How canreader/fastnn works
FastNN is a Rust deep-learning library. Its principal workflow loads and batches data, runs a neural-network model through tensor operations to produce a loss, backpropagates gradients into shared parameter slots, then updates those parameters with an optimizer. Separate branches cover transformer generation, CUDA execution, data sources, schedules, and checkpoint interchange. Relationships among the main training stages are documented in the README; implementation-level links are included only where the supplied excerpts or source index support them.
Main entry points: ops Tensor operations · loader.rs DataLoader · nn Neural layers · loss.rs Loss functions
Components 21
Data input
- MNIST dataset
src/data/mnist.rs - Dataset interface
src/data/dataset.rs - DataLoader
src/data/loader.rs
Model and tensors
- Neural layers
src/nn - Transformers and generation
src/nn - Loss functions
src/nn/loss.rs - Tensor operations
src/tensor/ops
Training
- Reverse-mode autograd
src/autograd/engine.rs - Backward rules
src/autograd/ops - Parameter slots
src/nn/param.rs - Optimizers
src/optim - Learning-rate schedules
src/optim/schedule.rs
Compute backend
- Tensor storage
src/tensor/storage.rs - CUDA kernels and bindings
src/cuda/kernels.rs - CUDA FFI
src/cuda/ffi.rs
Persistence
- Native checkpoints
src/serialize/checkpoint.rs - Training-state checkpoints
src/serialize/training.rs - Safetensors interchange
src/serialize/safetensors.rs - Checkpoint files
Other components
- Training caller
- CUDA device
Connections 28
- Training caller → MNIST dataset: loads data
- MNIST dataset → Dataset interface: implements
src/data/mnist.rs - Training caller → DataLoader: batches data
- DataLoader → Dataset interface: reads samples
src/data/loader.rs - DataLoader → Tensor operations: places batches
- Training caller → Neural layers: runs model
- Neural layers → Tensor operations: computes with
src/tensor/ops/conv.rs - Training caller → Loss functions: computes loss
- Loss functions → Tensor operations: uses operations
src/nn/loss.rs - Neural layers → Parameter slots: holds handles
src/nn/module.rs - Loss functions → Reverse-mode autograd: starts backward
- Reverse-mode autograd → Backward rules: applies rules
src/autograd/engine.rs - Reverse-mode autograd → Parameter slots: accumulates gradients
src/autograd/engine.rs - Training caller → Optimizers: updates parameters
- Optimizers → Parameter slots: reads and updates
src/optim/mod.rs - Learning-rate schedules → Optimizers: sets learning rate
src/optim/mod.rs - Training caller → Transformers and generation: builds or generates
- Tensor operations → Tensor storage: uses storage
src/tensor/ops/mod.rs - Tensor storage → CUDA kernels and bindings: allocates GPU buffers
src/tensor/storage.rs - Tensor operations → CUDA kernels and bindings: dispatches GPU ops
src/tensor/ops/mod.rs - CUDA kernels and bindings → CUDA FFI: calls bindings
src/cuda/kernels.rs - CUDA FFI → CUDA device: launches kernels
- Training caller → Native checkpoints: saves or loads
- Training caller → Training-state checkpoints: resumes training
- Training caller → Safetensors interchange: imports or exports
- Native checkpoints → Checkpoint files: reads and writes
- Training-state checkpoints → Checkpoint files: reads and writes
- Safetensors interchange → Checkpoint files: reads and writes