o
    ,h                     @   sV  U d dl Z d dlZd dlZd dlZd dlmZmZ d dlmZm	Z	m
Z
mZmZmZ d dlmZ d dlZd dlZd dlZd dlZd dlZd dlZd dlmZ d dlmZ d dlmZ d dlmZ d d	lmZ d d
l m!Z! erd dl"Z"d dl#Z#d dl$m%Z& d dl'Zd dl(Zd dl)Zd dl*Zd dl+Zda,ee- e.d< g dZ/de-fddZ0i Z1e2e3e4f e.d< 	dgde5deej6j7 de3fddZ8dee3 fddZ9de3fddZ:dd Z;de3fddZ<e=e>Z?G d d! d!eZ@dej6j7ddfd"d#ZAdeBe3d$f fd%d&ZCdej6j7deBed$f fd'd(ZDdej6j7defd)d*ZEdej6j7deBe3d$f fd+d,ZFd-eBe3d$f deBe3d$f fd.d/ZGd0eBeejHejIe4ejJeKejLe-f d$f deBd1 fd2d3ZMd4eBejHd$f d5eBd1 deBejHd$f fd6d7ZNd8edejHfd9d:ZOd;eejHe4eKe-f d<d=dejHfd>d?ZPd8ejHd@eejHejIe4ejJeKejLe-f deejHe4eKe-f fdAdBZQdCdDdEeBe3d$f dFeBejHd$f dGeBd1 dHeBe3d$f dIeBejHd$f dJeBd1 dKe-dLeBdM dNeBeejHejIe4ejJeKejLe-f d$f deBeejHe4eKe-f d$f fdOdPZRdCdDdEeBe3d$f dFeBejHd$f dGeBd1 dHeBe3d$f dIeBejHd$f dJeBd1 dKe-dLeBdM dNeBeejHejIe4ejJeKejLe-f d$f deBeejHe4eKe-f d$f fdQdRZSdSeTfdTdUZUG dVdW dWZVe jWG dXdY dYZXee3eBe3ee3ef f f ZYee.dZ< 	 e jWd[d\ed]d^G d_d` d`ZZed]d^G dadb dbZ[ed]d^ddcdej6j7ddeeeZee3ef f  fdedfZ\dS )h    N)MappingSequence)AnyCallableFinalOptionalTYPE_CHECKINGUnion)	TypeAlias)
FakeTensor)compatibility)FakeTensorProp)OperatorSupport)CALLABLE_NODE_OPS)_pytree_pybind_state_SUPPORT_ONNXRT)is_onnxrt_backend_supportedtorch_compile_backendOrtExecutionProviderOrtBackendOptions
OrtBackendreturnc                  C   s~   t du r=z,td td td ddl} ddl} ddl} ddlm}m}m	}m
} da W t S  ty<   da Y t S w t S )	a!  Returns ``True`` if ONNX Runtime dependencies are installed and usable
    to support TorchDynamo backend integration; ``False`` otherwise.

    Example::

        # xdoctest: +REQUIRES(env:TORCH_DOCTEST_ONNX)
        >>> import torch
        >>> if torch.onnx.is_onnxrt_backend_supported():
        ...     @torch.compile(backend="onnxrt")
        ...     def f(x):
        ...             return x * x
        ...     print(f(torch.randn(10)))
        ... else:
        ...     print("pip install onnx onnxscript onnxruntime")
        ...
    Nonnxruntimezonnxruntime.capi._pybind_state
onnxscriptr   )decomposition_tablefx_onnx_interpreterpasses
type_utilsTF)r   	importlibimport_module
torch.onnxtorch.onnx._internal%torch.onnx._internal._exporter_legacytorch.onnx._internal.fxr   r   r   r   ImportError)torchr   r   r   r    r(   [/var/www/html/scripts/venv/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.pyr   /   s    


r   _dumped_onnx_modelmodel_stringgraph_modulec                 C   s   t jdd}|sdS t|dd }| | d}t|d}||  W d   n1 s/w   Y  |t|< |durc| | d}t|d	d
d}|t|j W d   |S 1 s^w   Y  |S )a  Stores the onnx model into a file.
    The name is "{ONNXRT_DUMP_PATH}{N}.onnx"
    where *N* is the number of files already stored with
    this prefix.
    If graph_module is not None, the graph is stored as a string with
    the same filename except the extension (.txt).
    ONNXRT_DUMP_PATHN    z.onnxwbz.txtwzutf-8)encoding)osenvirongetr*   openwritestrgraph)r+   r,   prefixnfilenameffilename_txtr(   r(   r)   _dump_onnx_modelb   s"   

r@   c                   C   s   dgS )NCPUExecutionProviderr(   r(   r(   r(   r)   _infer_default_eps{   s   rB   namec                 C   s    t j rt jj|  dS dS )zIf PyTorch is installed with CUDA support, this starts NVTX range.

