AxonML
PyTorch-Equivalent ML Framework in Pure Rust
A complete machine learning framework with automatic differentiation, neural networks, optimizers, vision, audio, NLP, and LLM architectures. Features ONNX support, model quantization, kernel fusion, and a full CLI/TUI with Weights & Biases integration.
For more information, contact us at devops@automatanexus.com
19 Modular Crates
Use only what you need - each crate is independently publishable
axonml-tensor
N-dimensional tensors, broadcasting, SIMD
axonml-autograd
Dynamic computational graph, backprop
axonml-nn
Linear, Conv, RNN, LSTM, Attention, Loss
axonml-optim
SGD, Adam, AdamW, LR schedulers
axonml-vision
LeNet, ResNet, VGG, ViT architectures
axonml-llm
BERT, GPT-2, text generation
axonml-quant
INT4/INT8 quantization, ~8x compression
axonml-onnx
ONNX import/export, 40+ operators
Key Features
Complete Tensor Library
N-dimensional tensors with NumPy-style broadcasting, efficient views, slicing, and activation functions (ReLU, GELU, SiLU, Softmax).
Automatic Differentiation
Dynamic computational graph with reverse-mode autodiff. Gradient functions for all operations with no_grad context manager.
Neural Network Modules
Linear, Conv1d/2d, BatchNorm, LayerNorm, Dropout, RNN, LSTM, GRU, MultiHeadAttention, and all standard loss functions.
LLM Architectures
BERT encoder with classification and masked LM heads. GPT-2 decoder with top-k, top-p, temperature sampling for text generation.
Model Quantization
INT8 (Q8_0), INT4 (Q4_0, Q4_1), INT5, and F16 formats. Block-based quantization achieves ~8x model size reduction.
Kernel Fusion
Automatic fusion pattern detection. FusedLinear (MatMul + Bias + Activation) delivers up to 2x speedup for memory-bound ops.
Powerful CLI
Complete command-line interface for the entire ML workflow
$ axonml new my-model# Scaffold new project$ axonml train config.toml# Train from config$ axonml export model.axonml --onnx# Export to ONNX$ axonml quant convert model.axonml --type q4_0# Quantize to INT4$ axonml hub download resnet50# Download pretrained$ axonml tui --model model.axonml# Launch TUI dashboardInteractive TUI Dashboard
Terminal-based dashboard for model development
Architecture
├─ Linear(784, 256)
├─ ReLU
├─ Linear(256, 128)
├─ ReLU
└─ Linear(128, 10)
Parameters
Total: 235,146
Trainable: 235,146
Size: 940.6 KB
Status: Ready
Technical Specifications
Ready to Build ML in Rust?
Get started with AxonML and leverage Rust's performance, safety, and concurrency for your machine learning projects.
devops@automatanexus.com