Name	Size	Mode	Actions
accelerator_partitioner.cpython-36.pyc	29211	0644	editdlrm
const_fold.cpython-36.pyc	5899	0644	editdlrm
graph_gradual_typechecker.cpython-36.pyc	26871	0644	editdlrm
merge_matmul.cpython-36.pyc	5747	0644	editdlrm
normalize.cpython-36.pyc	5001	0644	editdlrm
optimization.cpython-36.pyc	14105	0644	editdlrm
partitioner_utils.cpython-36.pyc	9275	0644	editdlrm
refinement_types.cpython-36.pyc	921	0644	editdlrm
rewriter.cpython-36.pyc	4813	0644	editdlrm
schema_type_annotation.cpython-36.pyc	4037	0644	editdlrm
unify_refinements.cpython-36.pyc	2900	0644	editdlrm
__init__.cpython-36.pyc	157	0644	editdlrm

3 Eg>�@s�dd
lZddl
mZmZmZmZmZ
dd
lZdd
lZdd
lj Z ddlm Z mZ
ddlm Z mZmZmZ
ddlmZmZmZ
ddlmZ
Gdd �d e �ZGd d�de�Zd
S)�N)�Any�Callable�Dict�Tuple�Optional)�Transformer�Proxy)�Argument�Target�Node� map_aggregate)�normalize_module�normalize_function�create_type_hint�
)
�AnnotateTypesWithSchemac s�eZ
dZd
Zdejjed��f
dd� Ze e d��f
dd�Zdee ed feee fee e d feeee fd��f
dd � Zee ed feee fd��f
dd�Z�ZS)� NormalizeArgsa� Normalize arguments to Python targets. This means that `args/kwargs` will be matched up to the module/functional's signature and rewritten to exclusively kwargs in positional order if `normalize_to_only_use_kwargs` is true. Also populates default values. Does not support positional-only parameters or varargs parameters (*args, **kwargs). If the nodes have 'type' metadata, it will use it to disambiguate overloads. Otherwise, it will throw an error. Example usage: m = torchvision.models.resnet18() traced = torch.fx.symbolic_trace(m) traced = NormalizeArgs(traced).transform() T)�module�normalize_to_only_use_kwargscst�j
|
�

i|_||_dS)
N)�super�__init__�node_mapr)�selfrr)
� __class__��K/usr/local/lib64/python3.6/site-packages/torch/fx/experimental/normalize.pyr$s

zNormalizeArgs.__init__)�
n�returncs�|j�
�
\}}�
f
d
d��t
�
j��}t|t�s4t�
tdd�|D�
�
}�f
dd�|j�D�
}�
jdkr||j�
j ||||�}nt �j�
�
}�
jdkr��
|j|<�
j |j_ |S) Nc

s,t|t
j�r$d
�jkr �jd
SdSt|�
S)N�type)� isinstance�fxr�metar)
�arg)
rrr�get_type.s


z(NormalizeArgs.run_node..get_typec
Ssg|]}
t|
�
�qSr)
r)�.0�
irrr� 5sz*NormalizeArgs.run_node..c
si|]\}
}�|�
|
�qSrr)r$�
k�
v)
r#rr� 6sz*NormalizeArgs.run_node..� call_function�output)Zfetch_args_kwargs_from_envr�argsr�tuple�AssertionError�items�opr*�targetr�run_noderr!�node)rrr,�kwargs� arg_types�kwarg_types�out)
r)r#rrr2+s









zNormalizeArgs.run_nodeN.)r1r,r4r5r6c sRt|
�
st
�
t|
|||||j�}|r>|\}}|jjd
|
||�St�j|
||�SdS)Nr*)�callabler.rrZtracerZcreate_proxyrr*) rr1r,r4r5r6�new_args_and_kwargs�new_args� new_kwargs)
rrrr*@s









zNormalizeArgs.call_function)r1r,r4csRt|
t
�st�
t|j|
|||j�}|r>|\}}t�j|
||�St�j|
||�SdS)
N)r�strr.r rrr�call_module)rr1r,r4r9r:r;)
rrrr=Ys







zNormalizeArgs.call_module)
T)NN)�__name__� __module__�__qualname__�__doc__�torch�nn�Module�boolrrrr2r rr rr<rr*r=� __classcell__rr)
rrrs
Drcs�eZ
dZUd
Zejejejejejejej ej ejejej ejejejejejejejejejejejejejiZeeeegefeeegeffeeedfeeefd��f
dd�Z�ZS)�NormalizeOperatorsa�
Normalize callsites that are different ways of "spelling" the same invocation into a single, canonical call. Currently supports: 1. Normalize operators (e.g. operator.add) to the `torch` ops they ultimately invoke (e.g. torch.add) when it is possible to statically reason that Example usage: m = torchvision.models.resnet18() traced = torch.fx.symbolic_trace(m) traced = NormalizeOperators(traced).transform() .)r1r,r4csft|
�
st
�
|
|jkrVt|�
d
kr2t�j|
||�S|\}}t�j|j|
||fid�St�j|
||�S)N�)r1r,r4)r8r.�binary_magic_method_remap�lenrr*)rr1r,r4�lhs�rhs)
rrrr*�s





z NormalizeOperators.call_function)r>r?r@rArB�add�operator�mul�sub�div�truedivZfloor_divide�floordiv� remainder�mod�eq�ne�lt�le�gt�gerIrrrr rr r<r*rFrr)
rrrGks 










0rG)rN�typingrrrrrrBZtorch.fxr rrZ torch.fx.noder r rrZtorch.fx.operator_schemasr rrZschema_type_annotationrrrGrrrr�
s





Y