o Xiu@sddlmZddlZddlZddlZddlmZmZmZm Z m Z ddl m Z ddl Z ddl mZddlmZmZmZmZe dZe dZ d(d)ddZ  d*d+ddZd,ddZd-d d!Zd.d&d'ZdS)/) annotationsN)AnyCallableOptionalSequenceTypeVar) ParamSpec)ir) ast_utils converter irbuildervalues_R_Pfast.FunctionDefopset values.Opset global_namesdict[str, Any]sourcestr default_opsetOptional[values.Opset]returnirbuilder.IRFunctioncCstj||||d}||S)zACheck that a function falls into the ONNXScript subset of Python.)rrrr)r Convertertranslate_function_def)rrrrrconvertrM/home/ubuntu/.local/lib/python3.10/site-packages/onnxscript/_internal/main.py script_checks  r!kwargsr=Callable[[Callable[_P, _R]], onnxscript.OnnxFunction[_P, _R]]c s:ptddttjstdd fdd }|S) a8Main decorator. Declares a function as an onnx function. Args: opset: Opset the function belongs to (see :ref:`l-api-opsets`). default_opset: Opset to use for operators not in the function's opset. kwargs: Additional keyword arguments. Returns: an instance of :class:`onnxscript.values.OnnxFunction` Example: :: @script() def log2(x): one = op.Constant(value=make_tensor('one', TensorProto.FLOAT, [1], [1])) return op.Div(op.Log(x), op.CastLike(op.Log(cst), x)) Or: :: from onnxscript.onnx_opset import opset16 @script(opset16) def log2(x): one = op.Constant(value=make_tensor('one', TensorProto.FLOAT, [1], [1])) return op.Div(op.Log(x), op.CastLike(op.Log(cst), x)) thiszLScript parameter must be an opset. Did you use @script instead of @script()?rCallable[_P, _R]ronnxscript.OnnxFunction[_P, _R]csnt|s tdt|\}}t|}t|}|j}| |j t |||d}t |||S)Nz=The ONNXScript decorator should be applied to functions only.)r)inspect isfunction TypeErrorr get_src_and_ast getmodulegetclosurevars__dict__copyupdate nonlocalsr! onnxscript OnnxFunction)rsrcf_astmoduleclosureenvresultrr"rrr transformQs     zscript..transformN)rr&rr')r Opset isinstancer*)rrr"r;rr:r script)s" r>(Callable[[Callable], values.OnnxClosure]cs:tdtd}|jd}|jjd fdd }|S) aQA parametric decorator used to annotate nested-functions that are used as graph-attributes. Returns: A decorator that returns its input function, but attaches a graph_proto attribute representing the input function. The translation is not done at this time, but previously when the outer-level function was translated to an OnnxFunction. The decorator just looks up and retrieves the GraphProto representation previously generated. Example: :: @script() def cumulative_sum(X: INT64['N']): # Translation of cumulative_sum by @script will also translate Sum # into a GraphProto, which will be stored in the OnnxFunction generated # for cumulative_sum. At run-time (in eager-mode), the @graph decorator # retrieves the pre-computed GraphProto and attaches it to the Sum function. @graph() def Sum(sum_in, next): sum_out = sum_in + next scan_out = op.Identity(sum_out) return sum_out, scan_out zero = op.Constant(value_int=0) # The call to higher-order operator Scan below uses the above function # Sum as a graph-attribute. all_sum, result = op.Scan (zero, X, body=Sum, num_scan_inputs=1) return result r%selfrrrvalues.OnnxClosurecst|j|SN)r OnnxClosure__name__rfunction_framenested_functionsrr r;szgraph..transformN)rrrrB)sys _getframef_locals function_irrI) wrapper_frame onnx_functionr;rrGr graphes *  rPboolcCs t|tjS)zAReturn True if f is a function converted by onnxscript decorator.)r=r2r3rFrrr is_converted_funs rR functionsSequence[values.OnnxFunction]filenameNonecCs>tjtjgggddiddd|Dddd}t||dS) N)inputsoutputsnodes opset_importscSsg|] }tj|qSr)r serdedeserialize_functionto_function_proto).0rrrr sz#export_onnx_lib.. p2o)rS ir_version producer_name)r ModelGraphsave)rSrUmodelrrr export_onnx_libs  rjrC) rrrrrrrrrrrr)NN)rrrrr"rrr#)rr?)rrrrQ)rSrTrUrrrV) __future__rastr(rJtypingrrrrrtyping_extensionsrr2r onnxscript._internalr r r r rrr!r>rPrRrjrrrr s&     < 5