    Check torch.cuda.nvtx.range_push's document for more details.
    N)r'   cudais_availablenvtx
range_pushrC   r(   r(   r)   _nvtx_range_push   s   
rI   c                   C   s   t j rt jj  dS dS )zIf PyTorch is installed with CUDA support, this terminates NVTX range.

    Check torch.cuda.nvtx.range_pop's document for more details.
    N)r'   rD   rE   rF   	range_popr(   r(   r(   r)   _nvtx_range_pop   s   
rK   device_typec                 C   sN   ddl m} | dkr|j S | dkr|j S | dkr!|j S td|  )Nr   r   rD   cpumaiazUnsupported device type: )onnxruntime.capir   	OrtDevicerD   rM   npu
ValueError)rL   ORTCr(   r(   r)   _get_ort_device_type   s   


rT   c                       s`   e Zd ZdZdee deeef f fddZde	ee
jjf de
jjdef fd	d
Z  ZS )OrtOperatorSupporta0  Operator support for ONNXRuntime backend.

    It has two-level of support decision. One is via support_dict and the other one
    is via extra_support_dict. The logic of using support_dict is implemented in
    OrtOperatorSupport and extra_support_dict is used by OperatorSupport.is_node_supported.
    support_dictextra_support_dictc                    s   t  | || _d S N)super__init___onnx_support_dict)selfrV   rW   	__class__r(   r)   rZ      s   
zOrtOperatorSupport.__init__
submodulesnoder   c                    s   |j tvrdS |j dkr|j| jv rtd|jt|j dS t ||r3td|jt|j dS t	d|jt|j dS )NFcall_functionz0support_dict supports node.target: %s (type: %s)Tz6extra_support_dict supports node.target: %s (type: %s)zLsupport_dict and extra_support_dict don't support node.target: %s (type: %s))
opr   targetr[   loggerinfotyperY   is_node_supportedwarning)r\   r_   r`   r]   r(   r)   rg      s,   
z$OrtOperatorSupport.is_node_supported)__name__
__module____qualname____doc__setr   dictr9   rZ   r   r'   nnModulefxNodeboolrg   __classcell__r(   r(   r]   r)   rU      s    "rU   c                 C   sh   | j }g }d}|jD ]}|jdkr|| |du r!|jdkr!|}q
|du r(dS |D ]}|| q*dS )z
    In torch.fx.Graph, placeholder is a special assignment node. If it's not
    executed in the beginning, it could overwrite values computed by upstream
    nodes.
    Nplaceholder)r:   nodesrb   appendprepend)r,   r:   placeholdersfirst_not_placeholderr`   ru   r(   r(   r)   _move_placeholder_to_front   s   


r{   .c                  G   sP   g }| D ]}t |dr#|j}|jdkr|d q|jdkr#|d qt|S )zBReturn the first valid device (i.e., GPU or CPU) in argument list.devicerD   CUDAExecutionProviderrM   rA   )hasattrr|   rf   rw   tuple)argsepsargr|   r(   r(   r)   _infer_ep_from_device   s   



r   c                 C   sX   g }| j jD ]!}|jdkr't|dr"d|jv r"t|jd tjs"J || qt	|S )Nru   metaval)
r:   rv   rb   r~   r   
isinstancer'   Tensorrw   r   )r,   ry   r`   r(   r(   r)   _extract_graph_module_inputs   s   

r   c                 C   s.   | j jD ]}|jdkr|jd   S qtd)zHCollect "val" fields from outputs metadata in this torch.fx.GraphModule.outputr   z2No output node found in this torch.fx.GraphModule.)r:   rv   rb   r   rR   )r,   r`   r(   r(   r)   _extract_graph_module_outputs  s
   
