Environment variables


This section relates to environment variables that impact Numba’s runtime, for compile time environment variables see Build time environment variables and configuration of optional components.

Numba allows its behaviour to be changed through the use of environment variables. Unless otherwise mentioned, those variables have integer values and default to zero.

For convenience, Numba also supports the use of a configuration file to persist configuration settings. Note: To use this feature pyyaml must be installed.

The configuration file must be named .numba_config.yaml and be present in the directory from which the Python interpreter is invoked. The configuration file, if present, is read for configuration settings before the environment variables are searched. This means that the environment variable settings will override the settings obtained from a configuration file (the configuration file is for setting permanent preferences whereas the environment variables are for ephemeral preferences).

The format of the configuration file is a dictionary in YAML format that maps the environment variables below (without the NUMBA_ prefix) to a desired value. For example, to permanently switch on developer mode (NUMBA_DEVELOPER_MODE environment variable) and control flow graph printing (NUMBA_DUMP_CFG environment variable), create a configuration file with the contents:

developer_mode: 1
dump_cfg: 1

This can be especially useful in the case of wanting to use a set color scheme based on terminal background color. For example, if the terminal background color is black, the dark_bg color scheme would be well suited and can be set for permanent use by adding:

color_scheme: dark_bg

Jit flags

These variables globally override flags to the jit() decorator.


If set to 0 or 1, globally disable or enable bounds checking, respectively. The default if the variable is not set or set to an empty string is to use the boundscheck flag passed to the jit() decorator for a given function. See the documentation of @jit for more information.

Note, due to limitations in numba, the bounds checking currently produces exception messages that do not match those from NumPy. If you set NUMBA_FULL_TRACEBACKS=1, the full exception message with the axis, index, and shape information will be printed to the terminal.


These variables influence what is printed out during compilation of JIT functions.


If set to non-zero, developer mode produces full tracebacks and disables help instructions. Default is zero.


If set to non-zero, enable full tracebacks when an exception occurs. Defaults to the value set by NUMBA_DEVELOPER_MODE.


If set to non-zero, show resources for getting help. Default is zero.


Alters the way in which Numba captures and handles exceptions that do not inherit from numba.core.errors.NumbaError during compilation (e.g. standard Python exceptions). This does not impact runtime exception handling. Valid values are:

  • "old_style" (default): this is the exception handling behaviour that is present in Numba versions <= 0.54.x. Numba will capture and wrap all errors occurring in compilation and depending on the compilation phase they will likely materialize as part of the message in a TypingError or a LoweringError.

  • "new_style" this will treat any exception that does not inherit from numba.core.errors.NumbaError and is raised during compilation as a “hard error”, i.e. the exception will propagate and compilation will halt. The purpose of this new style is to differentiate between intentionally raised exceptions and those which occur due to mistakes. For example, if an AttributeError occurs in the typing of an @overload function, under this new behaviour it is assumed that this a mistake in the implementation and compilation will halt due to this exception. This behaviour will eventually become the default.


If set to non-zero error message highlighting is disabled. This is useful for running the test suite on CI systems.


Alters the color scheme used in error reporting (requires the colorama package to be installed to work). Valid values are:

  • no_color No color added, just bold font weighting.

  • dark_bg Suitable for terminals with a dark background.

  • light_bg Suitable for terminals with a light background.

  • blue_bg Suitable for terminals with a blue background.

  • jupyter_nb Suitable for use in Jupyter Notebooks.

Default value: no_color. The type of the value is string.


If set to non-zero and pygments is installed, syntax highlighting is applied to Numba IR, LLVM IR and assembly dumps. Default is zero.


If set to non-zero the issuing of performance warnings is disabled. Default is zero.


If set to non-zero, print out all possible debugging information during function compilation. Finer-grained control can be obtained using other variables below.


If set to non-zero, print out debugging information during operation of the compiler frontend, up to and including generation of the Numba Intermediate Representation.


If set to non-zero, print out debugging information at runtime about the use of Numba run time (NRT) reference count operations. If set to non-zero, this also switches on the filling of all NRT allocated regions with an identifiable “marker” byte pattern, 0xCB on allocation and 0xDE on deallocation, both to help with debugging memory leaks.


