diff --git a/tinygrad/runtime/ops_cuda.py b/tinygrad/runtime/ops_cuda.py index 6aa899269b..329e7c2025 100644 --- a/tinygrad/runtime/ops_cuda.py +++ b/tinygrad/runtime/ops_cuda.py @@ -55,14 +55,12 @@ class PTXCompiler(Compiler): def __init__(self, arch:str): self.arch = arch self.version = "7.8" if arch >= "sm_89" else "7.5" - #PTXCompiler.compiler_opts = replace(PTXCompiler.compiler_opts, has_tensor_cores=int(arch[3:]) >= 80) super().__init__(f"compile_ptx_{self.arch}") def compile(self, src:str) -> bytes: return src.replace("TARGET", self.arch).replace("VERSION", self.version).encode() class CUDACompiler(Compiler): def __init__(self, arch:str): self.arch = arch - #CUDACompiler.compiler_opts = replace(CUDACompiler.compiler_opts, has_tensor_cores=int(arch[3:]) >= 80) check(cuda.nvrtcVersion((nvrtcMajor := ctypes.c_int()), (nvrtcMinor := ctypes.c_int()))) self.compile_options = [f'--gpu-architecture={arch}', "-I/usr/local/cuda/include", "-I/usr/include", "-I/opt/cuda/include/"] if (nvrtcMajor.value, nvrtcMinor.value) >= (12, 4): self.compile_options.append("--minimal")