LoRA adapters for specialist models

Who this page is for

Read this when the method is already “LoRA” and the open questions are which file you store, how it attaches, and whether three products can share one box. Typical readers are a platform owner who will keep the frozen base, an SRE who will be paged when the wrong adapter attaches, and a specialist who still thinks “the model” is one object. This is not a PEFT notebook. It will not publish a rank-versus-quality table or a latency league for multi-adapter serving.

If labelled pairs do not exist yet, read supervised fine-tuning. If the frozen base does not fit in memory, read QLoRA. If someone wants to touch every weight, read full fine-tune versus PEFT. This URL owns adapter handover and multi-adapter operations.

What the paper and PEFT actually gave you

The LoRA paper keeps the pretrained network frozen and learns low-rank update matrices, so the number of trainable parameters drops sharply relative to a full update. Rank and alpha are capacity knobs on a named base for a named task. They are not a prestige ladder. We will not invent an llm.build winner among ranks.

Hugging Face PEFT documents the same idea in engineering form. Its quantisation guide is where the 4-bit sibling lives, including the original method’s large-model-on-one-GPU example. That sentence is their write-up. It is not a capacity plan for your sequence length. If memory is why you opened this tab, stay here only if the base already fits; otherwise go to the QLoRA page.

The adapter pack you actually hand over

A useful LoRA handover is the adapter files, the exact base checkpoint those files were trained against, how to load or merge them, and notes from unused evaluation work. Operators still need the base weights and a licence that allows the stack. An adapter without a revision hash is a souvenir. A merged blob without a recipe is a souvenir with a larger file size.

Merging produces a single-task image that is simpler to serve and harder to reuse. Keeping adapters separate keeps the frozen base as a shared asset. Choose from the operating plan, not from a tutorial default. If the next team cannot load what you shipped, you did not hand over a specialist model. You handed over a research leftover.

Multi-adapter serving is an operations object

One frozen base plus many adapters is attractive for several products on one box. It is also a working-set problem. How many adapters are hot. What the server actually allows. Who owns the incident when a claims request silently picks last week’s retail file. Do not promise “unlimited specialists on one GPU” in a sales deck.

Write the routing rule. Write the memory budget for hot adapters and the cold-load path. Write who can publish a new adapter into the serving set. Those sentences are the product. A notebook that shows two adapters attaching once is not an operating plan.

• Named base revision shared by every adapter in the set.
• Routing key from product to adapter, with a default refuse.
• Hot-set size and who may publish a new adapter.
• Incident owner when attach or merge fails.

LoRA is not a smaller legal surface

An adapter derived from Llama or Gemma materials is still in that licence’s world when you distribute it. File size does not create a new legal species. Apache or MIT bases are simpler because the notices are simpler, not because adapters magically become yours. Write the licence into the same pack as the adapter files.

This page will not reprint Community Licence clauses; read the Llama family page. It will not reprint Gemma Terms; read Gemma. Quantised storage and low rank do not move those duties. They also do not move inference. Prompts still go wherever you send them. Logs still sit wherever the server writes them.

What changes in the run

The foundation stays frozen. You train low-rank matrices and write adapter files, not a second copy of the foundation weights. GPU memory is lower than a full update because you are not storing optimiser state for every foundation parameter. Serving either loads those adapters on the named base or merges them once for a single-task image. A run that overfits the train slice still fails unused work. Low rank is not regularisation for bad labels.

The licence printed on the base is the licence you still have. Rank, alpha, and which modules you attach still decide the artefact. Write those settings into the same note as the files. An adapter without settings is a blog souvenir. We will not paste a training script here. This URL owns handover and operations, not a copy-paste recipe.

Insurance: one base, many document-family adapters

Retail: seasonal catalogue adapters

Merge versus separate is an operating decision

Merge when there is one product, one adapter, and a serving stack that only wants a single set of weights. Separate when more than one specialist must live on the same frozen base, or when you expect to roll an adapter back without touching the others. Neither choice is a quality upgrade. Both still need unused evaluation notes for each adapter you claim is ready.

A merged image is easier to copy into an air-gapped box and easier to lose track of. Write the recipe that produced it: base revision, adapter identifier, merge date. A separate-adapter setup is easier to audit and easier to mis-route. Write the attach test you run after every publish. Those are handover sentences, not optional extras from a blog.

Rank is capacity, not a quality setting

People treat a larger rank as a more serious project. Rank is how much update capacity you buy on a frozen base. If unused work fails, inspect labels and the task before you treat a larger rank as a prestige upgrade. A small rank on a clean schema beats a large rank on a disagreeing spreadsheet.

Write rank, alpha, and modules as facts of the run. If a later operator cannot reproduce the attach, the rank number in a slide is decoration. PEFT is the usual place those settings live in engineering form. We will not paste them into a recipe here.

Document-family adapters still share one incident rota

The insurance shape — one frozen checkpoint, a file per wording family — only works if publish rights and incident ownership are shared too. A wording team that can push an adapter the platform team cannot roll back is not multi-adapter ops. It is shadow IT with extra matrices. Write the publish window, the rollback test, and who is awake when attach fails at two in the morning.

The retail seasonal file has the same requirement. A peak-intake adapter that nobody is allowed to delete in January is how last year’s schema haunts the evergreen path. Retirement is an operations action. Rank does not retire files. People do, on a written rule.

Handover for a LoRA run

The pack is the adapter, the exact base revision, load or merge recipe, licence line from the family page, routing notes if more than one adapter exists, and unused-evaluation notes per specialist. Include rank and alpha as facts of the run, not as a quality claim. Include who may publish the next adapter.

If you cannot name the base revision, you do not have a LoRA handover. If you cannot say how the adapter attaches in the environment you actually operate, you have a training souvenir. We will ask for the pack before we treat the files as an engagement artefact.

Include a one-line attach test: load this revision, attach this adapter, send this fixture request, expect this schema. Without the fixture, the next operator will invent a demo. Demos are how the wrong adapter goes live. The fixture is boring on purpose. Boring is the handover.

When LoRA is the wrong conversation

If there are no labelled pairs, LoRA is a trainer looking for a job. If the base does not fit, you wanted QLoRA’s memory method, not a smaller rank. If adapters already failed a written bar and the labels are clean, you may be on the full-update page. If the only goal is a vendor UI style, you may be on the closed-API comparison. Do not start a LoRA run to avoid those sentences.

LoRA is also the wrong conversation when nobody will operate the frozen base. An adapter file on a laptop next to an unowned GPU booking is not a specialist model. It is a souvenir with a PEFT filename. Write the operator first, or keep the work exploratory and stop claiming handover.

Sources

1. Hu et al., LoRA: Low-Rank Adaptation of Large Language Models (Accessed September 2026). LoRA freezes the pretrained weights and trains low-rank update matrices, reducing the number of trainable parameters for adaptation.
2. Hugging Face PEFT, Quantization (Accessed September 2026). PEFT documents QLoRA as 4-bit quantisation of the frozen base plus LoRA adapters, and notes the pattern can fine-tune a 65B model on a single 48GB GPU in the original method write-up.

Questions we hear first