o
    wZh                     @   sH  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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$Z% d dl&Zd dl'Zd dl(Zd dl)Zd dl*Zd dl+Zda,ee- e.d< g dZ/de-fddZ0i Z1e2e3e4f e.d< 	ddde5deej6j7 de3fddZ8dee3 fddZ9de3fddZ:dd Z;de3fddZ<e=e>Z?G d d! d!eZ@dej6j7ddfd"d#ZAdeBe3d$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ZSG dSdT dTZTe jUG dUdV dVZVee3eBe3ee3ef f f ZWee.dW< 	 e jUdXdYedZd[G d\d] d]ZXedZd[G d^d_ d_ZYedZd[dd`dej6j7daeeeXee3ef f  fdbdcZZdS )e    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 rAz0td td td ddl} 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_stateZ
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.diagnosticstorch.onnx._internal.fxr   r   r   r   ImportError)torchr   r   r   r    r(   O/var/www/auris/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fZfilename_txtr(   r(   r)   _dump_onnx_modelc   s"   

r?   c                   C   s   dgS )NCPUExecutionProviderr(   r(   r(   r(   r)   _infer_default_eps|   s   rA   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nvtxZ
range_pushrB   r(   r(   r)   _nvtx_range_push   s   
rG   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'   rC   rD   rE   Z	range_popr(   r(   r(   r)   _nvtx_range_pop   s   
rH   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   rC   cpuZmaiazUnsupported device type: )onnxruntime.capir   	OrtDevicerC   rJ   Znpu
ValueError)rI   ORTCr(   r(   r)   _get_ort_device_type   s   


rO   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)selfrQ   rR   	__class__r(   r)   rU      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 )NFZ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   targetrV   loggerinfotyperT   is_node_supportedwarning)rW   rZ   r[   rX   r(   r)   ra      s,   
z$OrtOperatorSupport.is_node_supported)__name__
__module____qualname____doc__setr   dictr9   rU   r   r'   nnModulefxNodeboolra   __classcell__r(   r(   rX   r)   rP      s    "rP   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:   nodesr\   appendprepend)r,   r:   placeholdersZfirst_not_placeholderr[   ro   r(   r(   r)   _move_placeholder_to_front   s   


