warp.config

Global configuration settings for Warp.

This module provides settings to control compilation behavior, debugging, performance, and runtime behavior of Warp kernels and modules. Settings exist at the global, module, and kernel levels, with more specific scopes taking precedence.

Settings can be modified by direct assignment before or after calling warp.init(), though some settings only take effect if set prior to initialization. See individual setting documentation for details.

For information on module-level and kernel-level settings, see Configuration.

API

cache_kernels	Enable kernel caching between application launches.
compile_time_trace	Enable the generation of Trace Event Format files for runtime module compilation.
cuda_output	Preferred CUDA output format for kernel compilation.
enable_backward	Enable compilation of kernel backward passes.
enable_graph_capture_module_load_by_default	Enable automatic module loading before graph capture.
enable_mempools_at_init	Enable CUDA memory pools during device initialization when supported.
enable_tiles_in_stack_memory	Use stack memory instead of static memory for tile allocations on the CPU.
enable_vector_component_overwrites	Allow multiple writes to vector/matrix/quaternion components.
kernel_cache_dir	Directory path for storing compiled kernel cache.
line_directives	Enable Python source line mapping in generated code.
lineinfo	Enable the compilation of modules with line information.
llvm_cuda	Use Clang/LLVM compiler instead of NVRTC for CUDA compilation.
load_module_max_workers	Default number of worker threads for compiling and loading modules in parallel.
max_unroll	Maximum unroll factor for loops.
mode	Compilation mode for Warp kernels.
optimization_level	Optimization level for Warp kernels.
print_launches	Enable detailed kernel launch logging.
ptx_target_arch	Target architecture version for PTX generation, e.g., ptx_target_arch = 75.
quiet	Disable Warp module initialization messages.
use_precompiled_headers	Enable the use of precompiled headers during kernel compilation.
verbose	Enable detailed logging during code generation and compilation.
verbose_warnings	Enable extended warning messages with source location information.
verify_autograd_array_access	Enable warnings for array overwrites that may affect gradient computation.
verify_cuda	Enable CUDA error checking after kernel launches.
verify_fp	Enable floating-point verification for inputs and outputs.
version	Warp version string