a h@sddlZddlmZmZmZddlZddlmZmZe edfZ eee e fge edffZ gdZeedddZejjeeefd d d Zeeeeefd d dZeedddZdS)N)AnyCallableUnion)tree_flatten_with_pathtree_map.)normalize_source_namemodule_to_nested_dicttrack_dynamism_across_examplesclone_and_convert_to_meta)namereturncCstdd|S)Nz\.([a-zA-Z_][a-zA-Z0-9_]*)z['\1'])resub)r rM/var/www/auris/lib/python3.9/site-packages/torch/fx/experimental/_dynamism.pyrsr)moduler c Csi}i|d<i|d<t|D]v}zZ|dsxtt||sxt||}t|tjjsxt|tt tj frxt |t urx|||<Wqt yYqYq0q|jddD]\}}||d|<q|jddD]\}}||d|<q|D]\}}t||d|<q|S)ziRecursively converts an nn.Module into a nested dictionary with explicit 'parameters' and 'modules' keys. _parametersZ_modules_F)Zrecurse)dir startswithcallablegetattr isinstancetorchnnModuleintfloatTensortypeboolNotImplementedErrorZnamed_parametersZ named_buffersZnamed_childrenr)rZ self_dict attr_name attr_valuer parambuffer submodulerrrrs2        r)example_inputsr cCsi}|D]}d|vr8t|dtjjr8t|d|d<t|\}}|D]\}}t|tttjfsdqHt|tjrt |j }d}n |f}d}||vrddt t |D|f||<n2||\} } | |krt | t |kr| tqt|D]\} } ||d| | qqHqi} |D]`\}\} }t dd| D}d d d d|D}|dj}|| vrni| |<|| ||<q| S) a This function analyzes a list of example inputs to determine the dynamism of their shapes. It tracks whether the dimensions of tensors or non-tensor values change across different examples. The function returns a dictionary where each key represents a path to a value in the input examples, and the corresponding value is a tuple indicating which dimensions are dynamic (i.e., change across examples). This helps in understanding how the structure of data varies across different instances. selfTFcSsg|] }tqSr)set).0rrrr Vz2track_dynamism_across_examples..rcss|]}t|dkVqdS)N)len)r*srrr br,z1track_dynamism_across_examples..Lcss|]}t|VqdS)N)str)r*krrrr0cr,)rrrrrrrrrtupleshaperanger.appendr) enumerateadditemsjoinkey)r'ZtrackingexZleaves_with_pathsrZkey_pathvaluer6Z is_tensorZdim_setsflagiZdimoutputZ _is_tensorZ final_dynZkey_strr=rrrr ;s<          r ) example_inputr cCsttddd}t||S)z This function takes a list of example inputs and for each tensor, clones it and converts it to device=meta. For non-tensor values, it keeps the reference. It uses pytree to handle nested structures recursively. )r?r cSs t|tjr|jddS|S)Nmeta)Zdevice)rrrcloneto)r?rrr transform_fnqs z/clone_and_convert_to_meta..transform_fn)rr)rCrGrrrr ksr )r typingrrrrZtorch.utils._pytreerrr5ZKeyPathrrZNonTensorShapeFn__all__r3rrrdictrlistr r rrrrs $ 0