Trying to run cuda from torch inline extension, so file is not getting created while loading the module

C++
1

I am new to cuda programming and trying to run a hello world program through notebooks in python like this :

%load_ext wurlitzer

import torch, os
from torch.utils.cpp_extension import load_inline


os.environ['CUDA_LAUNCH_BLOCKING'] = '1'

def load_cuda(cuda_src, cpp_src, funcs, opt=False, verbose=False):
    return load_inline(cuda_sources=[cuda_src], cpp_sources=[cpp_src], functions=funcs,
                       extra_cuda_cflags=["-O2"] if opt else [], verbose=verbose, name="inline_ext")



cuda_begin = r"""
#include 
#include 
#include 

#define CHECK_CUDA(x) TORCH_CHECK(x.device().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)

inline unsigned int cdiv(unsigned int a, unsigned int b) { return (a + b - 1) / b;}
"""

cuda_src = cuda_begin + r"""
__global__ void rgb_to_grayscale_kernel(unsigned char* x, unsigned char* out, int n) {
    int i = blockIdx.x*blockDim.x + threadIdx.x;
    if (i>>(
        input.data_ptr(), output.data_ptr(), w*h);
    C10_CUDA_KERNEL_LAUNCH_CHECK();
    return output;
}"""


cpp_src = "torch::Tensor rgb_to_grayscale(torch::Tensor input);"
module = load_cuda(cuda_src, cpp_src, ['rgb_to_grayscale'], verbose=True)

This results in an error saying :

ImportError: /home/zeus/.cache/torch_extensions/py311_cu121/inline_ext/inline_ext.so: cannot open shared object file: No such file or directory

I have already tried deleting the cache dir and re-running the code but keep getting teh same error. Not sure how to debug this, can someone help?

2

Could you check the permissions of the cache folder and make sure your Python process has the needed rights?