rt   .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.devicerC   CUDAExecutionProviderrJ   r@   )hasattrru   r`   rq   tuple)argsepsargru   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 )Nro   metaval)
r:   rp   r\   rw   r}   
isinstancer'   Tensorrq   rx   )r,   rs   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:   rp   r\   ry   rM   )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~   rw   r}   ).0Z
output_argr(   r(   r)   
<listcomp>  s    
z/_infer_ep_from_graph_module.<locals>.<listcomp>)r   Ztree_flattenr   r|   )r,   Zflattened_output_args_Zselected_output_argsr(   r(   r)   _infer_ep_from_graph_module  s   r   rz   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 )Nr@      rv   r0   r   r(   )r   r(   r(   r)   get_execution_provider_priority"  s
   z2_sort_eps.<locals>.get_execution_provider_priorityT)keyreverse)r9   intrg   rx   sorted)rz   r   Z
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|  )NrJ   r   zUnsupported value type: )r   r'   r   rL   rO   ru   r`   Zdefault_memoryindexSymIntr   SymFloatfloatSymBoolrm   rM   r9   r   )rN   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 rS   r(   )r   r   )r   r(   r)   	<genexpr>S  s    z$_get_onnx_devices.<locals>.<genexpr>r0   )rK   r   r   r	   r'   r   r   r   r   r   rm   lenrx   )r   Zort_devicesr(   )rN   r   r   r)   _get_onnx_devices1  s   
r   tensorsdevicesc           	      C   s   ddl m} ddlm} | }|t|  g }g }g }| D ]}|||j  ||	  ||
  q|| |||| |S )Nr   r   )_TORCH_DTYPE_TO_NUMPY_DTYPE)rK   r   "torch.onnx._internal.fx.type_utilsr   OrtValueVectorZreserver   rq   dtypesizeZdata_ptrpush_back_batch)	r   r   rN   r   Z	ortvaluesZdtypesZshapesZ	data_ptrstensorr(   r(   r)   !_get_ortvalues_from_torch_tensorsY  s   r   r   c                 C   s*   | j rtdtj|  | j| jd}|S )Nz#sparse tensor is not yet supported.)r   ru   )Z	is_sparserM   r'   emptyr   r   ru   )r   outr(   r(   r)   _to_real_tensorn  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   r`   tensor_typeshapedimZsqueezer   r   Zint64r   Zfloat32rm   
contiguous)r   r   r(   r(   r)   _adjust_scalar_from_fx_to_onnxu  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   rm   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 td |
jjj|}tdd t||	D }t  |S )Nr   r   r   c                 s       | ]
\}}t ||V  qd S rS   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 rS   )r   r   r   )r   tr(   r(   r)   r     s    
run_with_ortvaluevectorZ'disable_synchronize_execution_providers1zafter run_with_ortvaluevectorc                 s   r   rS   r   r   onnx_outputprim_outputr(   r(   r)   r     r   c                 s   r   rS   r   r   r(   r(   r)   r     r   )r   rK   r   rG   rx   ziprH   r   r   Z
RunOptionsZadd_run_config_entryr   ZtrainingZ	ortmodule_utilsZ_ortvalues_to_torch_tensor)r   r   r   r   r   r   r   r   r   r   r   rN   Z
ort_inputspth_outputsort_outputsZrun_optionsr(   r(   r)   %_run_onnx_session_with_ortvaluevector  sN   

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   rS   r   r   r(   r(   r)   r     r   z/_run_onnx_session_with_fetch.<locals>.<genexpr>c                    s&   i | ]\}}| j |  qS r(   )ZOrtValueZortvalue_from_numpyrJ   numpy)r   rB   r   r   r(   r)   
<dictcomp>  s    z0_run_onnx_session_with_fetch.<locals>.<dictcomp>c                 s   s$    | ]\}}t t||V  qd S rS   )r   r'   Z
from_numpy)r   r   r   r(   r(   r)   r     s    
)r   rx   r   run)r   r   r   r   r   r   r   r   r   r   feedr   r   r(   r   r)   _run_onnx_session_with_fetch  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 rS   r   r   r   r   r   r   r   r   )	rW   r   r   r   r   r   r   r   r   r(   r(   r)   rU   +  s   z#OrtExecutionInfoPerSession.__init__c           	      G   s  ddl m}m} t|t| jkrdS t|| jD ]o\}}t|tjt	t
fs) dS t|t
t	tfrO|t|}||jjjkrA dS t|jjjjdkrN dS q||j }||jjjkr^ dS t|j|jjjjD ]\}}t|t
rz|j|ksy|jrzqht|tjr|jrqh  dS qdS )Nr   )(_TORCH_DTYPE_TO_ONNX_TENSOR_ELEMENT_TYPE,from_python_type_to_onnx_tensor_element_typeFT)r   r   r   r   r   r   r   r'   r   r   r   rm   r`   r   Z	elem_typer   r   r   Z	dim_valueZ	dim_paramr   )	rW   ry   r   r   r{   r   Z
onnx_dtyper   Zonnx_dimr(   r(   r)   is_supportedL  s2   

	z'OrtExecutionInfoPerSession.is_supportedN)rc   rd   re   rf   rx   r9   r	   r'   r   rU   r   r(   r(   r(   r)   r   (  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 rS   )execution_info_per_graph_module)rW   r(   r(   r)   rU   v  s   z+OrtExecutionInfoForAllGraphModules.__init__r,   c                 G   s8   || j vrd S | j | }|D ]}|j| r|  S qd S rS   )r   r   )rW   r,   ry   
candidates	candidater(   r(   r)   &search_reusable_session_execution_info}  s   


