ȅi %SrSSKrSSKrSSKJrJr SSKJr SSKJ r J r J r J r SSK r SSK Jr \R"\5r"SS\ 5r\\R&\\ R*/\4r0r\\\ \4\S '0r\\\4\S 'SS \ \S \ \S \\S\S\ 44Sjjr\R:"\SS9r\R:"\SS9rS \ \\4S\4Sjr SS\\4Sjjr!\RDSSj5r#\RDSSj5r$g)a This module implements TorchDynamo's backend registry system for managing compiler backends. The registry provides a centralized way to register, discover and manage different compiler backends that can be used with torch.compile(). It handles: - Backend registration and discovery through decorators and entry points - Lazy loading of backend implementations - Lookup and validation of backend names - Categorization of backends using tags (debug, experimental, etc.) Key components: - CompilerFn: Type for backend compiler functions that transform FX graphs - _BACKENDS: Registry mapping backend names to entry points - _COMPILER_FNS: Registry mapping backend names to loaded compiler functions Example usage: @register_backend def my_compiler(fx_graph, example_inputs): # Transform FX graph into optimized implementation return compiled_fn # Use registered backend torch.compile(model, backend="my_compiler") The registry also supports discovering backends through setuptools entry points in the "torch_dynamo_backends" group. Example: ``` setup.py --- from setuptools import setup setup( name='my_torch_backend', version='0.1', packages=['my_torch_backend'], entry_points={ 'torch_dynamo_backends': [ # name = path to entry point of backend implementation 'my_compiler = my_torch_backend.compiler:my_compiler_function', ], }, ) ``` ``` my_torch_backend/compiler.py --- def my_compiler_function(fx_graph, example_inputs): # Transform FX graph into optimized implementation return compiled_fn ``` Using `my_compiler` backend: ``` import torch model = ... # Your PyTorch model optimized_model = torch.compile(model, backend="my_compiler") ``` N)CallableSequence) EntryPoint)AnyOptionalProtocolUnion)fxcX\rSrSrS\R S\\R S44SjrSrg) CompiledFnJargsreturn.cg)N)selfrs Y/home/james-whalen/.local/lib/python3.13/site-packages/torch/_dynamo/backends/registry.py__call__CompiledFn.__call__KsrN) __name__ __module__ __qualname____firstlineno__torchTensortupler__static_attributes__rrrr r Js!LellLuU\\35F/GLrr _BACKENDS _COMPILER_FNS compiler_fnnametagsr.c Uc[R"[XS9$[U5(deU=(d URnU[ ;d SU35eU[ ;a S[ U'U[ U'[U5UlU$)a Decorator to add a given compiler to the registry to allow calling `torch.compile` with string shorthand. Note: for projects not imported by default, it might be easier to pass a function directly as a backend and not use a string. Args: compiler_fn: Callable taking a FX graph and fake tensor inputs name: Optional name, defaults to `compiler_fn.__name__` tags: Optional set of string tags to categorize backend with N)r"r#zduplicate name: ) functoolspartialregister_backendcallablerr rr_tags)r!r"r#s rr'r'Ts   !1HH K   ';''D } $?(8&?? $)# $%M$d K r)debug)r#) experimentalc[U[5(aaU[;a [5 U[;a SSKJn U"US9eU[ ;a$[UnUb[UR5US9 [ UnU$)z#Expand backend strings to functions)InvalidBackend)r")r!r") isinstancestrr _lazy_importexcr.r r'load)r!r. entry_points rlookup_backendr5wsm+s## i ' N i ' , k2 2 m +#K0K& [-=-=-?kR#K0 rc[5 [U=(d S5n[Vs/sH8nU[;d)UR [UR 5(aM6UPM: nn[ U5$s snf)zU Return valid strings that can be passed to: torch.compile(..., backend="name") r)r1setrr intersectionr)sorted) exclude_tagsexclude_tags_setr"backendss r list_backendsr=sn N<-2.D } $,,]4-@-F-FG   (  s 5A3 A3cVSSKJn SSKJn U"U5 SSKJn Uce[ 5 g)Nr-)r<)import_submodule)dynamo_minifier_backend)r<utilsr?repro.after_dynamor@_discover_entrypoint_backends)r<r?r@s rr1r1s%(X< " .. .!#rcSSKJn SnU"US9nURVs0sHo3X#_M nnUHnXE[U'M gs snf)Nr) entry_pointstorch_dynamo_backends)group)importlib.metadatarFnamesr)rF group_nameepsr"eps_dict backend_names rrDrDsN0(J Z (C,/II6IDciIH6 "*"8 ,!7sA)NNr))r*r+)rN)%__doc__r%loggingcollections.abcrrrIrtypingrrrr rr getLoggerrlogr GraphModulelistr CompilerFnrdictr0__annotations__r r'r&register_debug_backendregister_experimental_backendr5r=cacher1rDrrrr]sg:x.)11 !MMr~~tELL'9:JF G -/ 4Xj)) */') tCO$))-*% 3- 3-c3h :#**+;*M ) 1 1,! c:o 6:&T#Y$  $ $  9 9r