diff --git a/tinygrad/runtime/support/compiler_cuda.py b/tinygrad/runtime/support/compiler_cuda.py index 2fd7570751..8f83c34657 100644 --- a/tinygrad/runtime/support/compiler_cuda.py +++ b/tinygrad/runtime/support/compiler_cuda.py @@ -69,9 +69,7 @@ class PTXCompiler(Compiler): def disassemble(self, lib:bytes): cuda_disassemble(lib, self.arch) class NVPTXCompiler(PTXCompiler): - def __init__(self, arch:str): - nvrtc.nvJitLinkVersion(ctypes.byref(ctypes.c_int()), ctypes.byref(ctypes.c_int())) # try to get version to check if jitlink is available - super().__init__(arch, cache_key="nv_ptx") + def __init__(self, arch:str): super().__init__(arch, cache_key="nv_ptx") def compile(self, src:str) -> bytes: jitlink_check(nvrtc.nvJitLinkCreate(handle := nvrtc.nvJitLinkHandle(), 1, to_char_p_p([f'-arch={self.arch}'.encode()])), handle) jitlink_check(nvrtc.nvJitLinkAddData(handle, nvrtc.NVJITLINK_INPUT_PTX, ptxsrc:=super().compile(src), len(ptxsrc), "".encode()), handle)