o I h @sUddlZddlZddlmZmZddlmZddlZddlm Z m Z GdddZ de de d efd d Z d d Zeaeed<ejddZGdddZdddZdS)N)CallableOptional) deprecated)KernelRegistrationHandlec@s4eZdZdZdefddZdededefdd Zd S) FakeImplHolderz0A holder where one can register an fake impl to.qualnamecCs||_d|_d|_dSN)rkernellib)selfrr L/var/www/vscode/kcb/lib/python3.10/site-packages/torch/_library/fake_impl.py__init__s zFakeImplHolder.__init__funcsourcereturncsjdurtdjdjjdtjjdr$tdjdtjjdr5tdjdt||_jdurPj d d }tj |d _t j}j j|dfd d }t|S)z}Register an fake impl. Returns a RegistrationHandle that one can use to de-register this fake impl. Nz!register_fake(...): the operator z( already has an fake impl registered at .Metaz already has an DispatchKey::Meta implementation via a pre-existing torch.library or TORCH_LIBRARY registration. Please either remove that registration or don't call register_fake.CompositeImplicitAutograda% already has an implementation for this device type via a pre-existing registration to DispatchKey::CompositeImplicitAutograd.CompositeImplicitAutograd operators do not need an fake impl; instead, the operator will decompose into its constituents and those can have fake impls defined on them.z::rFRAGMENTcs jr jd_d_dSr )r _destroyr r r r rderegister_fake_classAs  z6FakeImplHolder.register..deregister_fake_class)r RuntimeErrorrrtorch_C%_dispatch_has_kernel_for_dispatch_keyrr splitlibraryLibraryconstruct_meta_kernelimplr)r rrns meta_kernelrr rrregisters0       zFakeImplHolder.registerN) __name__ __module__ __qualname____doc__strrrrr%r r r rr srrfake_impl_holderrcs.jdusJtjjfdd}|S)Ncs`jdusJjjfdd}t|j|i|WdS1s)wYdS)Ncstdd)Nz (a): You're trying to run this operator with meta Tensors (as opposed to FakeTensors), but this operator may return an output Tensor with data-dependent shape. Meta Tensors don't support operators with outputs that have data-dependent shapes but FakeTensors do. If your operator does not return an output with data-dependent shape, make sure the FakeTensor and/or meta kernel does not call torch.library.get_ctx(). Otherwise, please use FakeTensors.)rr )rrr r error_on_ctxRsz@construct_meta_kernel..meta_kernel..error_on_ctx)r rset_ctx_getter)argskwargsr,r+r)rrr$Ms  $z*construct_meta_kernel..meta_kernel)r functoolswrapsr)rr+r$r r0rr!Js r!cCsdSr r r r r rget_nonedsr3global_ctx_getterccs"t}z |adVW|adS|awr )r4) ctx_getterprevr r rr-ks r-c@sTeZdZdZddZededdddd ejfd d Z d ddd ejfd dZ dS) FakeImplCtxzO Context object for writing fake implementations for custom operators. cCs||_|j|_||_dSr ) _fake_mode shape_env _shape_env_op)r r8r;r r rr{s zFakeImplCtx.__init__zM`create_unbacked_symint` is deprecated, please use `new_dynamic_size` instead)categoryNminmaxrcCs|j||dSNr>)new_dynamic_sizer r?r@r r rcreate_unbacked_symintsz"FakeImplCtx.create_unbacked_symintrcCsv|jdus |jjstjj|jt|tjst|tjr(t d|d|d|dkr4t d|dt |j||S)a Constructs a new symint (symbolic int) representing a data-dependent value. This is useful for writing the fake implementation (which is necessary for torch.compile) for a CustomOp where an output Tensor has a size that depends on the data of the input Tensors. Args: min (int): A statically known inclusive lower bound for this symint. Default: 0 max (Optional[int]): A statically known inclusive upper bound for this symint. Default: None .. warning: It is important that the ``min`` and ``max`` (if not None) values are set correctly, otherwise, there will be undefined behavior under torch.compile. The default value of ``min`` is 2 due to torch.compile specializing on 0/1 sizes. You must also verify that your implementation on concrete Tensors (e.g. CPU/CUDA) only returns Tensors where the size that corresponds to the symint also has respects these constraint. The easiest way to do this is to add an assertion in the CPU/CUDA/etc implementation that the size follows these bounds. Example:: >>> # An operator with data-dependent output shape >>> lib = torch.library.Library("mymodule", "FRAGMENT") >>> lib.define("mymodule::custom_nonzero(Tensor x) -> Tensor") >>> >>> @torch.library.register_fake("mymodule::custom_nonzero") >>> def _(x): >>> # Number of nonzero-elements is data-dependent. >>> # Since we cannot peek at the data in an fake impl, >>> # we use the ctx object to construct a new symint that >>> # represents the data-dependent size. >>> ctx = torch.library.get_ctx() >>> nnz = ctx.new_dynamic_size() >>> shape = [nnz, x.dim()] >>> result = x.new_empty(shape, dtype=torch.int64) >>> return result >>> >>> @torch.library.impl(lib, "custom_nonzero", "CPU") >>> def _(x): >>> x_np = x.numpy() >>> res = np.stack(np.nonzero(x_np), axis=1) >>> return torch.tensor(res, device=x.device) Nzctx.new_dynamic_size(min=z, max=zZ): expected min and max to be statically known ints but got SymInt. This is not supported.rzc, ...): expected min to be greater than or equal to 0: this API can only create non-negative sizes.) r:allow_dynamic_output_shape_opsr _subclasses fake_tensorDynamicOutputShapeExceptionr; isinstanceSymInt ValueError allocate_sizerCr r rrBs 3 zFakeImplCtx.new_dynamic_size) r&r'r(r)rr FutureWarningrrJrDrBr r r rr7vsr7cCs"|}tjjjj|||d|SrA)rDrfx experimentalsymbolic_shapes_constrain_range_for_size)r9min_valmax_valresultr r rrLs  rL)rN) contextlibr1typingrrtyping_extensionsrrtorch._library.utilsrrrr*r!r3r4__annotations__contextmanagerr-r7rLr r r rs  ?   Z