Frozen vs Unfrozen Weights: What Adapters Actually Are in LoRA and QLoRA

2026-07-26 · Source: AI on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, medium

Summary

The article clarifies the distinction between "frozen" and "unfrozen" weights in neural networks, particularly within LoRA and QLoRA fine-tuning methods. Frozen weights remain constant during training, while unfrozen weights are updated by the optimizer. In full fine-tuning, all parameters, such as the 32 billion in a 32B LLM, are unfrozen, leading to high memory costs due to optimizer state tracking. LoRA, or Low-Rank Adaptation, addresses this by freezing the entire original base model and injecting small, new "adapter" matrices, typically into attention projections. Only these new adapter weights are unfrozen and trainable, significantly reducing memory and preventing catastrophic forgetting. QLoRA extends LoRA by storing the frozen base model in 4-bit precision instead of 16-bit, further optimizing memory without altering the frozen/unfrozen mechanism. This distinction directly impacts memory cost, model behavior drift, and portability.

Key takeaway

For Machine Learning Engineers optimizing LLM fine-tuning, understanding the frozen vs. unfrozen weight distinction is crucial. If you are struggling with VRAM limitations or catastrophic forgetting, you should prioritize LoRA or QLoRA. These methods allow you to train small, portable adapter files while preserving the base model's core capabilities and significantly reducing memory footprint compared to full fine-tuning.

Key insights

Frozen weights are static during training; unfrozen weights are updated, defining fine-tuning approaches like LoRA and QLoRA.

Principles

• Frozen weights protect original model capabilities.
• Optimizer state drives memory cost for unfrozen weights.
• Low-rank adaptation approximates full weight updates efficiently.

Method

LoRA injects small, new adapter matrices alongside frozen base model layers, training only these adapters. QLoRA adds 4-bit quantization to the frozen base.

In practice

• Use LoRA/QLoRA to reduce VRAM for fine-tuning.
• Employ adapters for smaller, portable model updates.
• Prevent catastrophic forgetting with frozen base models.

Topics

• LoRA
• QLoRA
• Fine-tuning
• Neural Network Weights
• Parameter Efficient Fine-tuning
• Large Language Models

Best for: Machine Learning Engineer, NLP Engineer, AI Student

Related on AIssential

• LoRA & QLoRA Mastery: The Beginner-to-Advanced Guide to Efficient LLM Fine-Tuning — LoRA and QLoRA enable efficient LLM fine-tuning by adapting a small fraction of parameters, drastically reducing memory and compute.
• Super Weights in LLMs and the Failure of Selective Training — Parameter importance does not imply trainability; effective LLM fine-tuning requires structured, layer-wide updates.
• How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning — LoRA adapter capacity and memorization are measurable in bits, revealing placement-dependent storage and privacy implications.
• From Full Fine-Tuning to LoRA and QLoRA: A Beginner-Friendly Guide to Adapting Large Language… — Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA and QLoRA enable cost-effective LLM adaptation by minimizing trainable parameters and base model memory.
• The Evolution of LoRA: 15+ Variants You Should Know — LoRA variants offer diverse strategies for efficient, stable, and adaptive AI model fine-tuning.
• Fine-Tuning a 7B LLM Required 4 A100s. LoRA Did It on One GPU. Here Is the Math Behind Why. — Fine-tuning LLMs involves low-dimensional weight updates, enabling efficient adaptation via LoRA and QLoRA.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI on Medium.