Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Use Unsloth branding in remaining user-facing text (#11254)

Michael Hanbe95e417b947
diff --git a/.github/ISSUE_TEMPLATE/install-blocked-windows.yml b/.github/ISSUE_TEMPLATE/install-blocked-windows.yml
index 77362090632..afaba86b31f 100644
--- a/.github/ISSUE_TEMPLATE/install-blocked-windows.yml
+++ b/.github/ISSUE_TEMPLATE/install-blocked-windows.yml
@@ -1,5 +1,5 @@
name: "Windows: antivirus or security software blocked the installer"
-description: The installer, updater or Studio was blocked, quarantined or deleted by security software on Windows
+description: The installer, updater or Unsloth Studio was blocked, quarantined or deleted by security software on Windows
title: "[Blocked] "
# Every one of these must exist in the repository. GitHub does not error on an unknown label, it
# silently drops it -- so a typo here costs the triage label the form exists to attach, and nothing
diff --git a/README.md b/README.md
index d734506088c..ceacf0d7139 100644
--- a/README.md
+++ b/README.md
@@ -167,9 +167,9 @@ docker run -d --name unsloth --gpus all --ipc=host \
-v unsloth-studio:/opt/unsloth-studio \
unsloth/unsloth && docker logs -f unsloth
```
-The log ends with your links and a generated JupyterLab password, plus a generated Studio password on the first run against a new `unsloth-studio` volume; change that one on first sign-in or Studio stops after an hour. A reused volume keeps the password already stored on it, and `docker exec unsloth unsloth studio reset-password --username unsloth` mints a new one and prints it. Ctrl-C stops following the log, not the container; `docker rm -f unsloth` deletes it. The Hugging Face cache keeps your models and the `unsloth-studio` volume keeps your accounts, chats and trained models, both across `docker rm`; a volume from an older image is migrated on first start, its old code kept under `.unsloth-studio-legacy/`. Those ports publish on every interface: on a cloud host add `-e UNSLOTH_STUDIO_SECURE=1`, drop `-p 8000:8000` and bind JupyterLab to `-p 127.0.0.1:8888:8888`, or bind both to `127.0.0.1` and use an SSH tunnel. Tags (`unsloth/unsloth:core` for notebooks only), GPU support and options: [Docker Hub](https://hub.docker.com/r/unsloth/unsloth).
+The log ends with your links and a generated JupyterLab password, plus a generated Unsloth Studio password on the first run against a new `unsloth-studio` volume; change that one on first sign-in or Unsloth Studio stops after an hour. A reused volume keeps the password already stored on it, and `docker exec unsloth unsloth studio reset-password --username unsloth` mints a new one and prints it. Ctrl-C stops following the log, not the container; `docker rm -f unsloth` deletes it. The Hugging Face cache keeps your models and the `unsloth-studio` volume keeps your accounts, chats and trained models, both across `docker rm`; a volume from an older image is migrated on first start, its old code kept under `.unsloth-studio-legacy/`. Those ports publish on every interface: on a cloud host add `-e UNSLOTH_STUDIO_SECURE=1`, drop `-p 8000:8000` and bind JupyterLab to `-p 127.0.0.1:8888:8888`, or bind both to `127.0.0.1` and use an SSH tunnel. Tags (`unsloth/unsloth:core` for notebooks only), GPU support and options: [Docker Hub](https://hub.docker.com/r/unsloth/unsloth).
-On AMD there is a separate image, [`unsloth/unsloth-rocm`](https://hub.docker.com/r/unsloth/unsloth-rocm), with the run command and the supported cards on its [Docker Hub page](https://hub.docker.com/r/unsloth/unsloth-rocm). It carries the training stack only, so there is no Studio or JupyterLab in it, and it needs native Linux: WSL exposes `/dev/dxg` rather than the `/dev/kfd` that ROCm needs.
+On AMD there is a separate image, [`unsloth/unsloth-rocm`](https://hub.docker.com/r/unsloth/unsloth-rocm), with the run command and the supported cards on its [Docker Hub page](https://hub.docker.com/r/unsloth/unsloth-rocm). It carries the training stack only, so there is no Unsloth Studio or JupyterLab in it, and it needs native Linux: WSL exposes `/dev/dxg` rather than the `/dev/kfd` that ROCm needs.
#### Remote HTTPS & LAN Access
Server-side tools are on by default - so **be careful**! Keep your password safe, or use `--disable-tools` when exposing Unsloth.
@@ -331,7 +331,7 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_SKIP_AUTOSTART=1 sh
$env:UNSLOTH_SKIP_AUTOSTART=1; irm https://unsloth.ai/install.ps1 | iex
```
-Keep the install-time package cache under the Studio directory instead of reusing an existing uv cache. Downloads are slower the first time, and an explicit `UV_CACHE_DIR` still wins over this:
+Keep the install-time package cache under the Unsloth Studio directory instead of reusing an existing uv cache. Downloads are slower the first time, and an explicit `UV_CACHE_DIR` still wins over this:
```bash
curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_ISOLATE_UV_CACHE=1 sh
```
diff --git a/docker/DOCKERHUB-ROCM.md b/docker/DOCKERHUB-ROCM.md
index 3a0a8eaa056..0d9339b8881 100644
--- a/docker/DOCKERHUB-ROCM.md
+++ b/docker/DOCKERHUB-ROCM.md
@@ -110,4 +110,4 @@ The build's own record is in the image: `cat /etc/unsloth-rocm-build` reports th
- Source and issues: [github.com/unslothai/unsloth](https://github.com/unslothai/unsloth)
- Docker files: [`docker/`](https://github.com/unslothai/unsloth/tree/main/docker)
- Documentation: [docs.unsloth.ai](https://docs.unsloth.ai)
-- Licences: the image is labelled `Apache-2.0 AND AGPL-3.0-only`. [Apache-2.0](https://github.com/unslothai/unsloth/blob/main/LICENSE) covers the repository and the training stack. The [AGPL-3.0](https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0) half is there for the one AGPL file this image ships, `/workspace/smoke_test_rocm.py`, which carries an `AGPL-3.0-only` header; Studio itself is not in this image.
