RARE: range-adaptive residue emulation of FP64 GEMM on integer tensor cores

79bc4c6 15 days ago

339 Bytes

	// Shim for kernel-builder's registration.h, for local JIT builds.
	// cpp_extension.load(is_python_module=False) loads via torch.ops.load_library,
	// so TORCH_LIBRARY runs at dlopen and no PyInit module object is needed.
	#pragma once

	#define TORCH_LIBRARY_EXPAND(NAME, MODULE) TORCH_LIBRARY(NAME, MODULE)
	#define REGISTER_EXTENSION(NAME)