v4.0.0

lyogavin/airllmv4.0.0Sep 5, 2026by lyogavin

AI Summary

This release introduces a memory-efficient training architecture known as Streamed LoRA, enabling the training of large models like Qwen3.8-Flash-Next on consumer-grade hardware with limited VRAM. It provides updated installation steps and specific usage examples for local dataset training.

Key Highlights

Streamed LoRA architecture moves frozen weights to disk to free up GPU memory.
Qwen3.8-Flash-Next can now be trained on an RTX 3060 Ti with under 6GB VRAM.
Qwen3.8-27B training is optimized to approximately 2GB VRAM at sequence length 512.
Requires the in-tree `qwen4_exp` directory from the latest `transformers` main branch.

New Features

Streamed LoRA implementation for reduced memory footprint.
Support for training Qwen3.8-Flash-Next on low-end GPUs.
Support for training Qwen3.8-27B with minimal VRAM requirements.
Updated example script for training Qwen models.

Full Release Notes

Streamed LoRA: frozen weights stay on disk, adapters stay on GPU. **Qwen3.8-Flash-Next** trains under **6GB** (RTX 3060 Ti); **Qwen3.8-27B** trains in **~2GB** at seq 512.

```bash
pip install -U airllm
pip install git+https://github.com/huggingface/transformers.git
```

Point `--data` at a JSONL file (`{"text": "..."}` or `{"prompt": "...", "completion": "..."}`):

```bash
python air_llm/examples/train_qwen38_flash_next_lora.py   --data my_data.jsonl   --seq-len 512   --epochs 1   --save-adapter qwen38-flash-next-lora.pt
```

Flash-Next still needs in-tree `qwen4_exp` (GitHub `transformers` main today).