Module optimizers 

GPU-resident optimizer steps executed through scirs2_core::gpu.

Every optimizer here implements crate::GpuOptimizer for f32 and runs a real compute shader: parameters and gradients are uploaded to device buffers, a WGSL kernel from crate::shaders::wgsl is dispatched, and the updated parameters are read back. The per-parameter optimizer state (Adam’s m/v, SGD’s momentum buffer, …) stays resident in device memory between steps; crate::GpuOptimizer::move_to_cpu genuinely downloads it and crate::GpuOptimizer::move_to_gpu genuinely uploads it again.

Backend support

Only the WebGPU backend (wgpu feature → Vulkan / Metal / DX12) has a complete compute path in scirs2-core 0.6.x. Constructing an optimizer with any other backend returns GpuOptimError::UnsupportedOperation rather than silently computing nothing — in particular the Metal backend registers empty kernel sources for the optimizer kernels, and CUDA was removed from scirs2-core in 0.6.x.

Precision

WGSL compute shaders are f32. The trait is therefore implemented for f32 only; there is no f64 GPU path and none is faked.

Structs

AdagradParams

Hyper-parameters for GpuAdagrad.

AdamParams

Hyper-parameters shared by GpuAdam and GpuAdamW.

GpuAdagrad

GPU Adagrad with learning-rate decay.

GpuAdam

GPU Adam with coupled L2 weight decay, numerically matching optirs_core::optimizers::Adam.

GpuAdamW

GPU AdamW with decoupled weight decay: the decay term is applied to the parameter and never enters the moment estimates.

GpuLamb

GPU LAMB with a real layer-wise trust ratio.

GpuOptimizerConfig

Configuration shared by every GPU optimizer.

GpuRmsprop

GPU RMSprop with optional centering and momentum.

GpuSgd

GPU stochastic gradient descent with optional momentum / Nesterov.

RmspropParams

Hyper-parameters for GpuRmsprop.

SgdParams

Hyper-parameters for GpuSgd.

Constants

SUPPORTED_BACKENDS

Backends that can execute the optimizer kernels, in preference order.