unsloth-rs-core
| Crates.io | unsloth-rs-core |
| lib.rs | unsloth-rs-core |
| version | 0.1.0 |
| created_at | 2026-01-17 16:00:43.291336+00 |
| updated_at | 2026-01-17 16:00:43.291336+00 |
| description | Fast, efficient implementation of Unsloth optimization kernels and training loop in Rust. |
| homepage | |
| repository | https://github.com/Moaisus-admin/unsloth-rs |
| max_upload_size | |
| id | 2050715 |
| size | 189,229 |
Moaisus (Moaisus-admin)
documentation
README
Unsloth Rust 🦥🦀
Unsloth, but faster. Written in Rust.
This library implements high-performance optimization kernels (Flash Attention, RoPE, RMSNorm) and training loops for LLMs using Rust and Candle. It provides a drop-in replacement API for the original Unsloth library.
Features
- Zero-Dependency Python Install: Pre-built wheels for Linux, macOS, and Windows.
- High Performance: Custom CUDA kernels and Rust-based orchestration.
- Native Compatibility: Includes
unsloth_nativeAPI that mimicsunsloth. - LoRA Support: Efficient LoRA injection and training.
- GGUF Export: Merge and save directly to Safetensors for GGUF conversion.
Installation
Python (Recommended)
pip install unsloth_rs
Rust
Add to your Cargo.toml:
[dependencies]
unsloth_rs = "0.1.0"
Usage
from unsloth_native import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
"TinyLlama/TinyLlama-1.1B-Chat-v1.0",
load_in_4bit=True
)
# ... training code ...
License
Apache-2.0