# LoRA training GPU cloud: why a US host is often fine

URL: https://cloudj.net/blog/lora-training-gpu-usa
Published: 2026-11-12 · Updated: 2026-11-12

> LoRA training on a GPU cloud is a batch job, so location matters little. Learn to sort US hosts by price, size VRAM, upload data once and pull results back.

LoRA training GPU cloud jobs have a shape that makes location almost irrelevant: you upload a dataset, the GPU works for an hour or a night with nobody watching, and you download a small adapter file at the end. That means you can choose a host for price, VRAM and reliability rather than distance. On Cloudj you can search broadly, across the United States or with no country set at all, and pick on those terms. This article explains how to do that well, whether you train image LoRAs for SDXL or Flux, or language LoRAs for an open-weight LLM.

## Why location barely matters for LoRA training

Compare a live chatbot with a LoRA run:

| | Live chatbot | LoRA training |
|---|---|---|
| Who waits on the GPU? | A user, on every message | Nobody, until the end |
| Data in | Small prompts, constantly | One dataset upload |
| Data out | Small responses, constantly | One adapter file and some samples |
| What distance adds | Delay on every request | A few extra minutes of transfer, once |

For training, a host on another continent costs you a little more transfer time at the start and end. It does not slow the training itself. [GPU near users vs cheapest GPU](https://cloudj.net/blog/gpu-near-users-vs-cheapest-gpu) explains this trade-off in general, and [batch inference in the cheapest region](https://cloudj.net/blog/batch-inference-cheapest-region) applies the same logic to inference jobs.

## Why search the US

The United States is a large country, and a US search is a reasonable broad starting point when you do not care about location. That is not a promise about stock: availability changes during the day everywhere. If your team is in North America, `US, CA` keeps interactive steps such as checking samples a little more comfortable. If you are elsewhere, you can equally leave the Country field empty and look at every country. [GPU cloud in the USA](https://cloudj.net/blog/gpu-cloud-usa) covers US searches in more depth.

## Choose VRAM for your LoRA

Rough, general guidance; actual needs depend on resolution, batch size, rank and the trainer you use:

- **SD 1.5 LoRAs:** a 24 GB card is more than enough; an [RTX 3090](https://cloudj.net/gpu-cloud/rtx-3090) is a sensible choice.
- **SDXL LoRAs:** 24 GB is comfortable; an [RTX 4090](https://cloudj.net/gpu-cloud/rtx-4090) trains faster than a 3090 in most setups.
- **Flux LoRAs:** memory-hungry; 24 GB can work with memory-saving options, and a 32 GB [RTX 5090](https://cloudj.net/gpu-cloud/rtx-5090) gives headroom.
- **LLM LoRA or QLoRA on a 7B to 14B model:** a 24 GB card usually works with 4-bit QLoRA; larger models or longer sequences need more memory or several GPUs in one machine.

All three cards are consumer GPUs available at the starting account level. Live "from" prices are on [GPU cloud](https://cloudj.net/gpu-cloud).

## Sort by price, then filter for safety

The console sorts results by total price, lowest first. The cheapest line is not automatically the right one for a multi-hour run. Use this order:

1. Open the [deploy page with the RTX 4090 preselected](https://app.cloudj.net/deploy/gpu?model=rtx-4090), or choose another model.
2. Enter `US` (or `US, CA`, or nothing) in the Country field.
3. Set a minimum reliability you are comfortable with.
4. Check **max duration**: the machine must stay available well beyond your expected run.
5. Check **download speed**: base models are large, and a slow link wastes paid GPU minutes.
6. Check **bandwidth price** if your dataset is big.
7. Pick the cheapest line that passes all of the above.

The console shows the exact price before you deploy, and that price is locked for the instance.

## Upload once, train, pull results

### Prepare locally

Clean and caption your images, or format your text dataset, before you rent anything. GPU time is billed per minute while the instance runs, so preparation on your own machine is free time.

### Upload the dataset once

```bash
rsync -avP -e "ssh -p PORT" ./dataset/ root@HOST:/workspace/dataset/
```

Download base models directly on the GPU host rather than uploading them from home; hosts usually have much better download speeds than a home upload. [Moving models and data between GPU regions](https://cloudj.net/blog/move-models-data-between-gpu-regions) covers larger transfers.

### Train inside tmux

```bash
ssh -p PORT root@HOST
tmux new -s train
accelerate launch train_lora.py --config config.toml
```

Replace the script and config with your trainer of choice. For language models, the [PEFT documentation](https://huggingface.co/docs/peft) explains LoRA settings. Run inside `tmux` so the job survives a dropped SSH connection.

### Pull results and destroy

```bash
rsync -avP -e "ssh -p PORT" root@HOST:/workspace/output/ ./lora-output/
```

Then destroy the instance. Destroy deletes the instance and disk and stops charges. Stopping instead keeps the disk at storage price but releases the GPU, and restarting may wait until a GPU is free.

## Protect long runs

- Save checkpoints every so often, and copy the latest one off the machine if the run is long.
- Note the "available until" date before you start.
- If the host goes offline, billing pauses, but the job does not move to another machine automatically. A recent checkpoint lets you resume on a new host.

## When to keep training close to home

Choose a nearby country instead of the US if your dataset is very large and your own upload path to the US is slow, if your rules require data to stay in certain countries, or if you plan many short interactive experiments. Otherwise, price, VRAM and reliability should drive the decision.

## FAQ

### Does training a LoRA in the US make it worse for users in Asia?

No. The adapter file is the same wherever it was trained. Location only affects where you serve it.

### Which GPU should I start with for SDXL LoRAs?

An RTX 4090 is a comfortable choice. An RTX 3090 also works with 24 GB.

### Should I always choose the cheapest machine?

Pick the cheapest one that also passes your reliability, max duration and download speed checks.

### Can I stop the instance overnight and resume?

You can, but stopping releases the GPU and restarting may wait. For short pauses, keep it running; for long ones, save checkpoints and redeploy later.

## Next step

Open the [RTX 4090 deploy page](https://app.cloudj.net/deploy/gpu?model=rtx-4090), enter `US`, apply your reliability and max duration filters, and start your LoRA on the cheapest machine that passes them.