Struct FusedAdamState 

pub struct FusedAdamState { /* private fields */ }

Interleaved (parameter, momentum, variance) layout used by the fused CPU update.

This is CPU code

Despite the “kernel fusion” vocabulary, nothing here touches a GPU: there is no allocation, no kernel launch and no device memory. What the module really provides is a blocked, cache-friendly, vectorizable Adam over an interleaved layout, and the accounting below describes that layout — see FusedAdamState::planned_layout_bytes.

Implementations

impl FusedAdamState

pub const MAX_LAYOUT_BYTES: usize

Largest layout this state will plan for, as a sanity bound on caller input.

pub fn new(config: KernelFusionConfig) -> Self

Creates a new fused GPU state.

pub fn allocate_parameter(&mut self, id: String, size: usize) -> Result<()>

Registers a parameter in the interleaved layout plan.

No memory is allocated: the buffer records the parameter’s size, alignment and offset so the fused update can walk it in cache-friendly blocks.

Errors

Returns an error when the requested layout exceeds FusedAdamState::MAX_LAYOUT_BYTES.

pub fn run_fused_adam_block( &mut self, param_id: &str, param: &mut [f32], grad: &[f32], lr: f32, betas: (f32, f32), eps: f32, weight_decay: f32, ) -> Result<()>

Runs the blocked, vectorizable Adam update for one parameter.

The work is done on the CPU in optimal_block_size-sized blocks over the interleaved layout; the block/grid arithmetic below mirrors the tiling a GPU kernel would use, but no kernel is launched.

pub fn launch_multi_param_kernel( &mut self, params: Vec<(&str, &mut [f32], &[f32])>, lr: f32, betas: (f32, f32), eps: f32, weight_decay: f32, ) -> Result<()>

Launches multi-parameter fused kernel.

pub fn fused_layout_stats(&self) -> FusedLayoutStats

Statistics about the interleaved layout plan.

Trait Implementations

impl Debug for FusedAdamState

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.