Crate any_gpu 

any-gpu — Tensor engine for every GPU. AMD, NVIDIA, Intel, Apple. One codebase, zero vendor lock-in. wgpu under the hood.

Modules

autograd

nanosign

optim

train

Structs

GpuBuffer

GPU-resident f32 buffer with element count metadata.

GpuDevice

GPU device handle. wgpu picks the right backend — Vulkan, Metal, DX12. One codepath, every vendor.

Tensor

GPU tensor with shape metadata. Wraps a GpuBuffer. Shape is stored inline (max 6 dims covers batch x channel x D x H x W + extra).