Struct CudaTrainer 

pub struct CudaTrainer { /* private fields */ }

CUDA-accelerated training context

Manages GPU resources and provides high-level training operations.

Implementations

impl CudaTrainer

pub fn new() -> Result<Self>

Create a new CUDA trainer on the default GPU

pub fn with_device(device_id: i32) -> Result<Self>

Create a new CUDA trainer on the specified GPU

pub fn context(&self) -> &Arc<CudaContext>

Get the CUDA context

pub fn stream(&self) -> &CudaStream

Get the CUDA stream

pub fn synchronize(&self) -> Result<()>

Synchronize the stream (wait for all operations to complete)

pub fn upload(&self, data: &[f32]) -> Result<GpuBuffer<f32>>

Allocate a GPU buffer from host data

pub fn zeros(&self, len: usize) -> Result<GpuBuffer<f32>>

Allocate a zero-initialized GPU buffer

pub fn free_memory_mb(&self) -> Option<u64>

Query free VRAM in MB (via cuMemGetInfo). Returns None if query fails.

pub fn download(&self, buffer: &GpuBuffer<f32>) -> Result<Vec<f32>>

Download GPU buffer to host

pub fn matmul_forward( &self, a: &GpuBuffer<f32>, b: &GpuBuffer<f32>, c: &mut GpuBuffer<f32>, m: u32, k: u32, n: u32, ) -> Result<()>

Matrix multiply forward pass: C = A @ B

Arguments

• a: Input matrix (m × k)
• b: Weight matrix (k × n)
• c: Output matrix (m × n)
• m, k, n: Matrix dimensions

pub fn matmul_backward( &self, a: &GpuBuffer<f32>, b: &GpuBuffer<f32>, grad_c: &GpuBuffer<f32>, grad_a: &mut GpuBuffer<f32>, grad_b: &mut GpuBuffer<f32>, m: u32, k: u32, n: u32, ) -> Result<()>

Matrix multiply backward pass for weight gradients

Given C = A @ B, computes:

• grad_A = grad_C @ B^T
• grad_B = A^T @ grad_C

pub fn adamw_step( &mut self, params: &mut GpuBuffer<f32>, grads: &GpuBuffer<f32>, m_state: &mut GpuBuffer<f32>, v_state: &mut GpuBuffer<f32>, lr: f32, beta1: f32, beta2: f32, eps: f32, weight_decay: f32, ) -> Result<()>

AdamW optimizer step on GPU

Updates weights in-place using the AdamW algorithm.

pub fn clip_gradients( &self, grads: &mut GpuBuffer<f32>, max_norm: f32, ) -> Result<()>

Apply gradient clipping

pub fn step_count(&self) -> u32

Get current optimizer step count

pub fn reset_step(&mut self)

Reset optimizer step count (for new training run)

pub fn device_name(&self) -> String

Get device name

pub fn total_memory(&self) -> usize

Get total GPU memory in bytes

Trait Implementations

impl Debug for CudaTrainer

Available on crate feature cuda only.

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

Formats the value using the given formatter.