U Kc%@sUddlZddlmZddlmZddlmZddlmZm Z m Z m Z m Z m Z ddlZddlZddlZddlmZddd gZiZe ejjefed<ejd d d Zdd dedddZe e ejjejjfe ejjefddd ZddlZddl ZdS)N) defaultdict)wraps)chain)CallableDict NamedTupleSequenceTupleUnion)tree_mapdecomposition_tableregister_decompositionget_decompositionsZatenZIMPLMetaF) disable_metacsttdfdd }|S)aS A decorator to register a function as a decomposition to the Python decomposition table. Use it like this:: @register_decomposition(torch.ops.aten.clamp_min) def clamp_min(x): return torch.clamp(self, min=min) If you are writing a new decomposition, consider contributing it directly to PyTorch in torch._decomp.decompositions. This API is experimental; we are almost certainly going to extend the API when we make decompositions eligible for use in transforms (e.g., autograd) and not just backend tracing, where we then need to know if a decomposition can be used to simulate a transform. By default, if the decomposition is for an operator that doesn't have a Meta implementation, we will register it to the dispatcher. Use `disable_meta` to disable this behavior. )freturncst}jd}|rt|ddtkr|jjtfdd}ddt |j D}t dd|j D|}tj||jd |_d d j D|_|D]}|j|j|j<q|dkrtfd d }t|S)Nout __origin__csVtfddD}|ddktfdd|Ds8szWregister_decomposition..decomposition_decorator.._fn..rc3s|]}|dkkVqdSrrr)is_nonerrr;sr)tupleallAssertionError)argsrZ out_kwargs)r out_names)rrr_fn6s zDregister_decomposition..decomposition_decorator.._fncSs&g|]\}}tj|tjjd|dqS)N)kinddefault annotation)inspect Parameter KEYWORD_ONLY)rrtrrr >szKregister_decomposition..decomposition_decorator..css|]\}}|dkr|VqdS)rNrrkvrrrrIszJregister_decomposition..decomposition_decorator..) parametersreturn_annotationcSsi|]\}}|dkr||qS)rrr-rrr OszKregister_decomposition..decomposition_decorator..csg}t|tjjr||n0t|tjjs0t|D]}|t||q8|D]}|krlt d||<| tj j j |jj}|jjr|d|jj7}sRtj |rRtj |dsRtdd|jjDrt dt|dt|qRdS)Nzduplicate registrations for .rcss"|]}|jdk o|jj VqdSr)Z alias_infoZis_write)rarrrr|szcregister_decomposition..decomposition_decorator..add_op_to_table..zB Attempting to register a python meta kernel for a view operator: a0. We shouldn't do this, because the output will report as not having aliased storages. All view ops have meta kernels in C++ today, so we should use those instead. If you're registering an operator through the `@register_decomposition` decorator, Please set `disable_meta=True`. ) isinstancetorch_ops OpOverloadappendOpOverloadPacketr! overloadsgetattr RuntimeErrorZpy_implZ_CZ DispatchKeyrZ_schemanameZ overload_nameZ_dispatch_has_kernelZ._dispatch_has_computed_kernel_for_dispatch_keyany argumentsstrmeta_libimpl)aten_opr;ol op_overloadr>)rfnregistryrradd_op_to_table[s@     zPregister_decomposition..decomposition_decorator..add_op_to_table)r( signature__annotations__getr<rr1_fieldsrzip__args__rr0items Signature __signature__r'r>r r )rsigZout_annotationr$Z out_paramsparamsrrIrDrrH)rrGr#rdecomposition_decorator,s:    2 z7register_decomposition..decomposition_decorator)r)rDrHrrVrrUrr sd)aten_opsrcCstt}tD]}||j|q i}|D]X}t|tjjr`||kr`||D]}t|||<qLq*t|tjj r*|tkr*t|||<q*|S)a, Retrieve a dictionary of decompositions corresponding to the list of operator overloads and overload packets passed as input. Overload packets will include all decomposed overloads in the packet. If there is no decomposition for a requested operator, it is silently ignored. This API is experimental; we are almost certainly going to give an alternate, more recommended formulation, where a user provides the set of operators they know how to implement, and we provide decompositions for everything not in this set. ) rlistr Zoverloadpacketr9r5r6r7r:r8)rWZpackets_to_overloadsZopoZdecompositionsoprFrrrrs )N)!r( collectionsr functoolsr itertoolsrtypingrrrrr r r6Z torch._opsZ torch.libraryZtorch.utils._pytreer __all__r r7r8rKZlibraryLibraryrBboolr r:rZtorch._decomp.decompositionsZ torch._refsrrrrs$      ~