r   c                 C   s(   t t| \}}dd |D }t| S )z[Return the all valid devices (i.e., GPU or CPU) among outputs of this torch.fx.GraphModule.c                 S   s*   g | ]}t |d rd|jv r|jd qS )r   r   r~   r   ).0
output_argr(   r(   r)   
<listcomp>  s    
z/_infer_ep_from_graph_module.<locals>.<listcomp>)r   tree_flattenr   r   )r,   flattened_output_args_selected_output_argsr(   r(   r)   _infer_ep_from_graph_module  s   r   r   c                 C   s,   dt dtfdd}t| }tt||ddS )z:Sort execution providers in eps based on pre-set priority.epr   c                 S   s   | dkrdS | dkrdS dS )NrA      r}   r0   r   r(   )r   r(   r(   r)   get_execution_provider_priority!  s
   z2_sort_eps.<locals>.get_execution_provider_priorityT)keyreverse)r9   intrm   r   sorted)r   r   
unique_epsr(   r(   r)   	_sort_eps  s   r   valueszORTC.OrtDevice.c              	      s   ddl m  dtdtfdddttjtjttjttj	t
f dtf fdd	t| dkr;tfd
d| D }|S dfS )Nr   r   	device_idr   c                 S   s   | pdS )Nr   r(   )r   r(   r(   r)   _device_id_or_zero:  s   z-_get_onnx_devices.<locals>._device_id_or_zerovaluec                    sx   t | tjr t| jj j | jjS t | tj	t
tjttjtfr2 td j dS tdtt|  )NrM   r   zUnsupported value type: )r   r'   r   rP   rT   r|   rf   default_memoryindexSymIntr   SymFloatfloatSymBoolrs   rR   r9   r   )rS   r   r(   r)   _map_tensor_or_sym_to_device=  s   

z7_get_onnx_devices.<locals>._map_tensor_or_sym_to_devicec                 3   s    | ]} |V  qd S rX   r(   )r   r   )r   r(   r)   	<genexpr>R  s    z$_get_onnx_devices.<locals>.<genexpr>r0   )rO   r   r   r	   r'   r   r   r   r   r   rs   lenr   )r   ort_devicesr(   )rS   r   r   r)   _get_onnx_devices0  s   
r   tensorsdevicesc           
      C   s   dd l }ddlm} tj|jtj|jtj|jtj|jtj|jtj	|j	tj
|j
tj|jtj|ji	}| }|t|  g }g }g }| D ]}	|||	j  ||	  ||	  qC|| |||| |S )Nr   r   )numpyrO   r   r'   float16float32float64uint8int8int16int32int64longlongrs   bool_OrtValueVectorreserver   rw   dtypesizedata_ptrpush_back_batch)
r   r   nprS   torch_dtype_to_numpy_dtype	ortvaluesdtypesshapes	data_ptrstensorr(   r(   r)   !_get_ortvalues_from_torch_tensorsX  s.   r   r   c                 C   s*   | j rtdtj|  | j| jd}|S )Nz#sparse tensor is not yet supported.)r   r|   )	is_sparserR   r'   emptyr   r   r|   )r   outr(   r(   r)   _to_real_tensorx  s   r   dynamo_value
value_infoonnx.ValueInfoProtoc                 C   s   t | tjrt|jjjjdkr| jdkrt| S t | t	r'tj
| tjdS t | tr4tj
| tjdS t | trAtj
| tjdS t | tjsIJ |  S )z9Helper function to wrap PyTorch variables as torch.Tensorr   )r0   )r   )r   r'   r   r   rf   tensor_typeshapedimsqueezer   r   r   r   r   rs   
contiguous)r   r   r(   r(   r)   _adjust_scalar_from_fx_to_onnx  s   





r   
prim_valuec                 C   s<   t | tjs
J dt |tjttjttjtfr| 	 S | S )zFHelper function to wrap ORT-produced torch.Tensor as PyTorch variableszORT's output must be tensor.)
r   r'   r   r   r   r   r   r   rs   item)r   r   r(   r(   r)   _adjust_scalar_from_onnx_to_fx  s   r   sessonnxruntime.InferenceSessioninput_namesinputsinput_devicesoutput_namesoutputsoutput_devicespreallocate_outputinput_value_infosr   .normalized_prim_outputsc
                 C   s"  dd l }
ddlm} td tdd t||D }t  td t||}|r7tdd |D }t||}n| }t  td |
	 }|
d	d
 | |||||| t  |rptd tdd t||	D }t  |S dd l}
td |
jjj|}tdd t||	D }t  |S )Nr   r   r   c                 s       | ]
\}}t ||V  qd S rX   r   r   r   r   r(   r(   r)   r     
    