zIOrtExecutionInfoForAllGraphModules.search_reusable_session_execution_infor_   c                 C   s.   || j vr|g| j |< d S | j | | d S rS   )r   rq   )rW   r,   r_   r(   r(   r)   cache_session_execution_info  s   
z?OrtExecutionInfoForAllGraphModules.cache_session_execution_info)r   N)
rc   rd   re   rU   r'   rk   GraphModuler   r   r   r(   r(   r(   r)   r   t  s    

r   r   T)frozenF)Z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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torch.onnx.ExportOptionsexport_optionszonnxruntime.SessionOptionsort_session_optionszonnx.ModelProtopre_ort_model_transforms)rc   rd   re   rf   r   r   r   r   __annotations__r   rm   r   r   r   r   r   r   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d u r,|j n| jj| _|jjjj| jj}d d d d d d}t||| _i | _t | _d| _d| _t|jdrct| _d S t| _d S )Nr   r   )getattrz_operator.getitemz_operator.mulz_operator.addz_operator.subFr   )rK   r   r!   r#   +torch.onnx._internal.fx.decomposition_tabler   _optionsonnx	_internalZ_exporter_legacyZResolvedExportOptionsr   ZExportOptions_resolved_onnx_exporter_optionsrk   r   Z'_create_onnx_supports_op_overload_tableonnx_registryrP   _supported_ops_partitioner_cacher   _all_ort_execution_info_assert_allclose_to_baselineexecution_countrw   r   r   r   r   )rW   r   rN   r'   rQ   rR   r(   r(   r)   rU     s<   


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   r   r|   r   r   r   r   rA   r   r9   rx   rq   )rW   r,   ry   Zinferred_epsZeps_from_argsZeps_from_graph_moduleZselected_epsr   r(   r(   r)   _select_epsG  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| }| jjrQd| _t|}dd }t||}nzt|j|i |}W n tyn   td| d| _ w |j| jjd}|| jj| }|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 }
t0|}t1|t,rt0|}nt0|f}t,dd |j-j.D }t,dd |j-j/D }t2||	||
||||d}| j3|| |  j4d7  _4t1|t5j6}|r)|fn|}t1|t,s3J t7dd |D s?J t8d | ||	|||
||| j"j||
}t9  | j:rt5j;j<j=|g|R ddi}|rn|fn|}t>||D ]\}}t5j?@|| qu|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)diagnostic_context)Zfx_graph_moduleonnxfunction_dispatcher)opset_versionr-   )r,   )Zpath_or_bytesZsess_options	providersc                 s       | ]}|j V  qd S rS   rF   r   inputr(   r(   r)   r         z2OrtBackend._ort_acclerated_call.<locals>.<genexpr>c                 s   r  rS   rF   r   r   r(   r(   r)   r     r  c                 s       | ]}|V  qd S rS   r(   r  r(   r(   r)   r         c                 s   r
  rS   r(   r	  r(   r(   r)   r     r  r   r0   c                 s   s$    | ]}t |tjtjtfV  qd S rS   )r   r'   r   r   r   )r   elemr(   r(   r)   r     s
    
