Expand description
The CUDA device backend for teenygrad — driver bindings (via
bindgen against the CUDA headers), device/runtime abstraction (device, runtime),
and the AOT/JIT kernel-compilation path (compiler) used by teeny-kernels’ cuda
feature.
Building this crate requires the CUDA toolkit headers/libs on the host (build.rs links
against them unconditionally); see the crate README for CUDA_INCLUDE_DIR/CUDA_LIB_DIR and
the separate, runtime-only teenyc/TEENYC_PATH requirement for actually compiling kernels.
Modules§
- compiler
- AOT/JIT kernel compilation.
- device
- Device and context management.
- errors
- Error types.
- model
- Loaded-model execution.
- runtime
- CUDA runtime abstraction (streams, memory, launches).
- testing
- Test helpers for
teeny-cuda’s own test suite.
Functions§
- cuda_
profiler_ ⚠start - Signal nsys (or any CUDA profiler) to start capturing.
- cuda_
profiler_ ⚠stop - Signal nsys (or any CUDA profiler) to stop capturing.