z8_run_onnx_session_with_ortvaluevector.<locals>.<genexpr>r   c                 s   s&    | ]}t |trt|n|V  qd S rX   )r   r   r   )r   tr(   r(   r)   r     s    
run_with_ortvaluevector'disable_synchronize_execution_providers1zafter run_with_ortvaluevectorc                 s   r   rX   r   r   onnx_outputprim_outputr(   r(   r)   r     r   c                 s   r   rX   r   r   r(   r(   r)   r     r   )r   rO   r   rI   r   ziprK   r   r   
RunOptionsadd_run_config_entryr   onnxruntime.trainingtraining	ortmodule_utils_ortvalues_to_torch_tensor)r   r   r   r   r   r   r   r   r   r   r   rS   
ort_inputspth_outputsort_outputsrun_optionsr(   r(   r)   %_run_onnx_session_with_ortvaluevector  sP   

r   c
                    s`   dd l  tdd t||D } fddt||D }
| ||
}tdd t||	D }|S )Nr   c                 s   r   rX   r   r   r(   r(   r)   r   !  r   z/_run_onnx_session_with_fetch.<locals>.<genexpr>c                    s&   i | ]\}}| j |  qS r(   )OrtValueortvalue_from_numpyrM   r   )r   rC   r   r   r(   r)   
<dictcomp>%  s    z0_run_onnx_session_with_fetch.<locals>.<dictcomp>c                 s   s$    | ]\}}t t||V  qd S rX   )r   r'   
from_numpy)r   r   r   r(   r(   r)   r   *  s    
)r   r   r   run)r   r   r   r   r   r   r   r   r   r   feedr   r   r(   r   r)   _run_onnx_session_with_fetch  s   
r  rf   c                 C   s.   ddl }t|jjt|jjt|jji}|| S )a=  
    Converts a Python type to the corresponding ONNX tensor element type.
    For example, `_from_python_type_to_onnx_tensor_element_type(float)` returns
    `onnx.TensorProto.FLOAT`.

    Args:
      type (type): The Python type to convert.

    Returns:
      int: The corresponding ONNX tensor element type.

    r   N)	onnxr   TensorProtoFLOATr   INT64rs   BOOLr6   )rf   r  (_PYTHON_TYPE_TO_ONNX_TENSOR_ELEMENT_TYPEr(   r(   r)   -_from_python_type_to_onnx_tensor_element_type4  s   
r  c                   @   sv   e Zd ZdZdddeedf ded deedf d	ed d
ed ded deeejdf ejf fddZ	dd Z
dS )OrtExecutionInfoPerSessionzWInformation required to execute torch.fx.GraphModule using onnxruntime.InferenceSessionsessionr   r   .r   r   r   output_value_infosr   r   r   example_outputsc	           	      C   s4   || _ || _|| _|| _|| _|| _|| _|| _d S rX   r  r   r   r   r  r   r   r  )	r\   r  r   r   r   r  r   r   r  r(   r(   r)   rZ   N  s   z#OrtExecutionInfoPerSession.__init__c           
      G   s  dd l }|jjtj|jjtj|jjtj|jj	tj
|jjtj|jjtj|jjtj|jjtj|jjtj|jjtj|jjtj|jjtj|jjtji}dd | D }t|t| jkr[dS t || jD ]o\}}t!|tj"t#t$fsq dS t!|t$t#tfrt%t&|}||j&j'j(kr dS t|j&j'j)j*dkr dS qa||j+ }||j&j'j(kr dS t |j)|j&j'j)j*D ]\}}	t!|t$r|	j,|ks|	j-rqt!|tj.r|	j-rq  dS qadS )Nr   c                 S   s   i | ]\}}||qS r(   r(   )r   r   r   r(   r(   r)   r    s    z;OrtExecutionInfoPerSession.is_supported.<locals>.<dictcomp>FT)/r  r  r  r'   r   FLOAT16r   
FLOAT8E5M2float8_e5m2FLOAT8E5M2FNUZfloat8_e5m2fnuzFLOAT8E4M3FNfloat8_e4m3fnFLOAT8E4M3FNUZfloat8_e4m3fnuzDOUBLEr   r
  rs   UINT8r   INT8r   INT16r   INT32r   r	  r   itemsr   r   r   r   r   r   r   r  rf   r   	elem_typer   r   r   	dim_value	dim_paramr   )
r\   r   r  (_onnx_tensor_element_type_to_torch_dtype(_torch_dtype_to_onnx_tensor_element_typer   r   
onnx_dtyper   onnx_dimr(   r(   r)   is_supportedo  sT   














