36 lines
1.4 KiB
Markdown
36 lines
1.4 KiB
Markdown
# Unsloth in a docker container
|
|
|
|
❗ Very important ❗
|
|
|
|
The complete docker image is designed for CUDA 12.1!
|
|
There must be a compatible version on the host such as <span style="color: red;">CUDA 12.2</span> and <span style="color: skyblue;">docker</span> has to be installed.
|
|
|
|
Run `nvidia-smi` on the Host to see if you are using the right Conda version.
|
|
|
|
|
|
|
|
Build the container with:
|
|
```shell
|
|
docker build -t cuda-unsloth-jupyter .
|
|
```
|
|
|
|
And mount your directory to the ```/app``` path when starting.
|
|
|
|
Setting default password to: `123`
|
|
```shell
|
|
docker run --rm --gpus=all -v path_to_files:/app -e JUPYTER_TOKEN='123' -p 8088:8088 -d cuda-unsloth-jupyter
|
|
```
|
|
|
|
```--rm```: This flag automatically deletes the container after it exits, freeing up disk space and keeping the environment clean. Add the `-it` flag to run the container in interactive mode, giving you direct access to its shell. Use `-d` to detach the container and run it in the background. Combining `-itd` allows you to interact with the container initially, then detach, leaving it running in the background.
|
|
|
|
|
|
Currently tested with:
|
|
|
|
| GPU Model | Memory (GB) |
|
|
|-------------|-------------|
|
|
| RTX 3090 Ti | 24 |
|
|
|
|
Please note:<br>
|
|
If you are using a different graphics card or a different Pytorch version, then you must adapt the Dockerfile accordingly. See:
|
|
[docs.unsloth.ai](https://docs.unsloth.ai/get-started/installation/pip-install)
|