DeepSeek-RS (Rust Implementation)

Note

DeepSeek-RS is a personal project implementing DeepSeek's architecture in Rust for learning and experimentation. This is not an official DeepSeek project.

DeepSeek-RS is a Rust-based deep-learning framework replicating DeepSeek’s architecture. It utilizes tch-rs (libtorch bindings) for tensor computation and supports large-scale transformer-based models.

---

Table of Contents

• Features
• Installation
• Usage
  • Model Loading
  • Inference
  • Training
• Dependencies
• License

---

Features

Warning

DeepSeek-RS is not stable and is currently unfinished. Expect bugs, missing features, and incomplete functionality. Use at your own risk!

• Transformer-based architecture - Supports large-scale models.
• Optimized tensor computations - Uses tch-rs (bindings to libtorch).
• 1:1 Python conversion - Structure is directly mapped from DeepSeek’s Python implementation.
• Efficient inference - Uses GPU acceleration via tch::Tensor.
• Training pipeline - Custom model training with dataset integration.

---

Installation

Prerequisites

• Rust (latest stable)
• Cargo package manager
• Libtorch (must be installed separately)

Clone the Repository

git clone https://github.com/rustyspottedcatt/deepseek-rs.git
cd deepseek-rs

Build the Project

cargo build --release

---

Usage

Model Loading

use tch::nn::{self, Module, OptimizerConfig};
use tch::{Device, Tensor};

let vs = nn::VarStore::new(Device::cuda_if_available());
let model = nn::seq()
    .add(nn::linear(vs.root() / "fc1", 768, 3072, Default::default()))
    .add_fn(|xs| xs.relu())
    .add(nn::linear(vs.root() / "fc2", 3072, 768, Default::default()));

println!("Model successfully loaded.");

---

Inference

use tch::Tensor;

let input_tensor = Tensor::randn(&[1, 768], (tch::Kind::Float, tch::Device::cuda_if_available()));
let output_tensor = model.forward(&input_tensor);

println!("Inference output: {:?}", output_tensor);

---

Training

use tch::nn::{self, OptimizerConfig};
use tch::Tensor;

let mut opt = nn::Adam::default().build(&vs, 1e-3).unwrap();
let loss = output_tensor.mse_loss(&target_tensor, tch::Reduction::Mean);

opt.backward_step(&loss);
println!("Training step completed.");

---

Dependencies

[dependencies]
tch = "0.19"
ndarray = "0.15"
serde = { version = "1.0", features = ["derive"] }
bincode = "1.3"
rayon = "1.7"

---

License

Distributed under the GNU AGPLv3 license.