def convert(
model: GraphModule, is_reference: bool = False,
convert_custom_config_dict: Dict[str, Any] = None,
is_standalone_module: bool = False,
_remove_qconfig_flag: bool = True,
convert_qconfig_dict: Dict[str, Any] = None,
backend_config_dict: Optional[Dict[str, Any]] = None) -> torch.nn.Module:
"""
We will convert an observed model (a module with observer calls) to a reference
quantized model, the rule is simple:
1. for each observer module call in the graph, we'll convert it to calls to
quantize and dequantize functions based on the observer instance
2. for weighted operations like linear/conv, we need to convert them to reference
quantized module, this requires us to know whether the dtype configured for the
weight is supported in the backend, this is done in prepare step and the result
is stored in observed_node_names, we can decide whether we need to swap the
module based on this set
standalone_module means it a submodule that is not inlined in
parent module, and will be quantized separately as one unit.
Returns a quantized standalone module, whether input/output is quantized is
specified by prepare_custom_config_dict, with
input_quantized_idxs, output_quantized_idxs, please
see docs for prepare_fx for details
"""
if convert_custom_config_dict is None:
convert_custom_config_dict = {}
node_name_to_scope, prepare_custom_config_dict, observed_node_names = restore_state(model)
qconfig_map: Dict[str, QConfigAny] = model._qconfig_map # type: ignore[assignment]
# TODO this should be removed now that gpu support for quantization is being supported.
# however in practice, as of 7/22/2021, certain functions that get called by convert expect
# only cpu arguments.
# As an example, in TestQuantizeFxModels.test_qat_functional_linear when device='cuda',
# fold_weight will call quantized::linear_prepack which doesn't support QuantizedCuda backend.
if not is_reference:
model.cpu()
# mapping from fully qualified module name to module instance
# for example,
# {
# '': Model(...),
# 'linear': Linear(...),
# 'linear.weight_fake_quant': PerChannelMinMaxObserver(...),
# }
# We use remove_duplicate=False here because torch.cat uses
# the same activation_post_process module instance but different names
modules = dict(model.named_modules(remove_duplicate=False))
# TODO refactor this code once we update the prepare logic to have additional information on
# which graph nodes have been observed and share that with convert to decide which observers to ignore.
if convert_qconfig_dict:
prepare_qconfig_dict: Dict[str, Dict[Any, Any]] = model._qconfig_dict # type: ignore[assignment]
modules_copy = copy.deepcopy(modules)
convert_dict_to_ordered_dict(convert_qconfig_dict)
if model._is_qat:
convert_qconfig_dict = update_qconfig_for_qat(convert_qconfig_dict, {})
convert_qconfig_dict = update_qconfig_for_fusion(model, convert_qconfig_dict)
compare_prepare_convert_qconfig_dict(prepare_qconfig_dict, convert_qconfig_dict) # type: ignore[arg-type]
convert_qconfig_map = generate_qconfig_map(model, modules_copy, model.graph, convert_qconfig_dict, node_name_to_scope)
# check the convert_qconfig_map generated and ensure that all the values either match what was set in prepare qconfig_map