	z'OrtExecutionInfoPerSession.is_supportedN)ri   rj   rk   rl   r   r9   r	   r'   r   rZ   r(  r(   r(   r(   r)   r  K  s(    

	
!r  c                   @   s>   e Zd ZdddZdejjfddZdejjdefd	d
Z	dS )"OrtExecutionInfoForAllGraphModulesr   Nc                 C   s
   i | _ d S rX   )execution_info_per_graph_module)r\   r(   r(   r)   rZ     s   z+OrtExecutionInfoForAllGraphModules.__init__r,   c                 G   s8   || j vrd S | j | }|D ]}|j| r|  S qd S rX   )r*  r(  )r\   r,   r   
candidates	candidater(   r(   r)   &search_reusable_session_execution_info  s   


zIOrtExecutionInfoForAllGraphModules.search_reusable_session_execution_infore   c                 C   s.   || j vr|g| j |< d S | j | | d S rX   )r*  rw   )r\   r,   re   r(   r(   r)   cache_session_execution_info  s   
z?OrtExecutionInfoForAllGraphModules.cache_session_execution_info)r   N)
ri   rj   rk   rZ   r'   rq   GraphModuler-  r  r.  r(   r(   r(   r)   r)    s    

r)  r   T)frozenF)is_backward_compatiblec                   @   s   e Zd ZU dZdZeee  ed< 	 dZ	e
ed< 	 dZeee  ed< 	 dZe
ed< 	 dZe
ed	< 	 dZed
 ed< 	 dZeeedgdf   ed< dS )r   aJ  Options for constructing an ``OrtBackend``, the ONNX Runtime
    backend (``"onnxrt"``) for ``torch.compile``.

    Example::

        >>> @torch.compile(
        ...     backend="onnxrt",
        ...     options=torch.onnx._OrtBackendOptions(...),
        ... )
        ... def ort_function(x):
        ...     return x ** x
    Npreferred_execution_providersTinfer_execution_providersdefault_execution_providersFr   use_aot_autogradzonnxruntime.SessionOptionsort_session_optionszonnx.ModelProtopre_ort_model_transforms)ri   rj   rk   rl   r2  r   r   r   __annotations__r3  rs   r4  r   r5  r6  r7  r   r(   r(   r(   r)   r     s&   
 	r   c                	   @   s   e Zd ZU dZddee fddZdejj	de
eeeeef f  fdd	Zdejj	fd
dZdejj	dejj	fddZdejj	dejj	fddZdZeed< g Zeed   ed< e	ddeeeeeef f  dd fddZedd Zedd ZdS )r   a	  A backend compiles (sub-)graphs in torch.fx.GraphModule to onnxruntime.InferenceSession calls.

    The compiler entry point is OrtBackend.compile, which
        1. partitions the original graph into supported sub-graphs (type: torch.fx.GraphModule) and unsupported
           sub-graphs.
        2. For each supported sub-graph, it replaces its _wrapped_call function with _ort_accelerated_call.
        3. Inside _ort_accelerated_call, it creates onnxruntime.InferenceSession and calls it to execute the sub-graph.
    Noptionsc                 C   s   ddl m} dd l}dd l}dd l}|d u rt n|| _|jjj	
 | _|jjjj| jj}d d d d d d}t||| _i | _t | _d| _d| _t|jdrUt| _d S t| _d S )Nr   r   )getattrz_operator.getitemz_operator.mulz_operator.addz_operator.subFr   )rO   r   r"   r$   +torch.onnx._internal.fx.decomposition_tabler   _optionsr  	_internal_exporter_legacyResolvedExportOptions_resolved_onnx_exporter_optionsrq   r   '_create_onnx_supports_op_overload_tableonnx_registryrU   _supported_ops_partitioner_cacher)  _all_ort_execution_info_assert_allclose_to_baselineexecution_countr~   r   r   r  r  )r\   r9  rS   r'   rV   rW   r(   r(   r)   rZ   !  s4   
zOrtBackend.__init__r,   r   c                 G   s   d}| j jrt|  }r|}nt| }r|}g }g | j jpg t|| j jp*t R D ]*}t|t	r9|i f}nt|t
rJ|d d u rJ|d i f}|d urW||vrW|| q-|S )Nr(   r0   r   )r<  r3  r   r   r2  r   r4  rB   r   r9   r   rw   )r\   r,   r   inferred_epseps_from_argseps_from_graph_moduleselected_epsr   r(   r(   r)   _select_epsu  s,   




