CUDA Python Bindings

13.2.0 · active · verified Sun Mar 29

cuda-bindings provides low-level Python wrappers for the NVIDIA CUDA C driver and runtime APIs. It is a core component of the broader NVIDIA 'CUDA Python' initiative, aiming to unify and simplify GPU-accelerated computing in Python. The current version is 13.2.0, with releases often tied to CUDA Toolkit versions and ongoing development to integrate Python as a first-class language in the CUDA ecosystem.

Warnings

breaking Mismatch between CUDA Toolkit, NVIDIA GPU driver, and `cuda-bindings` versions is a common source of runtime errors, including 'CUDA Driver Version Insufficient', 'No Kernel Image Available', or failure to find CUDA-enabled devices.
Fix: Ensure that your installed NVIDIA GPU driver, CUDA Toolkit, and `cuda-bindings` Python package are compatible. Consult the NVIDIA CUDA Python documentation for compatibility matrices.
gotcha Updating `cuda-python` (which `cuda-bindings` is a part of) from older versions (e.g., v12.6.2.post1 and below) using `pip install -U cuda-python` might fail.
Fix: Perform a clean re-installation by first uninstalling with `pip uninstall -y cuda-python` (or `pip uninstall -y cuda-bindings`) followed by a fresh `pip install cuda-python` (or `pip install cuda-bindings`).
gotcha `cuda-bindings` provides direct, low-level access to the CUDA C APIs. This requires explicit memory management, device context handling, and kernel configuration, which can be more complex than higher-level libraries like Numba CUDA or CuPy.
Fix: Be prepared to work with C types (e.g., `ctypes`) and manage GPU resources explicitly. For many common scientific computing or deep learning tasks, higher-level libraries might offer a simpler abstraction. Consider `cuda.core` for more Pythonic access to CUDA runtime functionalities if raw C API interaction is not strictly necessary.
gotcha Out-of-memory (OOM) errors or illegal memory access can occur when dealing with large datasets or complex models, especially on GPUs with limited VRAM, or due to incorrect memory operations within CUDA kernels.
Fix: Monitor GPU memory usage (`nvidia-smi`). Reduce batch sizes, optimize memory allocation patterns, and ensure correct buffer sizing and indexing in custom CUDA kernels. Explicitly free unused GPU memory if applicable.

Install

pip install cuda-bindings Install cuda-bindings

Imports

cuInit
```
import cuda.cuda as cu
# ... then call cu.cuInit(0)
```
The canonical import is `cuda.cuda` as the `cuda-bindings` package exposes the CUDA C APIs directly under this module.

Quickstart

This quickstart demonstrates how to initialize the CUDA driver, query the number of available CUDA devices, and print basic information for each, such as its name and total memory. It leverages the low-level CUDA C APIs exposed by `cuda-bindings`.

import cuda.cuda as cu
import cuda.cuda.runtime as rt
import ctypes # For C types like c_int, c_size_t

# Initialize CUDA Driver API
cu.cuInit(0)

# Get device count
count = ctypes.c_int()
cu.cuDeviceGetCount(ctypes.byref(count))
print(f"Found {count.value} CUDA devices.")

# Get properties for each device
for i in range(count.value):
    device = cu.CUdevice()
    cu.cuDeviceGet(ctypes.byref(device), i)

    name_buffer = ctypes.create_string_buffer(256)
    cu.cuDeviceGetName(name_buffer, len(name_buffer), device)
    print(f"  Device {i}: {name_buffer.value.decode().strip()}")

    total_mem = ctypes.c_size_t()
    cu.cuDeviceTotalMem(ctypes.byref(total_mem), device)
    print(f"    Total Memory: {total_mem.value / (1024**3):.2f} GB")

view raw JSON →