ó [T–h¬ãó’•%SSKrSSKJr SSKrSSKJr /r\\\S'"SS5r "SS5r S \S \4S jr S \S \ S \4S jr g)éN)ÚCallable)ÚTensorÚ__all__có"•\rSrSrS\4SjrSrg)Ú _CodeParseré Ú code_stringcó<•SnSnSnSnSnSnSnUS-U-U-U-U-U-U-U-U-U-U-U-n [R"X‘[R5n U c[S U35eU S UlU S UlU S UlU S UlU SUlg)Nz\s*z\s+z(?P\<.+\>)z(?P\w+)z(?P\w+)z(?P\(.+\))z(?P\{.+\})Útemplatez0Couldn't parse code, please check correctness: Útemplate_paramsÚ return_typeÚ function_nameÚfunction_paramsÚ function_body) ÚreÚmatchÚDOTALLÚ Exceptionr r rrr) Úselfr Ú optional_wsÚ required_wsr r rrrÚpatternÚresults ÚL/var/www/auris/envauris/lib/python3.13/site-packages/torch/cuda/jiterator.pyÚ__init__Ú_CodeParser.__init__ s€Øˆ ؈ Ø8ˆØ-ˆ Ø1ˆ Ø8ˆØ4ˆ ð Øñ àñ ðñ ðñ  ð ñ  ð ñ  ðñ ðñ ðñ ðñ ðñ ðñ ð ô —’Ø ¤"§)¡)ó ˆð ‰>ÜØCÀKÀ=ÐQóð ð &Ð&7Ñ8ˆÔØ! -Ñ0ˆÔØ# OÑ4ˆÔØ%Ð&7Ñ8ˆÔØ# OÑ4ˆÕó)rrrr r N)Ú__name__Ú __module__Ú __qualname__Ú__firstlineno__ÚstrrÚ__static_attributes__©rrrr s†ð&5 C÷&5rrcó8•\rSrSrS\S\S\4SjrS\4Sjr Sr g ) Ú_JittedFunctioné6r Ú return_by_refÚ num_outputsc óÚ•XlU(d US:XdS5eX lX0l[U5nURUlX@l[RR5Ul g)Néz.Return by value only works for single output. ) r r(r)rrÚ kernel_nameÚ kwargs_dictÚtorchÚcudaÚ is_availableÚis_cuda_available)rr r(r)ÚkwargsÚ parsed_codes rrÚ_JittedFunction.__init__7se€ð'Ôö ˜[¨AÓ-ð <à ;ó <Ø -à*ÔØ&Ôä! +Ó.ˆ Ø&×4Ñ4ˆÔà!ÔÜ!&§¡×!8Ñ!8Ó!:ˆÕrÚtensorscóž•UR(dS5e[U5S::dS5eURR5nUR 5H&upEX@R;aXSU'M[ US35e [ RRURURURURUU5$)NzFJiterator is only supported on CUDA and ROCm GPUs, none are available.éz.jiterator only supports up to 8 tensor inputs.z' is not declared in function definition) r1Úlenr-ÚcopyÚitemsÚKeyErrorr.Ú_CÚ)_cuda_jiterator_compile_and_launch_kernelr r,r(r))rr5r2Úexpanded_kwargsÚkeyÚvalues rÚ__call__Ú_JittedFunction.__call__HsÉ€ð × "× "ð Tà Só TØ "ô7‹|˜qÓ ÐRÐ"RÓRÐ à×*Ñ*×/Ñ/Ó1ˆØ Ÿ,™,ž.‰JˆCØ×&Ñ&Ó&Ø', Ó$ä # Ð&MÐNÓOÐOñ )ô x‰x×AÑAØ × Ñ Ø × Ñ Ø × Ñ Ø × Ñ Ø Ø ó  ð r)r r1r,r-r)r(N) rrr r!r"ÚboolÚintrrrAr#r$rrr&r&6s+†ð;Øð;Ø/3ð;ØBEô;ð"  ÷ rr&r Úreturnc ó •[U4SSS.UD6$)aY Create a jiterator-generated cuda kernel for an elementwise op. The code string has to be a valid CUDA function that describes the computation for a single element. The code string has to follow the c++ template pattern, as shown in the example below. This function will be inlined into elementwise kernel template, and compiled on the fly. Compiled kernel will be cached in memory, as well as local temp dir. Jiterator-generated kernels accepts noncontiguous tensors, and supports broadcasting and type promotion. Args: code_string (str): CUDA code string to be compiled by jiterator. The entry functor must return by value. kwargs (Dict, optional): Keyword arguments for generated function Example:: code_string = "template T my_kernel(T x, T y, T alpha) { return -x + alpha * y; }" jitted_fn = create_jit_fn(code_string, alpha=1.0) a = torch.rand(3, device='cuda') b = torch.rand(3, device='cuda') # invoke jitted function like a regular python function result = jitted_fn(a, b, alpha=3.14) code_string also allows multiple function definitions, and the last function will be treated as the entry function. Example:: code_string = "template T util_fn(T x, T y) { return ::sin(x) + ::cos(y); }" code_string += "template T my_kernel(T x, T y, T val) { return ::min(val, util_fn(x, y)); }" jitted_fn = create_jit_fn(code_string, val=0.0) a = torch.rand(3, device='cuda') b = torch.rand(3, device='cuda') # invoke jitted function like a regular python function result = jitted_fn(a, b) # using default val=0.0 Jiterator can be used together with python registration to override an operator's cuda kernel. Following example is overriding gelu's cuda kernel with relu. Example:: code_string = "template T my_gelu(T a) { return a > 0 ? a : 0; }" my_gelu = create_jit_fn(code_string) my_lib = torch.library.Library("aten", "IMPL") my_lib.impl('aten::gelu', my_gelu, "CUDA") # torch.nn.GELU and torch.nn.function.gelu are now overridden a = torch.rand(3, device='cuda') torch.allclose(torch.nn.functional.gelu(a), torch.nn.functional.relu(a)) .. warning:: This API is in beta and may change in future releases. .. warning:: This API only supports up to 8 inputs and 1 output .. warning:: All input tensors must live in CUDA device Fr+©r(r)©r&)r r2s rÚ_create_jit_fnrIbs€ôt ˜;Ð U°eÈÑ UÈfÑ UÐUrr)c ó •[U4SUS.UD6$)a\ Create a jiterator-generated cuda kernel for an elementwise op that supports returning one or more outputs. Args: code_string (str): CUDA code string to be compiled by jiterator. The entry functor must return value by reference. num_outputs(int): number of outputs return by the kernel kwargs (Dict, optional): Keyword arguments for generated function Example:: code_string = "template void my_kernel(T x, T y, T alpha, T& out) { out = -x + alpha * y; }" jitted_fn = create_jit_fn(code_string, alpha=1.0) a = torch.rand(3, device='cuda') b = torch.rand(3, device='cuda') # invoke jitted function like a regular python function result = jitted_fn(a, b, alpha=3.14) .. warning:: This API is in beta and may change in future releases. .. warning:: This API only supports up to 8 inputs and 8 outputs TrGrH)r r)r2s rÚ_create_multi_output_jit_fnrKŸs&€ô4 Øð Ø#'°[ñ ØDJñ ðr)rÚtypingrr.rrÚlistr"Ú__annotations__rr&rIrDrKr$rrÚrOskðä Ýã Ýð€ˆˆc‰Ó÷'5ñ'5÷T) ñ) ðX:V ð:V°(ô:VðzØðØ#&ðà õr