U zh'@sddlZddlZddlmZddlmZmZmZddlZddl m Z ddlm Z ddl m ZddlmZdd lmZdd lmZed d Zed dZeddZeddZeddddZddZeee ddZededeeddZededdd Zed edeed!Zed"edGd#d$d$e Z!Gd%d&d&e Z"eej#j$d'd(d)Z%eej#j$d'd*d+Z&eej#j$d'd,d-Z'eej#j$d'd.d/Z(ej)Gd0d1d1Z*ej#j$d'd2d3Z+Gd4d5d5Z,dS)6N) import_module)AnyListOptional)#min_cut_rematerialization_partition)_guards)config) ts_compile) aot_autograd)register_debug_backendcCs|jSN)forwardgmfake_tensor_inputsrR/var/www/html/venv/lib/python3.8/site-packages/torch/_dynamo/backends/debugging.pyeagersrcsfdd}|S)Nc sBz |WStk r<}ztjjd|W5d}~XYnXdS)Nz7Unexpected exception when running generated GraphModule) ExceptiontorchZ_dynamoexcZTorchDynamoException)argserrrinners zeager_noexcept..innerrrrrrrreager_noexcepts rcs4ddlm}fdd}||dd|}||S)Nr)make_fxcstjj|Sr rfx Interpreterrunrrrr runnable_gm.sz'pre_dispatch_eager..runnable_gmT)Z pre_dispatch)Z"torch.fx.experimental.proxy_tensorrZprint_readable)rrrr$Zpre_dispatch_gmrrrpre_dispatch_eager*s   r%csddlmfdd}|S)Nr)SchemaCheckModec s0 tjj|W5QRSQRXdSr rr#r&rrrr>szeager_debug..inner)Z#torch._subclasses.schema_check_moder&rrr'r eager_debug7s r(ts)namecCs tj|Sr )rZjitscriptrrrr torchscriptEsr,csfdd}d|_|S)Ncstj|Sr )rr r!Z boxed_runr#fx_grrr"Lszboxed_nop..runT)Z _boxed_call)r.example_inputsr"rr-r boxed_nopKs r0T) fw_compiler partition_fnkeep_inference_input_mutations aot_eager)r*Z compiler_fn)r1r3aot_eager_default_partitionerc CsHtjdd2tttddtjtddd||W5QRSQRXdS)NT)Zunlift_effect_tokenscSs tdS)Nztorch._inductor.compile_fx)rZselect_decomp_tablerrrrosz,aot_eager_decomp_partition..Zinductor)compiler)r1Z bw_compilerZdecompositionsr2)functorch_configpatchr r0 functoolspartialrrrrraot_eager_decomp_partitionhs r<)r1aot_tsc@s eZdZdS)ReluCompileErrorN__name__ __module__ __qualname__rrrrr>sr>c@s eZdZdS)TestingOnlyCompileErrorNr?rrrrrCsrCrcCs"|jjD]}|jtjkrtq|Sr )graphnodestargetrrelur>rr/noderrrrelu_compile_error_TESTING_ONLYs  rJcCs4|jjD]}|jtjkrtj|_d|_q||S)N)FZReluRuntimeError)rDrErFrrGZ_assertr recompilerHrrrrelu_runtime_error_TESTING_ONLYs   rLcCs>|jjD](}|jtjkrtj|_|jddf|_q||S)Nrr )rDrErFrrGaddrrKrHrrr relu_accuracy_error_TESTING_ONLYs   rNcCs8|jjD]}|jdkrq q|S|D]}|js$tq$|S)N call_function)rDrEopZis_leafrC)rr/rItrrr#non_leaf_compile_error_TESTING_ONLYs  rRc@seZdZUdZeejjed<e ed<e ed<ee ed<e ed<dZ e eejj ed<dZe eejed <dZe eed <d d ZdS) ExplainOutputzu This is the output of :func:`torch._dynamo.explain()` There is no reason to create this class directly. graphs graph_countgraph_break_count break_reasonsop_countN ops_per_graph out_guards compile_timesc Cs^d|jd}|d|jd7}|d|jd7}|d7}t|jD]R\}}|d|dd7}|d |jd7}|d 7}|jD]}|d |d7}qqD|jdk r|d 7}t|jD]6\}}|d |dd7}|D]}|d|d7}qq|jdk r<|d7}t|jD]0\}}|d|dd7}|dt |7}q |j dk rZ|d|j d7}|S)Nz Graph Count:  zGraph Break Count: z Op Count: zBreak Reasons: z Break Reason r z: z Reason: z User Stack: z zOps per Graph: z Ops z z Out Guards: z Guard zCompile Times: ) rUrVrX enumeraterWreasonZ user_stackrYrZstrr[) selfoutputidxZ break_reason frame_summaryopsrPiguardrrr__str__s0    zExplainOutput.__str__)r@rArB__doc__rrr GraphModule__annotations__intrrYrNoderZrZGuardr[r_rgrrrrrSs rScCsT||dd|jjD}|t|7}|||jjrF||j|||||fS)a This function is a utility which processes a torch.fx.GraphModule and accumulates information about its ops, graph breaks, and other details. It is intended to be used by the ExplainWithBackend class and `torch._dynamo.explain()` to provide details from Dynamo's graph capture. Parameters: gm (torch.fx.GraphModule): The GraphModule to be processed. graphs (list): A list that accumulates all the GraphModules processed. op_count (int): The total count of operations in all GraphModules processed so far. ops_per_graph (list): A list that accumulates the operations of each GraphModule. break_reasons (list): A list that accumulates the reasons for breaks in each GraphModule. Returns: tuple: A tuple containing the processed GraphModule, the updated lists of graphs, operations per graph, and break reasons, and the updated operation count. cSsg|]}|jdkr|jqS)rO)rPrF).0rIrrr s z)_explain_graph_detail..)appendrDrElenZcompile_subgraph_reasonZ graph_break)rrTrXrYrWrdrrr_explain_graph_details    rqc@s8eZdZdZddZejjdddZe ddd Z d S) ExplainWithBackenda This class is intended to be used as a backend for `torch.compile`. It is composable with other backends. When used in this way, it accumulates information about graph breaks, ops, and other info and provides a string representation summarizing this information. Attributes: backend (str): The name of the backend to use for optimization. graphs (list): A list of the graphs captured by TorchDynamo. op_count (int): The total number of operations in all optimized graphs. break_reasons (list): A list of graph break reasons with stack traces. Example Usage: def fn(x): x = torch.sigmoid(x) return x torch._dynamo.reset() eb = ExplainWithBackend("inductor") optimized_fn = torch.compile(fn, backend=eb) result = optimized_fn(torch.randn(5)) print(eb.output()) cCs,ddlm}|||_g|_d|_g|_dS)Nr )lookup_backendr)registryrsbackendrTrXrW)r`rursrrr__init__#s   zExplainWithBackend.__init__rcCs2t||j|jg|j\}|_|_}|_|||Sr )rqrTrXrWru)r`rr/_rrr__call__+szExplainWithBackend.__call__)returncCs(t|j}t|j||d|j|j}|S)Nr )rprTrSrWrX)r`rUrarrrra1s zExplainWithBackend.outputN) r@rArBrhrvrr rirxrSrarrrrrr srr)- dataclassesr: importlibrtypingrrrrZfunctorch.compilerrZtorch._functorchrr8Ztorch._functorch.compilersr commonr rtr Zregister_backendrrr%r(r,r0r4r5r<r=rr>rCr rirJrLrNrR dataclassrSrqrrrrrrsr                   1