Unsloth Fft

Overview

Full Fine-Tuning (FFT) in Unsloth allows for 100% exact weight updates, bypassing the low-rank approximations of LoRA. By utilizing Unsloth's optimized gradient checkpointing, FFT can fit significantly larger batch sizes while ensuring total model modification.

When to Use

• When performing base model pre-training or continued pre-training on large datasets.
• When model-wide behaviors need modification that adapters (LoRA) cannot fully capture.
• When sufficient VRAM is available to handle full model gradients.

Decision Tree

1. Do you need to modify 100% of the model weights?
   • Yes: Proceed with FFT.
   • No: Use [[unsloth-lora]].
2. Is VRAM limited (e.g., < 24GB for a 7B model)?
   • Yes: Enable use_gradient_checkpointing = 'unsloth' and adamw_8bit.
   • No: Use standard BF16 and high batch sizes.

Workflows

Initializing Full Fine-tuning

1. Load the model using FastLanguageModel.from_pretrained with load_in_4bit=False and load_in_8bit=False.
2. Pass full_finetuning=True in the initialization call to unlock all weight updates.
3. Apply the 'unsloth' gradient checkpointing via FastLanguageModel.get_peft_model(model, use_gradient_checkpointing='unsloth').

FFT Memory Management

1. Set per_device_train_batch_size to 1 to accommodate full weight gradients.
2. Increase gradient_accumulation_steps to simulate effective batch sizes of 16-32.
3. Use the adamw_8bit optimizer to reduce memory consumption by optimizer states.

Non-Obvious Insights

• Unsloth's gradient checkpointing implementation is critical for FFT, as it uses 30% less VRAM and allows for 2x larger batch sizes compared to standard Hugging Face implementations.
• Full fine-tuning is the preferred method for base model pre-training where the objective is knowledge injection rather than task instruction.
• FFT in Unsloth is strictly 100% exact; there is no numerical drift compared to standard training, only efficiency gains.

Evidence

• "To enable full fine-tuning (FFT), set full_finetuning = True." Source
• "use_gradient_checkpointing = 'unsloth' uses 30% less VRAM, fits 2x larger batch sizes!" Source

Scripts

• scripts/unsloth-fft_tool.py: Script to initialize a full fine-tuning session with Unsloth.
• scripts/unsloth-fft_tool.js: Node.js utility to generate FFT training parameters.

Dependencies

• unsloth
• torch
• transformers

References

• [[references/README.md]]