If set to non-zero, enable the Numba run time (NRT) statistics counters. These counters are enabled process wide on import of Numba and are atomic.


If set to non-zero, enable debug for the full application by setting the default value of the debug option in jit. Beware that enabling debug info significantly increases the memory consumption for each compiled function. Default value equals to the value of NUMBA_ENABLE_PROFILING.


If set to non-zero, extend the lifetime of variables to the end of the block in which their lifetime ends. This is particularly useful in conjunction with NUMBA_DEBUGINFO as it helps with introspection of values. Default is zero.


Set the gdb binary for use in Numba’s gdb support. This takes one of two forms: 1) a path and full name of the binary to explicitly express which binary to use 2) just the name of the binary and the current path will be searched using the standard path resolution rules. For example: /path/from/root/to/binary/name_of_gdb_binary or custom_gdb_binary_name. This is to permit the use of a gdb from a non-default location with a non-default name. The default value is gdb.


If set to non-zero, print out debugging information about type inference.


Enables JIT events of LLVM in order to support profiling of jitted functions. This option is automatically enabled under certain profilers.


If set to non-zero, trace certain function calls (function entry and exit events, including arguments and return values).


If defined, chrome tracing is enabled and this variable specifies the filepath of the chrome tracing json file output. The emitted file can be opened by a Chromium-based browser using the profile viewer at chrome://tracing/.


This feature is not supported in multi-process applications.


If set to non-zero, print out the Python bytecode of compiled functions.


If set to non-zero, print out information about the Control Flow Graph of compiled functions.


If set to non-zero, print out the Numba Intermediate Representation of compiled functions.


If set to non-zero, print out the Numba Intermediate Representation of compiled functions after conversion to Static Single Assignment (SSA) form.


Dump the Numba IR after declared pass(es). This is useful for debugging IR changes made by given passes. Accepted values are:

  • Any pass name (as given by the .name() method on the class)

  • Multiple pass names as a comma separated list, i.e. "foo_pass,bar_pass"

  • The token "all", which will print after all passes.

The default value is "none" so as to prevent output.


If set to non-zero, print out types annotations for compiled functions.


Dump the unoptimized LLVM assembly source of compiled functions. Unoptimized code is usually very verbose; therefore, NUMBA_DUMP_OPTIMIZED is recommended instead.


Dump the LLVM assembly source after the LLVM “function optimization” pass, but before the “module optimization” pass. This is useful mostly when developing Numba itself, otherwise use NUMBA_DUMP_OPTIMIZED.


Dump the LLVM assembly source of compiled functions after all optimization passes. The output includes the raw function as well as its CPython-compatible wrapper (whose name begins with wrapper.). Note that the function is often inlined inside the wrapper, as well.


Dump debugging information related to the processing associated with the parallel=True jit decorator option.


Dump debugging information related to the runtime scheduler associated with the parallel=True jit decorator option.


Dump statistics about how many operators/calls are converted to parallel for-loops and how many are fused together, which are associated with the parallel=True jit decorator option.


If set to an integer value between 1 and 4 (inclusive) diagnostic information about parallel transforms undertaken by Numba will be written to STDOUT. The higher the value set the more detailed the information produced.


Dump the native assembly code of compiled functions.


Set to 1 to enable recording of pass timings in LLVM; e.g. NUMBA_LLVM_PASS_TIMINGS=1. See Notes on timing LLVM.

Default value: 0 (Off)

Compilation options


The optimization level; typically this option is passed straight to LLVM. It may take one of the values 0, 1, 2 or 3 which correspond approximately to the -O{value} flag found in many command line compilation tools. The value max is also supported, this is Numba specific, it has the effect of running with the optimization level set at 3 both before and after a pass which in which reference count operation pruning takes place. In some cases this may increase performance, in other cases it may impede performance, the same can be said for compilation time. This option is present to give users the opportunity to choose a value suitable for their application.

Default value: 3


If set to non-zero, enable LLVM loop vectorization.

Default value: 1


If set to non-zero, enable LLVM superword-level parallelism vectorization. Note that use of this feature has occasionally resulted in LLVM producing miscompilations, hence it is off by default.

Default value: 0


If set to non-zero, enable AVX optimizations in LLVM. This is disabled by default on Sandy Bridge and Ivy Bridge architectures as it can sometimes result in slower code on those platforms.


