Chronicals v0.1.0 - Initial Release

Chronicals v0.1.0

High-Performance LLM Fine-Tuning Framework achieving 3.51x speedup over Unsloth.

Key Features

• Fused Triton Kernels: RMSNorm, SwiGLU, Cross-Entropy, QK-RoPE
• Cut Cross-Entropy: 37x memory reduction for loss computation
• LoRA+ Optimizer: Differential learning rates for adapter training
• Sequence Packing: Efficient batch utilization
• FlashAttention Integration: Memory-efficient attention

Installation

# Core installation
pip install chronicals

# With all GPU kernels
pip install chronicals[all]

# For Google Colab
pip install chronicals[kernels,training]

Quick Start

from chronicals import ChronicalsTrainer, ChronicalsConfig

config = ChronicalsConfig(use_flash_attention=True, use_fused_kernels=True)
trainer = ChronicalsTrainer(model, tokenizer, config)
trainer.train(dataset)

Benchmarks

Model	Unsloth	Chronicals	Speedup
Llama-3.2-1B	1.00x	3.51x	3.51x
Llama-3.2-3B	1.00x	2.89x	2.89x
Qwen2.5-7B	1.00x	2.45x	2.45x

Documentation

See README.md for full documentation.