Clorch: LibTorch Deep Learning for Clojure

Clorch brings PyTorch-style tensors, autograd, neural networks, and distributed CUDA training to Clojure via high-performance LibTorch bindings.

Clorch is a Clojure deep-learning library backed by LibTorch — the C++ engine that powers PyTorch. It gives you PyTorch-style tensors, automatic differentiation, neural-network modules, optimizers, data loaders, mixed-precision training, and NCCL distributed training through a REPL-friendly Clojure API.

Installation

Add Clorch to your project and configure the Java runtime

Quickstart

Build and train your first neural network in minutes

Tensors

Create, manipulate, and inspect LibTorch tensors from Clojure

Neural Networks

Build models with built-in layers and the defmodel macro

Distributed Training

Multi-GPU NCCL training with DDP, AMP, and checkpoints

LLM Components

RoPE, GQA, SwiGLU, KV cache, and autoregressive generation

What Clorch provides

Clorch maps directly to PyTorch’s conceptual model while presenting an idiomatic Clojure API. Every heavy operation executes in native C++ via JavaCPP, bypassing JVM overhead for tensor math.

Tensor Ops

~170 operations covering arithmetic, reductions, shape manipulation, and sampling

Autograd

Automatic differentiation with requires-grad, backward passes, and no-grad scopes

Memory Control

Deterministic native-memory scopes via with-torch and PointerScope

Optimizers

SGD, Adam, AdamW, RMSprop, and Adagrad wrapping LibTorch C++ implementations

Distributions

Probability distributions with sample, log-prob, mean, and variance

AMP

float16/bfloat16 autocast and dynamic gradient scaling for CUDA training

Getting started

1. Add the dependency

   Add Clorch to your deps.edn using a git coordinate pointing to the latest tag.

   deps.edn

   {:deps {io.github.antlobach/clorch
           {:git/tag "v0.2.0"
            :git/sha "07642acdbc522e8aa2a20cd223912247614d2239"}}}
   
2. Start a REPL

   Launch the Clojure REPL from your project directory. Clorch automatically selects CPU or CUDA natives on startup.

   clj
   
3. Create tensors and run autograd

   Require the core namespaces and start experimenting with tensors and gradients immediately.

   (require '[clorch.torch :as t]
            '[clorch.autograd :as autograd])
   
   (def x (t/tensor [2.0] {:requires-grad true}))
   (def y (t/pow x 3))
   (autograd/backward y)
   (autograd/grad x) ; => [12.0]
   
4. Build a neural network

   Compose layers with nn/sequential or define custom architectures with nn/defmodel.

   (require '[clorch.nn :as nn])
   
   (def model
     (nn/sequential
       (nn/linear 10 64)
       (nn/relu)
       (nn/linear 64 1)))
   
   (nn/forward model (t/randn [4 10]))
   

PyTorch namespace mapping

Clorch follows PyTorch’s namespace structure. If you know PyTorch, you can navigate Clorch immediately.

PyTorch	Clorch
torch	clorch.torch
torch.nn	clorch.nn
torch.nn.functional	clorch.nn.functional
torch.optim	clorch.optim
torch.autograd	clorch.autograd
torch.cuda	clorch.cuda
torch.amp	clorch.amp
torch.distributed	clorch.distributed
DistributedDataParallel	clorch.nn.parallel
torch.distributions	clorch.distributions
torch.linalg	clorch.linalg
Dataset / DataLoader	clorch.data

Clorch follows PyTorch concepts but does not expose every PyTorch symbol. Check the API documentation or source before translating a Python call directly.