# or are set to None in the convert_qconfig_map.
for k, v in qconfig_map.items():
assert k in convert_qconfig_map, 'Expected key {} in convert qconfig_map'.format(k)
if convert_qconfig_map[k] is not None:
assert qconfig_equals(v, convert_qconfig_map[k]), 'Expected k {} to have the same value in prepare qconfig_dict \
and convert qconfig_dict, found {} updated to {}.'.format(k, v, convert_qconfig_map[k])
qconfig_map = convert_qconfig_map
custom_module_classes = get_custom_module_class_keys(
convert_custom_config_dict,
"observed_to_quantized_custom_module_class")
custom_module_class_mapping = convert_custom_config_dict.get("observed_to_quantized_custom_module_class", {})
if model._equalization_qconfig_map is not None:
# If we want to do equalization then do the following:
# Calculate the equalization scale, update the observers with the scaled
# inputs, and scale the weight
weight_eq_obs_dict = update_obs_for_equalization(model, modules)
convert_eq_obs(model, modules, weight_eq_obs_dict)
# always run weight observers in the top level forward method
# for dynamic quant ops or weight only quant ops
run_weight_observers(model)
graph_inputs: List[str] = []
for node in model.graph.nodes:
if node.op == 'placeholder':
graph_inputs.append(node.name)
# TODO: move this outside of this function
def replace_observer_with_quantize_dequantize_node(
model: torch.nn.Module,
graph: Graph,
node: Node,
modules: Dict[str, torch.nn.Module],
node_name_to_scope: Dict[str, Tuple[str, type]],
qconfig_map: Dict[str, QConfigAny]) -> None:
""" Replace activation_post_process module call node with quantize and
dequantize node
Before:
... -> observer_0(x) -> ...
After:
... -> torch.quantize_per_tensor(x, ...) -> x.dequantize() -> ...
"""
assert modules is not None
assert isinstance(node.target, str)
module_path, prefix = get_module_path_and_prefix(node, node_name_to_scope, qconfig_map)
observer_module = modules[node.target]
maybe_quantize_node_info = get_quantize_node_info(observer_module)
# Skip replacing observers to quant/dequant nodes if the qconfigs of all
# consumers and producers of this observer are None
skip_replacement = all([
has_none_qconfig(n, qconfig_map) for n in
list(node.args) + list(node.users.keys())])
if skip_replacement or maybe_quantize_node_info is None:
# didn't find correponding quantize op and info for the observer_module
# so we just remove the observer
with graph.inserting_before(node):
node.replace_all_uses_with(node.args[0])
graph.erase_node(node)
else:
# otherwise, we can convert the observer moduel call to quantize/dequantize node
node_type, quantize_op, qparams = maybe_quantize_node_info
# replace observer node with quant - dequant node
with graph.inserting_before(node):
input_node = node.args[0]
inputs = [input_node]
for key, value in qparams.items():
# TODO: we can add the information of whether a value needs to
# be registered as an attribute in qparams dict itself
if key in ['_scale_', '_zero_point_']:
# For scale and zero_point values we register them as buffers in the root module.
# TODO: maybe need more complex attr name here
qparam_node = create_getattr_from_value(model, graph, module_path + prefix + key, value)
inputs.append(qparam_node)
else:
# for qparams that are not scale/zero_point (like axis, dtype) we store them as literals in the graph.
inputs.append(value)
quantized_node = graph.create_node(node_type, quantize_op, tuple(inputs), {})
dequantized_node = graph.call_method("dequantize", args=(quantized_node,))
node.replace_all_uses_with(dequantized_node)
graph.erase_node(node)
# this is a temporary hack for custom module, we may want to implement
# this properly after the custom module class design is finalized
def replace_observer_with_dequantize_node(node: Node, graph: Graph):
call_custom_module_node = node.args[0]
assert isinstance(call_custom_module_node, Node), \
f"Expecting the for call custom module node to be a Node, but got {call_custom_module_node}"
node.replace_all_uses_with(call_custom_module_node)
graph.erase_node(node)
insert_dequantize_node(call_custom_module_node, graph)
# additional state to override inputs to be quantized, if specified
# by the user
placeholder_node_seen_cnt = 0
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", [])
if backend_config_dict is None:
backend_config_dict = get_native_backend_config_dict()
root_module_to_quantized_reference_module = get_root_module_to_quantized_reference_module(backend_config_dict)
# convert tuples so that it can work with isinstance(module, tuple_of_classes)
root_module_classes = tuple(root_module_to_quantized_reference_module.keys())
qat_module_classes = get_qat_module_classes(backend_config_dict)
fused_module_classes = get_fused_module_classes(backend_config_dict)
statically_quantized_custom_module_nodes: Set[Node] = set()
for node in list(model.graph.nodes):
if node.op == 'placeholder':
cur_placeholder_node_idx = placeholder_node_seen_cnt
placeholder_node_seen_cnt += 1
if cur_placeholder_node_idx in input_quantized_idxs:
# Inputs are assumed to be quantized if the user specifid the
# input_quantized_idxs override.
# we need to dequantize the inputs since all operators took
# floating point inputs in reference quantized models
insert_dequantize_node(node, model.graph)
elif node.op == "output":
# If the argument is empty we don't need to do anything
if len(output_quantized_idxs) == 0:
continue
# Result are kept quantized if the user specified the
# output_quantized_idxs override.
# Remove the dequantize operator for the node in the end if any
return_node = node
output = node.args[0]
# outputs can be Node, list, tuple, dict, other cases are not supported yet
if isinstance(output, (list, tuple)):
for idx in output_quantized_idxs:
maybe_recursive_remove_dequantize(output[idx], return_node, model.graph)
elif isinstance(output, (Node, dict)):
# we treat dict as a single argument currently, but it can be extended
# to support {"key": dtype} after we change output_quantized_idxs to
# dict
if 0 in output_quantized_idxs:
maybe_recursive_remove_dequantize(output, return_node, model.graph)
else:
warnings.warn(f"Unsupported node type for output_quantized_idxs: {type(output)}")
elif node.op == "call_module":
if is_activation_post_process(modules[node.target]):
observed_node = node.args[0]
if observed_node in statically_quantized_custom_module_nodes:
replace_observer_with_dequantize_node(node, model.graph)
else:
replace_observer_with_quantize_dequantize_node(
model, model.graph, node, modules, node_name_to_scope,
qconfig_map)
elif is_observed_standalone_module(modules[node.target]):
convert_standalone_module(
node, modules, model, is_reference, backend_config_dict)
elif type(modules[node.target]) in set(
root_module_classes).union(qat_module_classes).union(fused_module_classes):
# extra check for fused module classes to make sure they are fused module classes
# of target modules
if type(modules[node.target]) in fused_module_classes and \
type(modules[node.target][0]) not in root_module_classes:
continue
convert_weighted_module(
node, modules, observed_node_names, qconfig_map, backend_config_dict)
elif type(modules[node.target]) in custom_module_classes:
convert_custom_module(
node, model.graph, modules, custom_module_class_mapping,
statically_quantized_custom_module_nodes)
preserved_attributes = set(convert_custom_config_dict.get("preserved_attributes", []))
model = QuantizedGraphModule(model, copy.deepcopy(model.graph), preserved_attributes)
# remove deadcode after converting observers to quant/dequant ops
model.graph.eliminate_dead_code()
model.recompile()
# TODO: maybe move this to quantize_fx.py
if not is_reference:
model = duplicate_dequantize_node(model)
model = duplicate_quantize_dynamic_node(model)
model = lower_to_fbgemm(model, qconfig_map, node_name_to_scope)
model = remove_quant_dequant_pairs(model)
model = remove_extra_dequantize(model)
# TODO: this looks hacky, we want to check why we need this and see if we can
# remove this
# removes qconfig and activation_post_process modules
if _remove_qconfig_flag:
_remove_qconfig(model)
return model
def convert_weighted_module(
node: Node,
modules: Dict[str, torch.nn.Module],
observed_node_names: Set[str],
qconfig_map: Dict[str, QConfigAny],
backend_config_dict: Dict[str, Any]):
""" Convert a weighted module to reference quantized module in the model
If the QConfig of a QAT module is not set, the module will still be converted to
a float module.
