iteration_guard += 1 if iteration_guard > 10000: raise AssertionError('Unable to find observer of previous node') parent_name, name = _parent_name(input_arg.target) setattr(modules[parent_name], name, obs_mod_to_use) # set the output observer node to use that module for output_obs_node, _ in node.users.items(): assert is_activation_post_process_node(output_obs_node, modules) parent_name, name = _parent_name(output_obs_node.target) setattr(modules[parent_name], name, obs_mod_to_use) # TODO(future PR): delete the orphaned observer modules return True def remove_output_observer( node: Node, model: torch.nn.Module, modules: Dict[str, torch.nn.Module]): items = list(node.users.items()) for output_obs_node, _ in items: assert is_activation_post_process_node(output_obs_node, modules) output_obs_node.replace_all_uses_with(node) model.graph.erase_node(output_obs_node) # type: ignore[union-attr, operator] def swap_custom_module_to_observed( node: Node, qconfig: QConfigAny, modules: Dict[str, torch.nn.Module], prepare_custom_config_dict: Dict[str, Any]): custom_module = modules[node.target] # type: ignore[index] custom_module_class_mapping = prepare_custom_config_dict.get( "float_to_observed_custom_module_class", {}) observed_custom_module_class = \ get_swapped_custom_module_class( custom_module, custom_module_class_mapping, qconfig) observed_custom_module = \ observed_custom_module_class.from_float(custom_module) parent_name, name = _parent_name(node.target) setattr(modules[parent_name], name, observed_custom_module) def insert_observers_for_model( model: GraphModule, modules: Dict[str, torch.nn.Module], matches: Dict[str, MatchResult], qconfig_map: Dict[str, QConfigAny], graph: Graph, prepare_custom_config_dict: Dict[str, Any], equalization_config_map: Dict[str, Any], input_quantized_idxs: List[int], output_quantized_idxs: List[int], backend_config_dict: Optional[Dict[str, Any]], observed_node_names: Set[str], is_qat: bool, ) -> Optional[Node]: """ Inserts observers, using the following high level algorithm: For each node in the graph: 1. determine the target dtype of this node in the quantized graph, and save it for future steps 2. determine the target dtype or all args and kwargs of this node 3. if any arg or kwarg's target dtype does not match the current node's dtype, insert an observer 4. if the current node needs an output observer, insert it For example: - starting graph: x0 -> linear -> x1 - observed graph after processing x0: x0(fp32) - observed graph after processing linear: x0(fp32) -> x0_obs0(int8) -> linear(int8) -> linear_obs0(int8) - observed graph after processing x1: x0(fp32) -> x0_obs0(int8) -> linear(int8) -> linear_obs0(int8) -> x1 After a node is processed, the naive observer placement is guaranteed to be complete for that node and all of its predecessors. There can be future passes which optimize the graph by deduplicating observers, etc. """ # name of Node in original FX Graph to the target dtype information # that's derived from qconfig for the Node, for example, if we have # a conv2d node that has a qconfig # { # # information for input and bias node omitted # # for getattr node # # weight = getattr(self, 'weight') # 'weight': { # 'output_activation_dtype': torch.float, # } # # for conv2d node # # conv2d = call_function[target=torch.nn.functional.conv2d]( # # args=(input, weight, bias)) # 'conv2d': { # 'input_activation_dtype': torch.quint8, # 'weight_dtype': torch.qint8, # 'bias_dtype': torch.float, # 'output_activation_dtype': torch.quint8, # } # } # # TODO: rename this to node_name_to_target_dtype_info node_name_to_target_dtype: Dict[str, Dict[str, Optional[Union[torch.dtype, type]]]] = defaultdict(dict) cache_for_no_tensor_check: Dict[Node, bool] = dict() inputs_seen_counter = 0 outputs_seen_counter = 0 # first, populate the dtype map based only on qconfig and qhandler # this assumes: # graph inputs are fp32 by default, and int8 where overriden # other nodes output dtype is specified by the qconfig modules = dict(model.named_modules(remove_duplicate=False)) for node in model.graph.nodes: root_node, _, pattern, qhandler, qconfig = matches.get( node.name, (None, None, None, None, None)) node_name_to_target_dtype[node.name] = get_target_activation_dtype_for_node( node, qconfig, inputs_seen_counter, outputs_seen_counter, input_quantized_idxs, output_quantized_idxs, qhandler, modules, cache_for_no_tensor_check) if node.op == "placeholder": inputs_seen_counter += 1 if node.op == "output": outputs_seen_counter += 1 # Second, for nodes