Skip to main content
Version: v1.5.0

Global Settings

Every Taichi program starts with ti.init(). You can customize your Taichi runtime by passing arguments to ti.init() or setting environment variables. Each argument or environment variable controls one specific behavior of the Taichi runtime. For example, the argument arch specifies the backend, and the argument debug decides whether to run the program in debug mode.

Taichi executes the following to initialize a specific configuration in the ti.init() method call. Taking the arch argument as an example:

  1. Taichi first looks for the arguments passed to ti.init(). In this case, after Taichi reads ti.init(arch=cuda), it chooses CUDA as the backend and ignores the corresponding environment variable for backend setting, namely, TI_ARCH.
  2. If no argument is found, Taichi checks the corresponding environment variable. In this case, if arch is not specified but the environment variable is set to export TI_ARCH=cuda, Taichi still chooses CUDA as the backend.
  3. If no customized setting is found, Taichi uses a default configuration. In this case, if neither the argument arch is specified nor an environment variable TI_ARCH is found, Taichi adopts the default backend arch=ti.cpu.

Following are some frequently-used configurations in ti.init():

Customize Taichi runtime via arguments

[Backend Options]

arch: [ti.cpu, ti.gpu, ti.cuda, ti.vulkan, ...]
Specify which architecture to use.

device_memory_GB: float
Specify the pre-allocated memory size for CUDA.

[Compilation Options]

advanced_optimization: bool
Enable/disable advanced optimization. Turning off the setting can save compile time and reduce possible errors.

fast_math: bool
Enable/disable fast math. Turning off the setting can prevent possible undefined math behavior.

print_ir: bool
Turn on/off the printing of the intermediate IR generated.

[Runtime Options]

cpu_max_num_threads: int
Set the number of threads used by the CPU thread pool.

debug: bool
Run your program in debug mode.

default_cpu_block_dim: int
Set the number of threads in a block on CPU.

default_gpu_block_dim: int
Set the number of threads in a block on GPU.

default_fp: [ti.f32, ti.f64]
Set the default precision of floating-point numbers in the Taichi scope.

default_ip: [ti.i32, ti.i64]
Set the default precision of integers in the Taichi scope.

kernel_profiler: bool
Turn on/off kernel performance profiling.

offline_cache: bool
Enable/disable offline cache of the compiled kernels.

offline_cache_file_path: str
Set a directory for holding the offline cached files.

random_seed: int
Set a custom seed for the random number generator.

[Logging Options]

log_level: [ti.INFO, ti.TRACE, ti.WARN, ti.ERROR, ti.CRITICAL, ti.DEBUG]
Set the logging level.

verbose: bool
Turn on/off verbose outputs. For example, `ti.init(verbose=False)` prevents verbose outputs.

[Develop Options]

gdb_trigger: bool
Enable/disable GDB when Taichi crashes. For example, `ti.init(gdb_trigger=True)` enables GDB.

Customize Taichi runtime via environment variables

Taichi allows a number of environment variables for runtime customization. Some provide an alternative to the ti.init() arguments listed above. For example, as mentioned above, arch is interchangeable with TI_ARCH. But not all of the environment variables can be replaced by arguments, and vice versa.

[Backend Options]

Specify which GPU to use for CUDA: `export CUDA_VISIBLE_DEVICES=[gpuid]`.

Specify which architecture to run the program. For example, `export TI_ARCH=cuda` designates CUDA as the backend.

Disable a backend upon startup. For example, set `export TI_ENABLE_CUDA=0` to disable the CUDA backend.

Specify which GPU to use for VULKAN: `export TI_VISIBLE_DEVICES=[gpuid]`.

[Runtime Options]

Turn on/off debug mode. For example, `export TI_DEBUG=1` activates debug mode.

Enable/disable the import of torch upon startup. For example, `export TI_ENABLE_TORCH=0` prohibits the use of torch.
The default value is 1.

Enable/disable the import of paddle upon startup. For example, `export TI_ENABLE_PADDLE=0` prohibits the use of paddle.
The default value is 1.

[Develop Options]

Enable/disable the caching of compiled runtime bitcode in developer mode. For example, `export TI_CACHE_RUNTIME_BITCODE=1` enables the program to cache compiled runtime bitcode.
Turning off the setting can save startup time.

