unsloth-candle-cuda 2026.4.1

High-performance LLM fine-tuning library built with Candle and Rust.

🚀 Unsloth-Candle

High-performance LLM fine-tuning library built with Rust 🦀 and Candle.

---

Unsloth-Candle brings the blazing fast performance of Unsloth to the Candle ecosystem. By leveraging optimized Rust kernels and efficient memory management, it enables 2x faster training and 70% less memory usage compared to standard implementations.

✨ Core Advantages

• Zero Learning Curve: 1:1 API compatibility with Unsloth's Python interface.
• Hardware Optimized: Native support for CUDA, Metal (Apple Silicon), and AVX/Neon (CPU).
• Memory Efficient: Native 4-bit NF4 quantization and gradient checkpointing.
• Unified Support: One engine for Llama 3.2, Mistral, Qwen 2.5, DeepSeek-V3, and more.

📦 Installation

Via Pip (Recommended)

pip install unsloth-candle

Build from Source

git clone https://github.com/unslothai/unsloth-candle.git
cd unsloth-candle
pip install -e .

To enable GPU acceleration:

• CUDA: pip install -e . --features cuda
• Metal: pip install -e . --features metal

🛠 Usage

1. Load Model & Tokenizer

from unsloth_candle import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Llama-3.2-1B-Instruct",
    max_seq_length = 2048,
    load_in_4bit = True,
)

2. Apply LoRA/DoRA

model = FastLanguageModel.get_peft_model(
    model,
    r = 16,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_alpha = 16,
    use_gradient_checkpointing = True,
    use_dora = False, # Set to True for DoRA
)

3. Fine-tuning with SFTTrainer

from unsloth_candle import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = 2048,
    args = SFTConfig(
        max_steps = 60,
        learning_rate = 2e-4,
        logging_steps = 1,
    ),
)
trainer.train()

4. Save & Export

# Save as merged HF weights
model.save_pretrained_merged("output_hf", tokenizer)

# Save as GGUF (for Ollama/llama.cpp)
model.save_pretrained_gguf("output_gguf", tokenizer, quantization_type="q4_k_m")

🗺️ Model Catalog

Model	Architecture	4-bit	LoRA	DoRA
Llama 3.2	LlamaForCausalLM	✅	✅	✅
Mistral Nemo	MistralForCausalLM	✅	✅	✅
Qwen 2.5	Qwen2ForCausalLM	✅	✅	✅
DeepSeek V3	DeepSeekV3 (MLA)	✅	✅	✅
Gemma 3	Gemma3 (GeGLU)	✅	✅	✅
Phi 4	Phi4	✅	✅	✅

📜 License

Licensed under the Apache License, Version 2.0.

---

Built with 💖 by the Unsloth Community and Antigravity.