Z$run_onnx_session_with_ortvaluevectorexecutorZaten)Ar   r%   r   r   r   r   r   r   r   r   r   r   r   r   ZMovePlaceholderToFrontr   r  r   dynamic_shapesr   r   r   Ztree_mapr   	propagate	Exceptionr^   rb   ZFxOnnxInterpreterZInsertTypePromotionr  Zto_model_protor   r  r   r   ZSerializeToStringr4   r5   r6   r?   ZInferenceSessionr   r   rx   r:   r  r   r   r   r   r   r   r'   r   allrG   rH   r   Z_primsr  executer   testingZassert_close)rW   r,   ry   kwargsr   r   r   Z!cached_execution_info_per_sessionZonnx_sessionr   r   r   r   r   r   Zprim_outputsZextracted_outputsr   Zfx_interpreterZexportedZ
onnx_modelZ	transformZonnx_model_bytesZexecution_info_per_sessionZis_single_tensor_outputr   Zonnx_outputsZbaseline_outputsZnormalized_baseline_ouptutsr   Zbaseline_outputr(   r(   r)   _ort_acclerated_callh  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)Zallows_single_node_partitionZcall_moduleZfused_)Z!torch.fx.passes.infra.partitionerr  r   r   Zpartition_and_fuser:   rp   r\   rB   r   r  Z_wrapped_call)	rW   r,   ry   r  Zpartitioned_prim_graph_moduleZprim_graph_moduleZpartitionerr[   Zfused_moduler(   r(   r)   compile1  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)Zfw_compilerZpartition_fnZdecompositions)	r   r   Zfunctorch.compiler  Ztorch._dynamo.backends.commonr  r  r   r   )rW   r,   ry   r  r  r(   r(   r)   __call__g  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   s   | j |j ks$| j|jks$| j|jks$| j|jks$| j|jks$| j|jkr&dS | jd us0|jd ur2dS | j|ju r:dS | jd urd|jd urd| jj|jjkoc| jj	|jj	koc| jj
|jj
u oc| jj|jju S dS )NFT)r   r   r   r   r   r   r   r   r  Zdiagnostic_optionsr   Zfake_context)r  r  r(   r(   r)   reusable  s,   	z<OrtBackend.get_cached_instance_for_options.<locals>.reusablec                 3   s     | ]}|j  r|V  qd S rS   )r   )r   r  r   r   r(   r)   r     s    z=OrtBackend.get_cached_instance_for_options.<locals>.<genexpr>NzNo more than z instances of z allowed. Please instantiate `z` explicitly to pass to `torch.compile`. See https://github.com/pytorch/pytorch/pull/107973#discussion_r1306144795 for discussion.r(   )r   r   nextr   r  r   r  rq   )r   backendr(   r!  r)   get_cached_instance_for_options|  s"   
#
z*OrtBackend.get_cached_instance_for_optionsc                   C   s   t j  d S rS   )r   r  clearr(   r(   r(   r)   clear_cached_instances  s   z!OrtBackend.clear_cached_instancesc                   C   s
   t tjS rS   )rx   r   r  r(   r(   r(   r)   get_cached_instances  s   
zOrtBackend.get_cached_instancesrS   )rc   rd   re   rf   r   r   rU   r'   rk   r   r   rx   r9   r   r   r   r  r  r  r  r   r   r  liststaticmethodr	   r$  r&  r'  r(   r(   r(   r)   r     s:   
 	X
! J6
E
r   )r   r   c                C   s   t || |S rS   )r   r$  )r,   ry   r   r(   r(   r)   r     s   r   rS   )[dataclassesr   loggingr4   collections.abcr   r   typingr   r   r   r   r   r	   Ztyping_extensionsr
   r'   Ztorch._CZ
torch._opsZtorch._prims.executorZtorch.fxZtorch._subclasses.fake_tensorr   Ztorch.fx._compatibilityr   Z torch.fx.passes.fake_tensor_propr   Z torch.fx.passes.operator_supportr   Ztorch.fx.passes.tools_commonr   Ztorch.utilsr   r   r   rK   r   rN   r!   r"   r#   r$   r   Ztorch.onnx._internal.fx.passesr   rm   r   __all__r   r*   rh   r9   r   bytesrk   r   r?   rA   rG   rH   rO   	getLoggerrc   r^   rP   rt   rx   r|   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   r   	dataclassr   r   r   r   r   r(   r(   r(   r)   <module>   sp  
 	1

		
2

"
(

 


	

Q

	

&L$!
?   i