Specify the number of threads to run a test. For example, set `export TI_TEST_THREADS=4` to allocate four threads.
Alternatively, you can run `python tests/ -t4`.

[Logging Options]

Set the logging level. For example, `export TI_LOG_LEVEL=trace` enables the TRACE level.


  • To specify which architecture to use: ti.init(arch=ti.cuda) designates CUDA as the backend. This argument is equivalent to the environment variable TI_ARCH.
  • To specify the pre-allocated memory size for CUDA: ti.init(device_memory_GB=0.5) allocates 0.5 GB size of memory.
  • To specify which GPU to use for CUDA: export CUDA_VISIBLE_DEVICES=[gpuid].
  • To specify which GPU to use for VULKAN: export TI_VISIBLE_DEVICE=[gpuid].
  • To disable a backend (CUDA, METAL, OPENGL) upon startup: For example, export TI_ENABLE_CUDA=0 disables CUDA.

If you want to use CUDA and Taichi's GGUI system at the same time on a machine with multiple GPU cards, ensure that CUDA_VISIBLE_DEVICES matches TI_VISIBLE_DEVICE. In principle, CUDA_VISIBLE_DEVICES and TI_VISIBLE_DEVICE should point to the same GPU device identified with UUID. Use nvidia-smi -L to retrieve the details of your GPU devices.


  • To disable advanced optimization: ti.init(advanced_optimization=False), which helps save compile time and reduce possible errors.
  • To disable fast math: ti.init(fast_math=False), which helps prevent possible undefined math behavior.
  • To print the intermediate IR generated: ti.init(print_ir=True). Note that compiled kernels are cached by default. To force compilation and IR emission, use ti.init(print_ir=True, offline_cache=False).


  • To restart the entire Taichi system (and erase all fields and kernels): ti.reset().
  • To start a program in debug mode: ti.init(debug=True). Alternatively, you can set the environment variable TI_DEBUG or run your code via ti debug
  • To disable the import of torch upon startup: export TI_ENABLE_TORCH=0.
  • To disable the import of paddle upon startup: export TI_ENABLE_PADDLE=0.
  • To set a custom seed for the random number generator used by ti.random(): ti.init(random_seed=seed). seed should be an integer. An example: ti.init(random_seed=int(time.time())).
  • To set the default precision of floating-point numbers of Taichi runtime to ti.f64: ti.init(default_fp=ti.i64).
  • To set the default precision of floating-point numbers of Taichi runtime to ti.i32: ti.init(default_ip=ti.i32).
  • To disable the offline cache of compiled kernels: ti.init(offline_cache=False). See the Offline cache for more information.
  • To enable the use of variables as indices to access vector/matrix elements in the Taichi scope: ti.init(dynamic_index=True).
  • To turn on kernel profiling: ti.init(kernel_profiler=True). See the Profiler for more information.


  • To set the logging level: ti.init(log_level=ti.TRACE) or ti.set_logging_level(ti.TRACE) enables the TRACE level. The environment variable TI_LOG_LEVEL serves the same purpose.
  • To eliminate verbose outputs: ti.init(verbose=False).


  • To trigger GDB when Taichi crashes: ti.init(gdb_trigger=True).
  • To cache compiled runtime bitcode in dev mode: export TI_CACHE_RUNTIME_BITCODE=1, which saves startup time.
  • To allocate four threads to run a test: export TI_TEST_THREADS=4 or python tests/ -t4.

If ti.init is called twice, the configuration in the first call is discarded. For example:

print(ti.cfg.debug) # True
print(ti.cfg.debug) # False

Going high precision

The default setting fast_math=True may cause unexpected precision problems, which can be hard to debug. For example, the function ti.sqrt (and those that rely on it, such as ti.norm) may have a lower precision on the CUDA backend than on the CPU backends. See this issue for more information. To work around this problem, set the default precision to ti.f64 and turn off the fast_math option:

ti.init(default_fp=ti.f64, fast_math=False)

If you encounter other precision loss problems, set ti.init(fast_math=False) and see if it works. If not, report an issue to us.