Expand description
Native CUDA execution backend for the WGSL Forge differential oracle.
This bridges PTX modules emitted by crate::wgsl_forge::emit::ptx onto the
NVIDIA driver API via cudarc 0.19. It implements the same stateless
QualiaCompute contract as the wgpu backend: a single persistent device
slab is pre-allocated, and dispatches receive lightweight byte offsets
(BufferView) rather than allocating in the hot loop.
cudarc 0.19 (driver redesign): allocation, copies, module loading and launch
all hang off a [CudaContext]/[CudaStream] pair. With the default
fallback-dynamic-loading feature the crate compiles without the CUDA
toolkit present and resolves libraries at runtime, so an absent toolkit
degrades to a runtime ForgeError::GpuUnavailable rather than a build
failure.
Structsยง
- Captured
Cuda Graph - Instantiated CUDA graph owned by a prepared runtime.
- Cuda
Compute Context - Cuda
Pipeline