with known input dtypes, propagate them throughout the # graph. For example, if there is a call such as # x1 = x0.masked_fill(mask, 1) # we propagate the type of mask to be torch.bool propagate_dtypes_for_known_nodes( model.graph, node_name_to_target_dtype, matches) # After this point, the current node and all of its arguments # have a dtype assigned. Now, we insert observers for inputs # of this node (if needed for this node), and the output of this node # (if needed for this node). # Since we are mutating the graph as we go, we iterate over the original # nodes before observer insertion, instead of model.graph.nodes. nodes_before_observation = list(model.graph.nodes) # reset inputs/outputs counters inputs_seen_counter = 0 outputs_seen_counter = 0 results_node = None for node in nodes_before_observation: if node.op == 'placeholder': # if a graph input is in fp32, it does not need observation # if a graph input is in int8, we assume the observation happens # outside of the graph, and no additional observation is needed pass elif node.op in ('call_module', 'call_method', 'call_function', 'output'): # check for matches last_node, matched_node_pattern, pattern, qhandler, qconfig = matches.get( node.name, (None, None, None, None, None)) equalization_qconfig = equalization_config_map.get(node.name, None) this_node_dtype = node_name_to_target_dtype[node.name] output_not_a_tensor = this_node_dtype is None # TODO(future PR): consider stopping matching getitem is_getitem = node.op == 'call_function' and \ node.target == operator.getitem skip_inserting_observers = ( (qconfig is None) or output_not_a_tensor or is_getitem ) and ( not node.op == 'output' ) is_supported_by_backend = is_pattern_dtype_config_supported_by_backend( pattern, matched_node_pattern, node_name_to_target_dtype, backend_config_dict) if not skip_inserting_observers and is_supported_by_backend: modules = dict(model.named_modules(remove_duplicate=False)) if node.op != 'output': assert matched_node_pattern is not None # add matched nodes to the observed node name set add_matched_node_name_to_set(matched_node_pattern, observed_node_names) # This is currently only used for equalization. # Checks if the current node is in a branch in which the two # first layers are both being quantized. # # ex. conv2 # / # x -> conv1 # # If this is the case, we will not apply equalization to the # initial two layers. is_quantized_branch = False if ( len(node.args) > 0 and isinstance(node.args[0], Node) and len(node.args[0].users) > 1 ): for user in node.args[0].users: # Checks if there exists another user being quantized is_user_quantized = ( qconfig_map.get(user.name, None) is not None or (user.op == 'call_module' and isinstance(modules[str(user.target)], ObserverBase)) ) if user != node and is_user_quantized: is_quantized_branch = True # TODO: this only works for sequential fusion right now, extend it # it to automatically detect all input nodes based on the pattern # need to change find_matches function to return this information root_node = _default_root_node_getter(matched_node_pattern) is_input_node_of_the_pattern = node is root_node if is_input_node_of_the_pattern: # this modifies node inplace maybe_insert_input_observers_for_node( node, qconfig, model, modules, graph, node_name_to_target_dtype, qhandler, prepare_custom_config_dict, backend_config_dict) # Insert equalization input observers if needed maybe_insert_input_equalization_observers_for_node( node, equalization_qconfig, model, modules, graph, node_name_to_target_dtype, is_quantized_branch) is_last_node_of_pattern = node is last_node is_general_tensor_value_op = \ (qhandler is not None and qhandler.is_general_tensor_value_op()) is_reuse_input_qconfig_ = is_reuse_input_qconfig(qconfig) if is_last_node_of_pattern: # this returns the new observer node if it was needed maybe_output_obs_node = maybe_insert_output_observer_for_node( node, model, modules, graph, matches, node_name_to_target_dtype, pattern, qhandler, is_qat) if maybe_output_obs_node is not None: # Update users of original node to use the output observer # instead. For example, change # # next_node # / # cur_node -> obs # # to # # next_node # / # cur_node -> obs # # We need to save orig users before updating uses because # the list of users will change as we update uses orig_users = list(node.users.keys()) for user_node in orig_users: if user_node is maybe_output_obs_node: continue user_node.replace_input_with(node, maybe_output_obs_node) is_observer_in_same_graph_ = is_observer_in_same_graph(node, modules, node_name_to_target_dtype) # for general tensor value ops, we modify the graph # to make all inputs and outputs use the first input's # observer if (is_general_tensor_value_op and is_observer_in_same_graph_) or \ is_reuse_input_qconfig_: if not maybe_make_input_output_share_observers(node, model, modules): remove_output_observer(node, model, modules) if qhandler is not None and qhandler.is_custom_module(): swap_custom_module_to_observed(node, qconfig, modules, prepare_custom_config_dict) else: # output maybe_insert_observers_before_graph_output( node, output_quantized_idxs, node_name_to_target_dtype, qconfig_map, model, modules, graph) # # After this point, the current node has input and output observers # that it needs for itself inserted. # # increment the counters, so future inputs and outputs are assigned # correct dtypes if node.op == 'placeholder': inputs_seen_counter += 1 elif node.op == 'output': outputs_seen_counter += 1 results_node = node return results_node def run_prepare_fx_on_standalone_modules( model: torch.nn.Module, is_qat: bool, modules: Dict[str, torch.nn.Module], matches: Any, prepare_custom_config_dict: Dict[str, Any], backend_config_dict: Optional[Dict[str, Any]], ) -> None: """ Runs prepare_fx on each standalone module. Note: this does not modify the graph, it just replaces the unobserved modules with their observed versions. """ for ( node_name, (root_node, _, pattern, qhandler, qconfig), ) in matches.items(): if qhandler is None: continue elif not qhandler.is_standalone_module(): continue sm_qconfig_dict, sm_prepare_config_dict, sm_backend_config_dict = \ prepare_get_standalone_module_configs( root_node, modules, prepare_custom_config_dict, qconfig, backend_config_dict) standalone_module = modules[root_node.target] prepare = \ torch.ao.quantization.quantize_fx._prepare_standalone_module_fx # type: ignore[attr-defined] observed_standalone_module = \ prepare( standalone_module, sm_qconfig_dict, is_qat, sm_prepare_config_dict, backend_config_dict=sm_backend_config_dict) preserved_attributes = \ set(sm_prepare_config_dict.get("preserved_attributes", [])) observed_standalone_module = ObservedStandaloneGraphModule( observed_standalone_module, observed_standalone_module.graph, preserved_attributes) parent_name, name = _parent_name(root_node.target) setattr(modules[parent_name], name, observed_standalone_module) modules[root_node.target] = observed_standalone_module def save_state( observed: GraphModule, qconfig_map: Dict[str, QConfigAny], node_name_to_scope: Dict[str, Tuple[str, type]], prepare_custom_config_dict: Dict[str, Any], equalization_qconfig_map: Dict[str, Any], qconfig_dict: Dict[str, Dict[Any, Any]], is_qat: bool, observed_node_names: Set[str], ) -> None: observed._qconfig_map = qconfig_map # type: ignore[assignment] observed._prepare_custom_config_dict = \ prepare_custom_config_dict # type: ignore[assignment] observed._node_name_to_scope = node_name_to_scope # type: ignore[assignment] observed._equalization_qconfig_map = equalization_qconfig_map # type: ignore[assignment] observed._qconfig_dict = qconfig_dict # type: ignore[assignment] observed._is_qat = is_qat # type: ignore[assignment] observed._observed_node_names = observed_node_names # type: ignore[assignment] def prepare( model: GraphModule, qconfig_dict: Any, is_qat: bool, node_name_to_scope: Dict[str, Tuple[str, type]], prepare_custom_config_dict: Optional[Dict[str, Any]] = None, equalization_qconfig_dict: Optional[Dict[str, Any]] = None, backend_config_dict: Optional[Dict[str, Any]] = None, is_standalone_module: bool = False) -> ObservedGraphModule: """ 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 Args: node_name_to_scope: mapping from node name to the scope of the module which contains the node. The scope is a tuple of fully qualified path of the module and the type of the module Returns: model(GraphModule): prepared standalone module attributes: _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 indexs for the graph output that is quantized same as input_quantized_idxs configuration provided for the standalone module """ if prepare_custom_config_dict is None: prepare_custom_config_dict = {} if equalization_qconfig_dict is None: equalization_qconfig_dict = {} # mapping from a tuple of nodes in reverse order to uninitialized # QuantizeHandler subclass. For example, # { # # match a single node # (: # ), # # match multiple nodes in reverse order # ((, ): # ), # } # TODO: rename to pattern_to_quantize_handler patterns: Dict[Pattern, QuantizeHandler] = {} if backend_config_dict is None: backend_config_dict = get_native_backend_config_dict() patterns = get_pattern_to_quantize_handlers(backend_config_dict) patterns = sorted_patterns_dict(patterns) # TODO: make WEIGHT_INDEX_DICT and BIAS_INDEX_DICT an argument to the functions that needs them # TODO: refactor this part to return WEIGHT_INDEX_DICT and BIAS_INDEX_DICT pattern_to_input_type_to_index = get_pattern_to_input_type_to_index(backend_config_dict) for pattern, input_type_to_index in pattern_to_input_type_to_index.items(): for input_type, index in input_type_to_index.items(): index_dicts = { "weight": WEIGHT_INDEX_DICT, "bias": BIAS_INDEX_DICT, "input": {} # not used right now } assert input_type in index_dicts.keys(), \ f"input type must be one of {index_dicts.keys()} but got: {input_type}" index_dict = index_dicts[input_type] if pattern in index_dict: # type: ignore[operator] index_dict[pattern].append(index) # type: ignore[index] else: index_dict[pattern] = [index] # type: ignore[index] root_node_getter_mapping = \ get_fusion_pattern_to_root_node_getter(backend_config_dict) convert_dict_to_ordered_dict(qconfig_dict) convert_dict_to_ordered_dict(equalization_qconfig_dict) qconfig_dict = update_qconfig_for_fusion(model, qconfig_dict) equalization_qconfig_dict = update_qconfig_for_fusion(model, equalization_qconfig_dict) flattened_qconfig_dict = get_flattened_qconfig_dict(qconfig_dict) # TODO: support regex as well propagate_qconfig_(model, flattened_qconfig_dict, prepare_custom_config_dict) if is_qat: module_to_qat_module = get_module_to_qat_module(backend_config_dict) qat_swap_modules(model, module_to_qat_module) qconfig_dict = update_qconfig_for_qat(qconfig_dict, {}) # mapping from fully qualified module name to module instance # for example, # { # '': Model(...), # 'linear': Linear(...), # 'linear.weight_fake_quant': PerChannelMinMaxObserver(...), # } modules = dict(model.named_modules(remove_duplicate=False)) # fill qconfig_map, a map from node name to qconfig, used in find_matches equalization_qconfig_map = generate_qconfig_map(model, modules, model.graph, equalization_qconfig_dict, node_name_to_scope) qconfig_map = generate_qconfig_map(model, modules, model.graph, qconfig_dict, node_name_to_scope) # match the patterns that will get quantized standalone_module_name_configs = prepare_custom_config_dict.get( "standalone_module_name", []) standalone_module_class_configs = prepare_custom_config_dict.get( "standalone_module_class", []) standalone_module_names = [config[0] for config in standalone_module_name_configs] standalone_module_classes = [config[0] for config in standalone_module_class_configs] custom_module_classes = get_custom_module_class_keys( prepare_custom_config_dict, "float_to_observed_custom_module_class") matches = find_matches( model.graph, modules, patterns, root_node_getter_mapping, qconfig_map, standalone_module_names, standalone_module_classes, custom_module_classes) input_quantized_idxs: List[int] = prepare_custom_config_dict.get( "input_quantized_idxs", []) output_quantized_idxs: List[int] = prepare_custom_config_dict.get( "output_quantized_idxs", []) run_prepare_fx_on_standalone_modules( model, is_qat, modules, matches, prepare_custom_config_dict, backend_config_dict) # record names for the set of observed node, so that in convert step # we know whether we need to convert a floating point module to reference # quantized module or not observed_node_names: Set[str] = set() result_node = insert_observers_for_model( model, modules, matches, qconfig_map, model.graph, prepare_custom_config_dict, equalization_qconfig_map, input_quantized_idxs, output_quantized_idxs, backend_config_dict, observed_node_names, is_qat) save_state(model, qconfig_map, node_name_to_scope, prepare_custom_config_dict, equalization_qconfig_map, qconfig_dict, is_qat, observed_node_names) preserved_attributes = set(prepare_custom_config_dict.get("preserved_attributes", [])) model = ObservedGraphModule(model, model.graph, preserved_attributes) if is_standalone_module: assert result_node is not None assert isinstance(result_node.args[0], Node), \ "standalone module only supports returning simple value currently"\ "(not tuple, dict etc.)" # these inputs are observed in parent # converting List[int] to Tensor since module attribute is # Union[Tensor, Module] model._standalone_module_input_quantized_idxs = \ torch.tensor(input_quantized_idxs) model._standalone_module_output_quantized_idxs = torch.tensor(output_quantized_idxs) return model