Args:
- node: The call_module node of the observed standalone module
- modules: named_module of original model
- observed_node_names: names for the set of observed fx node, we can skip
this conversion if the node is not observed
"""
original_module = modules[str(node.target)]
qconfig: QConfigAny = original_module.qconfig # type: ignore[assignment]
weight_post_process = None
qat_module_classes = get_qat_module_classes(backend_config_dict)
if isinstance(
original_module,
qat_module_classes):
# Converting qat module to a float module, we need to attch
# weight fake_quant to the module, weight fake_quant is assumed to be run during
# QAT so we don't need to run it again here
weight_post_process = original_module.weight_fake_quant
original_module = original_module.to_float() # type: ignore[operator]
# change qat module to float module
parent_name, name = _parent_name(node.target)
setattr(modules[parent_name], name, original_module)
is_observed = node.name in observed_node_names
# If a qconfig is not defined for this node, then skip converting to a reference module
if qconfig is None or has_none_qconfig(node, qconfig_map) or not is_observed:
return
# skip converting to reference quantized module if the qconfig is not supported
pattern_to_dtype_configs = get_pattern_to_dtype_configs(backend_config_dict)
dtype_configs = pattern_to_dtype_configs.get(type(original_module), [])
if not is_qconfig_supported_by_dtype_configs(qconfig, dtype_configs):
return
# TODO: rename weight_is_statically_quantized to weight_is_int8_quantized
is_weight_quantized = weight_is_quantized(qconfig)
# the condition for swapping the module to reference quantized module is:
# weights need to be quantized
if not is_weight_quantized:
return
fused_module = None
float_module = original_module
# extract the inidividual float_module and fused module
if isinstance(original_module, torch.nn.intrinsic._FusedModule):
fused_module = float_module
float_module = fused_module[0] # type: ignore[index]
# TODO: move this to the reference quantized module
# weight_qparams or weight_qparams dict
wq_or_wq_dict = {}
if isinstance(float_module, torch.nn.RNNCellBase):
weight_post_process_ih = qconfig.weight() # type: ignore[union-attr, operator]
weight_post_process_hh = qconfig.weight() # type: ignore[union-attr, operator]
weight_post_process_ih(float_module.weight_ih)
weight_post_process_hh(float_module.weight_hh)
weight_qparams_ih = get_qparam_dict(weight_post_process_ih)
weight_qparams_hh = get_qparam_dict(weight_post_process_hh)
wq_or_wq_dict = {
"weight_ih": weight_qparams_ih,
"weight_hh": weight_qparams_hh,
}
elif isinstance(float_module, torch.nn.LSTM):
# format for wq_or_wq_dict (flattened attributes):
# {"weight_ih_l0_scale": ..., "weight_ih_l0_qscheme": ..., ...}
for wn in float_module._flat_weights_names:
if hasattr(float_module, wn) and wn.startswith("weight"):
weight = getattr(float_module, wn)
weight_post_process = qconfig.weight() # type: ignore[union-attr, operator]
if weight_post_process.dtype == torch.qint8: # type: ignore[union-attr]
weight_post_process(weight) # type: ignore[operator, misc]
wq_or_wq_dict[wn] = get_qparam_dict(weight_post_process)
else:
# weight_post_process is None means the original module is not a QAT module
# we need to get weight_post_process from qconfig in this case
if weight_post_process is None:
weight_post_process = qconfig.weight() # type: ignore[union-attr, operator]
# 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
# Issue: https://github.com/pytorch/pytorch/issues/73941
weight_post_process(float_module.weight) # type: ignore[operator]
wq_or_wq_dict = get_qparam_dict(weight_post_process)
# We use the same reference module for all modes of quantization: static, dynamic, weight_only
# root_module_to_quantized_reference_module: module mapping from root (floating point) module class
# to quantized reference module class, e.g. nn.Conv2d to nn.quantized._reference.Conv2d
root_module_to_quantized_reference_module = get_root_module_to_quantized_reference_module(backend_config_dict)
ref_qmodule_cls = root_module_to_quantized_reference_module.get(type(float_module), None)
assert ref_qmodule_cls is not None, f"No reference quantized module class configured for {type(float_module)}"
ref_qmodule = ref_qmodule_cls.from_float(float_module, wq_or_wq_dict) # type: ignore[attr-defined]
if fused_module is not None:
fused_module[0] = ref_qmodule # type: ignore[operator]
else:
parent_name, name = _parent_name(node.target)
setattr(modules[parent_name], name, ref_qmodule)