zOrtBackend._select_epsc                 O   s  ddl }ddlm}m} | jj|g|R  }|r1|j}|j}	|j}
|j	}|j
}|j}|j}|j}n|| }| jjrNd| _t|}dd }t||}nzt|j|i |}W n tyk   td| d| _ w | }|| }|j|| jjd}|j| jjj d	}| j!j"r| j!j"D ]}|| q|# }t$j%&d
drt'||d |j(|| j!j)| j*|g|R  d}t+dd |j,j-D }	t+dd |j,j.D }
t/|}t0|t+rt/|}nt/|f}t+dd |j,j-D }t+dd |j,j.D }t1||	||
||||d}| j2|| |  j3d7  _3t0|t4j5}|r|fn|}t0|t+s)J t6dd |D s5J t7d | ||	|||
||| j!j||
}t8  | j9rxt4j:j;j<|g|R ddi}|rd|fn|}t=||D ]\}}t4j>?|| qk|r|d S |S )a  This function replaces GraphModule._wrapped_call in compiled model.

        The _wrapped_call is the underlying implementation of forward method. Replacing
        it means we delegate the computation to _ort_acclerated_call and therefore
        onnxruntime.InferenceSession.
        r   N)r   r   Fc                 S   s"   t | drd| jv r| jd S | S )Nr   r   r   r   r(   r(   r)   maybe_map_to_meta_val  s   
z>OrtBackend._ort_acclerated_call.<locals>.maybe_map_to_meta_valzFakeTensorProb failed for %s)fx_graph_moduleonnxfunction_dispatcher)opset_versionr-   )r,   )path_or_bytessess_options	providersc                 s       | ]}|j V  qd S rX   rH   r   inputr(   r(   r)   r         z2OrtBackend._ort_acclerated_call.<locals>.<genexpr>c                 s   rT  rX   rH   r   r   r(   r(   r)   r     rW  c                 s       | ]}|V  qd S rX   r(   rU  r(   r(   r)   r         c                 s   rY  rX   r(   rX  r(   r(   r)   r     rZ  r  r0   c                 s   s$    | ]}t |tjtjtfV  qd S rX   )r   r'   r   r   r   )r   elemr(   r(   r)   r   7  s
    
$run_onnx_session_with_ortvaluevectorexecutoraten)@r   r%   r   r   rE  r-  r  r   r   r   r  r   r   r  MovePlaceholderToFrontr  r@  dynamic_shapesr   r   r   tree_mapr   	propagate	Exceptionrd   rh   FxOnnxInterpreterInsertTypePromotionrO  to_model_protorB  rP  r<  r7  SerializeToStringr4   r5   r6   r@   InferenceSessionr6  rL  r   r:   rV  r   r   r   r  r.  rG  r'   r   allrI   rK   rF  _primsr]  executer   testingassert_close)r\   r,   r   kwargsr   r   r   !cached_execution_info_per_sessiononnx_sessionr   r   r   r  r   r   prim_outputsextracted_outputsrM  fx_interpreterexported
onnx_model	transformonnx_model_bytesexecution_info_per_sessionis_single_tensor_outputr   onnx_outputsbaseline_outputsnormalized_baseline_ouptutsr   baseline_outputr(   r(   r)   _ort_acclerated_call  s   	

