NVIDIA CUDA NVCC Compiler

13.2.78 verified Fri May 01 auth: no python

Provides the nvcc CUDA compiler as a Python package for building CUDA kernels. Version 13.2.78 corresponds to CUDA 12.8. This package bundles the nvcc compiler and dependencies, simplifying CUDA compilation in Python environments without requiring a system CUDA installation.

pip install nvidia-cuda-nvcc

Common errors

error ModuleNotFoundError: No module named 'nvidia.cuda' ↓

cause Missing or incorrect import path. The package `nvidia-cuda-nvcc` must be installed, and the import path changed after version 12.

fix

Run pip install nvidia-cuda-nvcc and use from nvidia.cuda.nvcc import nvcc_path.

error RuntimeError: nvcc not found. Set CUDA_HOME or ensure nvcc is on PATH. ↓

cause Calling subprocess with 'nvcc' without using the bundled path. The system may lack a CUDA installation.

fix

Use nvcc_path() from the package to get the correct path and pass it to subprocess.

error FileNotFoundError: [Errno 2] No such file or directory: 'nvcc' ↓

cause Similar to above: nvcc is not on PATH. The package's bin directory is not automatically added.

fix

Use from nvidia.cuda.nvcc import nvcc_path; nvcc = nvcc_path() and then subprocess.run([nvcc, ...]).

Warnings

breaking The package name and import structure changed in version 12.x. Previously, the package was installed as `cuda-nvcc` with import `cuda.cc.nvcc`. Now it's `nvidia-cuda-nvcc` with import `nvidia.cuda.nvcc`. Old imports will break. ↓

fix Update import to `from nvidia.cuda.nvcc import nvcc_path` and install the new package.

gotcha The nvcc binary is bundled inside the Python package and not added to PATH. You must use `nvcc_path()` or call it via subprocess; calling `nvcc` directly from shell will not work unless you manually add the package's bin directory to PATH. ↓

fix Always use `from nvidia.cuda.nvcc import nvcc_path` to get the path, then invoke via subprocess.

deprecated Some older versions used `from nvidia.cuda import nvcc` (without the .nvcc submodule). This still works in some versions but is deprecated and may be removed. ↓

fix Switch to explicit `from nvidia.cuda.nvcc import ...`.

Imports

nvcc_path
wrong
```
import nvcc
```
correct
```
from nvidia.cuda.nvcc import nvcc_path
```
nvcc is not a top-level importable symbol; use the correct submodule.

get_cc_version

wrong

from nvidia.cuda.nvcc import cc_version

correct

from nvidia.cuda.nvcc import get_cc_version

The function is named get_cc_version, not cc_version.

Quickstart

Locate the bundled nvcc and compile a minimal CUDA kernel.

from nvidia.cuda.nvcc import nvcc_path
import subprocess
import os

# Get the path to the bundled nvcc
nvcc = nvcc_path()
print(f"Using nvcc at: {nvcc}")

# Simple compile and run a CUDA kernel
code = '''
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}
'''
with open('kernel.cu', 'w') as f:
    f.write(code)
result = subprocess.run([nvcc, 'kernel.cu', '-o', 'kernel'], capture_output=True, text=True)
if result.returncode == 0:
    print("Compilation succeeded")
else:
    print("Compilation failed:", result.stderr)