CUDA Runtime#

cutlass.cuda provides host-side CUDA helpers for context setup, stream management, memory copies, device queries, and loading or launching compiled kernels.

These APIs are host runtime utilities. They are not jittable DSL operations and should be used from Python host code, not from inside @cute.jit or @cute.kernel bodies.

Error-checking helpers and internal context-retain wrappers are intentionally omitted from this page. They are used by the runtime implementation rather than being part of the recommended cutlass.cuda user surface.

Context and Device#

class cutlass.cuda.DeviceInfo(*args: Any, **kwargs: Any)#

cutlass.cuda.initialize_cuda_context()#

cutlass.cuda.context_destroy()#

cutlass.cuda.get_device_info()#

cutlass.cuda.get_compute_capability_major_minor()#

cutlass.cuda.get_current_device()#

cutlass.cuda.get_device()#

cutlass.cuda.get_device_attribute()#

cutlass.cuda.get_driver_version()#

Streams#

cutlass.cuda.default_stream()#

cutlass.cuda.stream_create()#

cutlass.cuda.stream_destroy()#

cutlass.cuda.stream_sync()#

Memory#

cutlass.cuda.allocate()#

cutlass.cuda.deallocate()#

cutlass.cuda.memcpy_h2d()#

cutlass.cuda.memcpy_d2h()#

Modules and Kernel Launch#

cutlass.cuda.load_cubin_module()#

cutlass.cuda.load_cubin_module_data()#

cutlass.cuda.unload_cubin_module()#

cutlass.cuda.get_kernel_function()#

cutlass.cuda.load_library()#

cutlass.cuda.load_library_data()#

cutlass.cuda.load_library_from_file()#

cutlass.cuda.unload_library()#

cutlass.cuda.get_library_kernel()#

cutlass.cuda.get_function_from_kernel()#

cutlass.cuda.set_kernel_attribute()#

cutlass.cuda.launch_kernel()#