	


zOrtBackend._ort_acclerated_callc           	      C   s   ddl m} || jv r| j| }|S |}||| jdd}| }|| j|< |jjD ]}|jdkr?d|jv r?t	||j}| j
|_q)|S )Nr   )CapabilityBasedPartitionerT)allows_single_node_partitioncall_modulefused_)!torch.fx.passes.infra.partitionerr  rD  rC  partition_and_fuser:   rv   rb   rC   r:  r~  _wrapped_call)	r\   r,   r   r  partitioned_prim_graph_moduleprim_graph_modulepartitionerr`   fused_moduler(   r(   r)   compileZ  s$   


zOrtBackend.compilec                 C   sF   | j jrddlm} ddlm} || j|| jjd||S | ||S )zIf ``OrtBackendOptions.use_aot_autograd`` is ``True``, the `auto_autograd` compiler
        will be invoked, wrapping this ``OrtBackend`` instance's ``compile`` method. Otherwise,
        the ``compile`` method is invoked directly.r   )#min_cut_rematerialization_partition)aot_autograd)fw_compilerpartition_fndecompositions)	r<  r5  functorch.compiler  torch._dynamo.backends.commonr  r  r@  r   )r\   r,   r   r  r  r(   r(   r)   __call__  s   zOrtBackend.__call__   %_OrtBackend__instance_cache_max_count_OrtBackend__instance_cachec                    s   dt dt fddt t st di  pi  t fddtjD d}|du rJttjtjk s@J dtj d	t d
t dtjt  } |S )a  Returns a possibly cached instance of an ``OrtBackend``. If an existing
        backend was created previously through this function with the same options,
        it will be returned. Otherwise a new backend will be created, cached, and
        returned.

        Note: if ``options`` sets ``ort_session_options``, a new ``OrtBackend``
        will always be returned, since ``onnxruntime.SessionOptions`` cannot
        participate in caching.abc                 S   sh   | j |j ks$| j|jks$| j|jks$| j|jks$| j|jks$| j|jkr&dS | jd us0|jd ur2dS dS )NFT)r2  r3  r4  r   r5  r7  r6  )r  r  r(   r(   r)   reusable  s   z<OrtBackend.get_cached_instance_for_options.<locals>.reusablec                 3   s     | ]}|j  r|V  qd S rX   )r<  )r   r  r9  r  r(   r)   r     s    z=OrtBackend.get_cached_instance_for_options.<locals>.<genexpr>NzNo more than z instances of z allowed. Please instantiate `z` explicitly to pass to `torch.compile`. See https://github.com/pytorch/pytorch/pull/107973#discussion_r1306144795 for discussion.r(   )r   r   nextr   r  r   r  rw   )r9  backendr(   r  r)   get_cached_instance_for_options  s"   

z*OrtBackend.get_cached_instance_for_optionsc                   C   s   t j  d S rX   )r   r  clearr(   r(   r(   r)   clear_cached_instances  s   z!OrtBackend.clear_cached_instancesc                   C   s
   t tjS rX   )r   r   r  r(   r(   r(   r)   get_cached_instances  s   
zOrtBackend.get_cached_instancesrX   )ri   rj   rk   rl   r   r   rZ   r'   rq   r/  r   r   r9   r   r   rL  r~  r  r  r  r   r8  r  liststaticmethodr	   r  r  r  r(   r(   r(   r)   r     s:   
 	T
! E6
6
r   )r9  r9  c                C   s   t || |S rX   )r   r  )r,   r   r9  r(   r(   r)   r     s   r   rX   )]dataclassesr    loggingr4   collections.abcr   r   typingr   r   r   r   r   r	   typing_extensionsr
   r'   torch._C
torch._opstorch._prims.executortorch.fx!torch.onnx._internal._lazy_importtorch._subclasses.fake_tensorr   torch.fx._compatibilityr    torch.fx.passes.fake_tensor_propr    torch.fx.passes.operator_supportr   torch.fx.passes.tools_commonr   torch.utilsr   r  r   rO   r   rS   r"   r#   r$   r;  torch.onnx._internal.fx.passesr   rs   r8  __all__r   r*   rn   r9   r   bytesrq   r/  r@   rB   rI   rK   rT   	getLoggerri   rd   rU   r{   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r  rf   r  r  	dataclassr)  r   r   r   r   r(   r(   r(   r)   <module>   sr  
 	0

		
2

"
(
 
 


	

S

	

&^$!
<   Q