nexlora 0.2.1

Next-generation LLM fine-tuning engine with self-optimizing training

NexLoRA

Merged package combining FastLoRA v4.0 and NexusTrain v1.0 into a single installable package.

Authors:

• FastLoRA: FastLoRA Contributors
• NexusTrain: Ömür Bera Işık

Installation

pip install nexlora

Usage

FastLoRA Usage

from nexlora import fastlora

model, tokenizer = fastlora.FastLoRA("meta-llama/Llama-3.2-3B").load()
# ... training ...

NexusTrain Usage

from nexlora import nexustrain

nt = nexustrain.NexusTrain(model, tokenizer, nexustrain.NexusConfig())
trainer = nt.patch_trainer(trainer)
nt.engage()
# ... training ...

Environment Check

nexlora-check

Features

FastLoRA (v4.0)

• LoRA / QLoRA fine-tuning
• 2-bit quantization (independent of bitsandbytes)
• OOM recovery with auto-resume
• Custom Triton kernels (RMSNorm, SwiGLU, RoPE)
• Mixture of Experts (MoE) support
• Optuna hyperparameter tuning
• CPU/NVMe offloading (ZeRO-Infinity style)
• Smart VRAM guard

NexusTrain (v1.0)

• CrystalCore™ - Runtime kernel crystallization
• MorphicMemory™ - Adaptive memory morphism
• SpectraOptimizer™ - Frequency domain optimizer
• ResonanceScheduler™ - Self-tuning learning rate
• ChromaticPrecision™ - Layer-wise dynamic precision
• GradientHarmonics™ - Wavelet-based gradient processing
• NeuralProfiler™ - LSTM-based training predictor
• CrystalPipeline™ - Self-reconfiguring training pipeline
• ZeroWaste™ - Unnecessary computation elimination

License

MIT License - See LICENSE file for details.