If set to non-zero and Intel SVML is available, the use of SVML will be disabled.


Disable JIT compilation entirely. The jit() decorator acts as if it performs no operation, and the invocation of decorated functions calls the original Python function instead of a compiled version. This can be useful if you want to run the Python debugger over your code.


Override CPU and CPU features detection. By setting NUMBA_CPU_NAME=generic, a generic CPU model is picked for the CPU architecture and the feature list (NUMBA_CPU_FEATURES) defaults to empty. CPU features must be listed with the format +feature1,-feature2 where + indicates enable and - indicates disable. For example, +sse,+sse2,-avx,-avx2 enables SSE and SSE2, and disables AVX and AVX2.

These settings are passed to LLVM for configuring the compilation target. To get a list of available options, use the llc commandline tool from LLVM, for example:

llc -march=x86 -mattr=help


To force all caching functions (@jit(cache=True)) to emit portable code (portable within the same architecture and OS), simply set NUMBA_CPU_NAME=generic.


Override the size of the function cache for retaining recently deserialized functions in memory. In systems like Dask, it is common for functions to be deserialized multiple times. Numba will cache functions as long as there is a reference somewhere in the interpreter. This cache size variable controls how many functions that are no longer referenced will also be retained, just in case they show up in the future. The implementation of this is not a true LRU, but the large size of the cache should be sufficient for most situations.

Note: this is unrelated to the compilation cache.

Default value: 128


Turns on the LLVM pass level reference-count pruning pass and disables the regex based implementation in Numba.

Default value: 1 (On)


When NUMBA_LLVM_REFPRUNE_PASS is on, this allows configuration of subpasses in the reference-count pruning LLVM pass.

Valid values are any combinations of the below separated by , (case-insensitive):

  • all: enable all subpasses.

  • per_bb: enable per-basic-block level pruning, which is same as the old regex based implementation.

  • diamond: enable inter-basic-block pruning that is a diamond shape pattern, i.e. a single-entry single-exit CFG subgraph where has an incref in the entry and a corresponding decref in the exit.

  • fanout: enable inter-basic-block pruning that has a fanout pattern, i.e. a single-entry multiple-exit CFG subgraph where the entry has an incref and every exit has a corresponding decref.

  • fanout_raise: same as fanout but allow subgraph exit nodes to be raising an exception and not have a corresponding decref.

For example, all is the same as per_bb, diamond, fanout, fanout_raise

Default value: “all”


Whether llvmlite’s built-in memory manager is enabled. The default is to enable it on 64-bit ARM platforms (macOS on Apple Silicon and Linux on AArch64), where it is needed to ensure ABI compliance, specifically conformance with the requirements for GOT and text segment placement in the large code model.

This environment variable can be used to override the default setting and force it to be enabled (1) or disabled (0). This should not normally be required, but it is provided as an option for debugging and potential workaround situations.

Default value: None (Use the default for the system)


Turns on the experimental RVSDG frontend. It depends on the numba-rvsdg package and only supports Python 3.11 partially. This option will be removed when the RVSDG frontend fully replaces the old frontend.

Default value: 0 (Off)

Caching options

Options for the compilation cache.


If set to non-zero, print out information about operation of the JIT compilation cache.


Override the location of the cache directory. If defined, this should be a valid directory path.

If not defined, Numba picks the cache directory in the following order:

  1. In-tree cache. Put the cache next to the corresponding source file under a __pycache__ directory following how .pyc files are stored.

  2. User-wide cache. Put the cache in the user’s application directory using appdirs.user_cache_dir from the Appdirs package.

  3. IPython cache. Put the cache in an IPython specific application directory. Stores are made under the numba_cache in the directory returned by IPython.paths.get_ipython_cache_dir().

Also see docs on cache sharing and docs on cache clearing

GPU support


If set to non-zero, disable CUDA support.


If set, force the CUDA compute capability to the given version (a string of the type major.minor), regardless of attached devices.


The default compute capability (a string of the type major.minor) to target when compiling to PTX using cuda.compile_ptx. The default is 5.0, which is the lowest non-deprecated compute capability in the most recent version of the CUDA toolkit supported (12.4 at present).


