Environment variables#

For runtime#

Here are the environment variables that CuPy uses at runtime.


Path to the directory containing CUDA. The parent of the directory containing nvcc is used as default. When nvcc is not found, /usr/local/cuda is used. See Working with Custom CUDA Installation for details.


Default: ${HOME}/.cupy/kernel_cache

Path to the directory to store kernel cache. See Performance Best Practices for details.


Default: 0

If set to 1, CUDA source file will be saved along with compiled binary in the cache directory for debug purpose. Note: the source file will not be saved if the compiled binary is already stored in the cache.


Default: 0

If set to 1, CUPY_CACHE_DIR and CUPY_CACHE_SAVE_CUDA_SOURCE will be ignored, and the cache is in memory. This environment variable allows reducing disk I/O, but is ignoed when nvcc is set to be the compiler backend.


Default: 0

If set to 1, headers loaded by Jitify would not be cached on disk (to CUPY_CACHE_DIR). The default is to always cache.


Default: 0

If set to 1, when CUDA kernel compilation fails, CuPy dumps CUDA kernel code to standard error.


Default: 0

If set to 1, CUDA kernel will be compiled with debug information (--device-debug and --generate-line-info).


Default: 0 (unlimited)

The amount of memory that can be allocated for each device. The value can be specified in absolute bytes or fraction (e.g., "90%") of the total memory of each GPU. See Memory Management for details.


Set the seed for random number generators.


Default: 0

If set to 1, the following syntax is enabled:

cupy_ndarray[:] = numpy_ndarray

Default: "cub" (In ROCm HIP environment, the default value is "". i.e., no accelerators are used.)

A comma-separated string of backend names (cub, cutensor, or cutensornet) which indicates the acceleration backends used in CuPy operations and its priority (in descending order). By default, all accelerators are disabled on HIP and only CUB is enabled on CUDA.


Default: 0

If set to 1, it allows CUDA libraries to use Tensor Cores TF32 compute for 32-bit floating point compute.


Default: 1

This controls CuPy’s behavior as a Consumer. If set to 0, a stream synchronization will not be performed when a device array provided by an external library that implements the CUDA Array Interface is being consumed by CuPy. For more detail, see the Synchronization requirement in the CUDA Array Interface v3 documentation.


Default: 3

This controls CuPy’s behavior as a Producer. If set to 2, the CuPy stream on which the data is being operated will not be exported and thus the Consumer (another library) will not perform any stream synchronization. For more detail, see the Synchronization requirement in the CUDA Array Interface v3 documentation.


Default: 0.6

This controls CuPy’s DLPack support. Currently, setting a value smaller than 0.6 would disguise managed memory as normal device memory, which enables data exchanges with libraries that have not updated their DLPack support, whereas starting 0.6 CUDA managed memory can be correctly recognized as a valid device type.


Default: nvcc

Define the compiler to use when compiling CUDA source. Note that most CuPy kernels are built with NVRTC; this environment variable is only effective for RawKernel/RawModule with the nvcc backend or when using cub as the accelerator.


Default: 0

If set to 1, CuPy will use the CUDA per-thread default stream, effectively causing each host thread to automatically execute in its own stream, unless the CUDA default (null) stream or a user-created stream is specified. If set to 0 (default), the CUDA default (null) stream is used, unless the per-thread default stream (ptds) or a user-created stream is specified.


Default: 0

By default, CuPy directly compiles kernels into SASS (CUBIN) to support CUDA Enhanced Compatibility If set to 1, CuPy instead compiles kernels into PTX and lets CUDA Driver assemble SASS from PTX. This option is only effective for CUDA 11.1 or later; CuPy always compiles into PTX on earlier CUDA versions. Also, this option only applies when NVRTC is selected as the compilation backend. NVCC backend always compiles into SASS (CUBIN).

CUDA Toolkit Environment Variables

In addition to the environment variables listed above, as in any CUDA programs, all of the CUDA environment variables listed in the CUDA Toolkit Documentation will also be honored.


When CUPY_ACCELERATORS or NVCC environment variables are set, g++-6 or later is required as the runtime host compiler. Please refer to Installing CuPy from Source for the details on how to install g++.

For installation#

These environment variables are used during installation (building CuPy from source).


Path to the cuTENSOR root directory that contains lib and include directories. (experimental)


Default: 0

If set to 1, CuPy is built for AMD ROCm Platform (experimental). For building the ROCm support, see Installing Binary Packages for further detail.


Default: 0

If set to 1, CuPy is built using CUDA Python.


Build CuPy for a particular CUDA architecture. For example:


For specifying multiple archs, concatenate the arch=... strings with semicolons (;). If current is specified, then it will automatically detect the currently installed GPU architectures in build time. When this is not set, the default is to support all architectures.


Default: 4

To enable or disable parallel build, sets the number of processes used to build the extensions in parallel.


Default: 2

To enable or disable nvcc parallel compilation, sets the number of threads used to compile files using nvcc.

Additionally, the environment variables CUDA_PATH and NVCC are also respected at build time.