i SrSSKJrJrJrJrJr SSKrSSKJ s J r SSK J r JrJr SSKJrJrJrJrJrJrJrJr SSKJr S\R6S\\R8R:\R644S jrSS \S \\R6\\R644S \\\R6S \S\R@RB4Sjjr"S\S \R64S\S \\R64S\ RFS\\\R6\44Sjr$S\\R6S 4S\S \S \\R6\\R64444Sjr%g)a  This module contains tools for rewriting a dynamic PyTorch program such that the dynamic part (e.g. control flow) can be properly captured by DispatchTracer. The core idea is annotating all branches in the graph with unique keys, and using a dictionary of supplemental inputs as arguments to these local branches so that every path gets a canonical input during tracing. For example, consider the following usage of Python if statement: .. code-block:: python if pred: ... ret = a else: ... ret = b To rewrite the code to be tracable, users may use tracing_key decorator and cond operator: .. code-block:: python @control_flow.tracing_context(inputs) def branch_true(args): ... return a @control_flow.tracing_context(inputs) def branch_false(args): ... return b ret = control_flow.cond(pred, branch_true, branch_false, args) and we can use the usual exir.capture() function. .. code-block:: python exir.capture(module, args) )CallableListOptionalTupleUnionN) ExportErrorExportErrorTypeinternal_assert)DispatchTracerflattened_dispatch_trace PythonTensor tree_returnunwrap_functional unwrap_proxy using_tracerValue)update_with_proxyxreturnc*[R"5nUc UR$[U5n[ U[ 5(d&[ [RS[U535e[R"[UR5[RS9n[UR5H up4XBU'M [RR R"R%UR&5n[ [U5U5$)zL A helper function for capturing the shape as a tensor from a tensor value. zEexir custom shape function only takes EXIR dispatch tensor, but got: )dtype)r getshaper isinstancer rr INVALID_INPUT_TYPEtypetorchemptylenint64 enumerateopsaten_shape_as_tensordefaultproxy)rtracertmpisr&s [/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/executorch/exir/control_flow.pyrrGs    !F ~ww!A a & &  . .STXYZT[S\ ]   ++c!''l%++ 6C!''"A# IINN + + 3 3AGG while_loop..s E4Dqz!U\\**4D')zRcontrol_flow.while_loop() expects all inputs values to be tensors, actual inputs: Nc3V# UHn[U[R5v M! g7frIrJ)rLos r+rMrNs F4Eqz!U\\**4ErOzUcontrol_flow.while_loop() expects all returned values to be tensors, actual outputs: T)r/ call_function)pytree tree_flattenallrr rrINVALID_OUTPUT_TYPEr rrCr>r create_proxy while_looprr) rDrErFflattened_inputsrAvalflattened_outputsr'gm_condgm_bodyvproxiesr&s r+rXrXsT!--h7 E4DE E E  . .`ai`j k  d sm3-Csmm  "..s3 F4EF F F  / /cdgch i     !F ~ gT:Gg&G.>?.>\!_.>?@G    7#  E s#4 55;  (@sEE Einputsc ^S[S[[R4S[S[[R[[R444U4SjjnU$)z A decorator function to annotate code path that we conditionally run during tracing. We need to annotate these paths for now because during exir.capture(), the tracer does not know what's the proper local inputs to be passed to the untaken path. f.rc>^S[RS[[RS[[R4U4SjjnTUlU$)Nr=kwargsrcP>U(a[[RS5eT"U6$)NzBkwargs are not supported for @tracing_context decorated functions.)rr NOT_SUPPORTED)r=rdrbs r+wrapper3tracing_context..decorator..wrappers,!#11X d8Or,)rrKrr1)rbrgr`s` r+ decorator"tracing_context..decoratorsE << +0+>  5<<  &,"r,)rrrrKr)r`ris` r+tracing_contextrksT Cu||,, - #uU\\5+>>?? @" r,)NF)&__doc__typingrrrrrrtorch.utils._pytreeutils_pytreerSexecutorch.exir.errorrr r executorch.exir.tracerr r r rrrrrexecutorch.exir.wraprrK_CSizerboolfx GraphModulerCPyTreerXrkr,r+r{s~*X:9 $$OO   37U\\7eEHHMM5<<$?@7459eELL% *==>>?d5<<01 XX D-6 c5<<' (-6 c5.. /-6mm-6 5  %& -6` %,,# $ c8Cu||U5<<5H'H!IIJJKr,