ferrotorch

Top-level re-export crate for the ferrotorch deep learning framework.

What it provides

This is the umbrella crate that re-exports all ferrotorch sub-crates through a single dependency. Add ferrotorch to your Cargo.toml and access everything via submodules:

Always included

ferrotorch-core — Tensor, autograd, 80+ differentiable ops, complex / sparse / named tensors, FFT, signal, masked, quantization, einops
ferrotorch-nn — Module trait, 30+ layers (Linear, Conv1d/2d, LSTM, GRU, Attention, norms, activations, lazy modules), losses, LoRA
ferrotorch-optim — 18 optimizers (SGD, Adam, AdamW, Adamax, NAdam, RAdam, Adagrad, Adadelta, Adafactor, RMSprop, Rprop, ASGD, SparseAdam, L-BFGS, Muon, K-FAC, EMA, SWA), 12+ LR schedulers, gradient clipping
ferrotorch-data — Dataset, DataLoader (parallel via rayon), DistributedSampler, collation, transforms, Apache Arrow / Polars interop (feature arrow)
ferrotorch-vision — 10 classification (ResNet, VGG, ViT-B/16, EfficientNet-B0, ConvNeXt-T, Swin-T, MobileNetV2, MobileNetV3, DenseNet-121, InceptionV3), 3 detection (Faster R-CNN, Mask R-CNN, SSD300), 2 segmentation (DeepLabV3, FCN), plus U-Net and YOLO; MNIST/CIFAR/ImageFolder datasets; image I/O

Default features (opt-out with `default-features = false`)

ferrotorch-train — Learner training loop, callbacks, metrics, checkpointing
ferrotorch-serialize — ONNX export (via IrGraph), PyTorch .pt import, SafeTensors, GGUF (parse/load/dequantize)
ferrotorch-jit — Tracing JIT, IR graph, optimization passes, code generation
ferrotorch-jit-script — #[script] proc macro for source-based graph capture
ferrotorch-distributions — Probability distributions for sampling and VI
ferrotorch-profiler — Performance profiling with Chrome trace export
ferrotorch-hub — Model hub for downloading and caching pretrained weights
ferrotorch-tokenize — HuggingFace tokenizer wrapper (BPE, WordPiece, Unigram)

Optional features (opt-in)

gpu — NVIDIA CUDA backend with PTX kernels, cuBLAS, cuSOLVER, cuFFT (cargo add ferrotorch --features gpu)
cubecl — Portable GPU via CubeCL: CUDA + WGPU/AMD + ROCm (--features cubecl)
mps — Apple Silicon Metal Performance Shaders backend (--features mps)
xpu — Intel Arc / Data Center GPU Max via CubeCL wgpu (--features xpu)
distributed — DDP, collective ops, TCP / Gloo backends (--features distributed)
llama — Llama 3 model composition + GPU bf16 inference (--features llama)
ml — Sklearn-compatible adapter and classic-ML datasets (--features ml)

Quick start

use ferrotorch::prelude::*;

fn main() -> FerrotorchResult<()> {
    // Autograd
    let a = scalar(2.0f32)?.requires_grad_(true);
    let b = scalar(3.0f32)?.requires_grad_(true);
    let c = (&a * &b)?;
    c.backward()?;
    println!("{}", a.grad()?.unwrap()); // tensor(3.)

    // Neural network
    let model = Sequential::new(vec![
        Box::new(Linear::new(784, 256, true)?),
        Box::new(ReLU::default()),
        Box::new(Linear::new(256, 10, true)?),
    ]);
    let x = rand::<f32>(&[32, 784])?;
    let out = model.forward(&x)?;
    Ok(())
}

Submodules

// Always-on submodules (matching the `[features]` defaults):
use ferrotorch::nn::*;            // layers, losses, activations
use ferrotorch::optim::*;         // optimizers, schedulers
use ferrotorch::data::*;          // datasets, dataloaders, samplers
use ferrotorch::vision::*;        // models, transforms
use ferrotorch::train::*;         // Learner, callbacks, metrics
use ferrotorch::serialize::*;     // SafeTensors, GGUF, ONNX, PyTorch import
use ferrotorch::jit::*;           // tracing JIT, IR graph
use ferrotorch::distributions::*; // probability distributions
use ferrotorch::profiler::*;      // Profiler, Chrome trace export
use ferrotorch::hub::*;           // pretrained model registry + cache
use ferrotorch::tokenize::*;      // HuggingFace tokenizer wrapper
// Feature-gated submodules (opt-in via the matching feature flag):
// use ferrotorch::gpu::*;         // `gpu`         — CUDA backend
// use ferrotorch::cubecl::*;      // `cubecl`      — portable CUDA/WGPU/ROCm
// use ferrotorch::mps::*;         // `mps`         — Apple Silicon Metal
// use ferrotorch::xpu::*;         // `xpu`         — Intel Arc via CubeCL wgpu
// use ferrotorch::distributed::*; // `distributed` — DDP, collectives, TCP/Gloo
// use ferrotorch::llama::*;       // `llama`       — Llama 3 model + inferencer
// use ferrotorch::ml::*;          // `ml`          — sklearn-compatible adapter

Part of ferrotorch

This is the top-level crate of the ferrotorch workspace. See the workspace README for full documentation.

License

MIT OR Apache-2.0

ferrotorch 0.6.2