If set, don’t compile and execute code for the GPU, but use the CUDA Simulator instead. For debugging purposes.


Whether to synchronize on streams provided by objects imported using the CUDA Array Interface. This defaults to 1. If set to 0, then no synchronization takes place, and the user of Numba (and other CUDA libraries) is responsible for ensuring correctness with respect to synchronization on streams.


For debugging purposes. If no other logging is configured, the value of this variable is the logging level for CUDA API calls. The default value is CRITICAL - to trace all API calls on standard error, set this to DEBUG.


By default the CUDA API call logs only give the names of functions called. Setting this variable to 1 also includes the values of arguments to Driver API calls in the logs.


Path of the directory in which the CUDA driver libraries are to be found. Normally this should not need to be set as Numba can locate the driver in standard locations. However, this variable can be used if the driver is in a non-standard location.


Buffer size for logs produced by CUDA driver API operations. This defaults to 1024 and should not normally need to be modified - however, if an error in an API call produces a large amount of output that appears to be truncated (perhaps due to multiple long function names, for example) then this variable can be used to increase the buffer size and view the full error message.


Whether the CUDA driver should produce verbose log messages. Defaults to 1, indicating that verbose messaging is enabled. This should not need to be modified under normal circumstances.


When set to 1, the default stream is the per-thread default stream. When set to 0, the default stream is the legacy default stream. This defaults to 0, for the legacy default stream. See Stream Synchronization Behavior for an explanation of the legacy and per-thread default streams.

This variable only takes effect when using Numba’s internal CUDA bindings; when using the NVIDIA bindings, use the environment variable CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM instead.

See also

The Default Stream section in the NVIDIA Bindings documentation.


Enable warnings if the grid size is too small relative to the number of streaming multiprocessors (SM). This option is on by default (default value is 1).

The heuristic checked is whether gridsize < 2 * (number of SMs). NOTE: The absence of a warning does not imply a good gridsize relative to the number of SMs. Disabling this warning will reduce the number of CUDA API calls (during JIT compilation), as the heuristic needs to check the number of SMs available on the device in the current context.


Enable warnings if a kernel is launched with host memory which forces a copy to and from the device. This option is on by default (default value is 1).


When set to 1, Numba will attempt to use the NVIDIA CUDA Python binding to make calls to the driver API instead of using its own ctypes binding. This defaults to 0 (off), as the NVIDIA binding is currently missing support for Per-Thread Default Streams and the profiler APIs.


The location of the CUDA include files. This is used when linking CUDA C/C++ sources to Python kernels, and needs to be correctly set for CUDA includes to be available to linked C/C++ sources. On Linux, it defaults to /usr/local/cuda/include. On Windows, the default is $env:CUDA_PATH\include.

Threading Control


If set, the number of threads in the thread pool for the parallel CPU target will take this value. Must be greater than zero. This value is independent of OMP_NUM_THREADS and MKL_NUM_THREADS.

Default value: The number of CPU cores on the system as determined at run time. This can be accessed via numba.config.NUMBA_DEFAULT_NUM_THREADS.

See also the section on Setting the Number of Threads for information on how to set the number of threads at runtime.


This environment variable controls the library used for concurrent execution for the CPU parallel targets (@vectorize(target='parallel'), @guvectorize(target='parallel') and @njit(parallel=True)). The variable type is string and by default is default which will select a threading layer based on what is available in the runtime. The valid values are (for more information about these see the threading layer documentation):

  • default - select a threading layer based on what is available in the current runtime.

  • safe - select a threading layer that is both fork and thread safe (requires the TBB package).

  • forksafe - select a threading layer that is fork safe.

  • threadsafe - select a threading layer that is thread safe.

  • tbb - A threading layer backed by Intel TBB.

  • omp - A threading layer backed by OpenMP.

  • workqueue - A simple built-in work-sharing task scheduler.


This environment variable controls the order in which the libraries used for concurrent execution, for the CPU parallel targets (@vectorize(target='parallel'), @guvectorize(target='parallel') and @njit(parallel=True)), are prioritized for use. The variable type is string and by default is tbb omp workqueue, with the priority taken based on position from the left of the string, left most being the highest. Valid values are any permutation of the three choices (for more information about these see the threading layer documentation.)