+- Licences: the image is labelled `Apache-2.0 AND AGPL-3.0-only`. [Apache-2.0](https://github.com/unslothai/unsloth/blob/main/LICENSE) covers the repository and the training stack. The [AGPL-3.0](https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0) half is there for the one AGPL file this image ships, `/workspace/smoke_test_rocm.py`, which carries an `AGPL-3.0-only` header; Unsloth Studio itself is not in this image.
diff --git a/docker/DOCKERHUB.md b/docker/DOCKERHUB.md
index c904cd2e489..d98cec86b9d 100644
--- a/docker/DOCKERHUB.md
+++ b/docker/DOCKERHUB.md
@@ -1,6 +1,6 @@
# Unsloth Docker Image
-Fine-tune and run LLMs, vision, audio and diffusion models with no setup. Every image carries the training stack (PyTorch 2.11 with CUDA 12.8, [Unsloth](https://github.com/unslothai/unsloth), unsloth-zoo, bitsandbytes, TRL, PEFT, plus xformers on `linux/amd64`), JupyterLab with the [Unsloth notebooks](https://github.com/unslothai/notebooks), and prebuilt llama.cpp. The `latest` image adds whisper.cpp for Studio's speech-to-text.
+Fine-tune and run LLMs, vision, audio and diffusion models with no setup. Every image carries the training stack (PyTorch 2.11 with CUDA 12.8, [Unsloth](https://github.com/unslothai/unsloth), unsloth-zoo, bitsandbytes, TRL, PEFT, plus xformers on `linux/amd64`), JupyterLab with the [Unsloth notebooks](https://github.com/unslothai/notebooks), and prebuilt llama.cpp. The `latest` image adds whisper.cpp for Unsloth Studio's speech-to-text.
Source: [`docker/`](https://github.com/unslothai/unsloth/tree/main/docker). Guide: [docs.unsloth.ai](https://docs.unsloth.ai/get-started/install/docker).
@@ -17,15 +17,15 @@ docker run -d --name unsloth --gpus all --ipc=host \
unsloth/unsloth && docker logs -f unsloth
```
-That starts the container and follows its startup, which ends after about a minute with your links and a generated JupyterLab password. On the first run against a new `unsloth-studio` volume it generates a Studio password too; change that one on first sign-in or Studio stops after an hour. Reusing an existing volume keeps the Studio password already stored on it, which the log says rather than printing one. Ctrl-C stops following the log, not the container.
+That starts the container and follows its startup, which ends after about a minute with your links and a generated JupyterLab password. On the first run against a new `unsloth-studio` volume it generates an Unsloth Studio password too; change that one on first sign-in or Unsloth Studio stops after an hour. Reusing an existing volume keeps the Unsloth Studio password already stored on it, which the log says rather than printing one. Ctrl-C stops following the log, not the container.
Those ports publish on every interface, which is fine on a laptop and not on a cloud host. Three ways to close that, in order of least work:
-- `-e UNSLOTH_STUDIO_SECURE=1`, with `-p 8000:8000` dropped and 8888 changed to `-p 127.0.0.1:8888:8888`: Studio is served only over a Cloudflare HTTPS link, printed in the log, and binds to loopback inside the container so no raw port exists. It fails closed, so no tunnel means no link rather than serving in the clear. Move JupyterLab too, or it stays published on every interface over plain HTTP, which is a shell on your GPU box with its password in the clear. On a new volume the public page will not hand out the generated first-boot password, on purpose; read it from the log, where it is printed for exactly this case, or set one yourself with `UNSLOTH_STUDIO_PASSWORD`. `UNSLOTH_STUDIO_CLOUDFLARE=1` keeps the local port as well.
+- `-e UNSLOTH_STUDIO_SECURE=1`, with `-p 8000:8000` dropped and 8888 changed to `-p 127.0.0.1:8888:8888`: Unsloth Studio is served only over a Cloudflare HTTPS link, printed in the log, and binds to loopback inside the container so no raw port exists. It fails closed, so no tunnel means no link rather than serving in the clear. Move JupyterLab too, or it stays published on every interface over plain HTTP, which is a shell on your GPU box with its password in the clear. On a new volume the public page will not hand out the generated first-boot password, on purpose; read it from the log, where it is printed for exactly this case, or set one yourself with `UNSLOTH_STUDIO_PASSWORD`. `UNSLOTH_STUDIO_CLOUDFLARE=1` keeps the local port as well.
- Publish to `127.0.0.1` and tunnel: `-p 127.0.0.1:8000:8000 -p 127.0.0.1:8888:8888`, then `ssh -L 8000:localhost:8000 -L 8888:localhost:8888 user@your-host`.
- Keep the ports and set real passwords with `-e UNSLOTH_STUDIO_PASSWORD=...` and `-e JUPYTER_PASSWORD=...`.
-Studio reports on start whether the port answered from the public internet, so you are told either way.
+Unsloth Studio reports on start whether the port answered from the public internet, so you are told either way.
### Before that, on a new machine
@@ -62,7 +62,7 @@ docker rm -f unsloth # stop and delete the container
docker ps -a # find it again, running or not
```
-`docker rm -f` deletes the container, not your work: models stay in the Hugging Face cache, your files in the directory you mounted, and Studio's accounts and chats on the `unsloth-studio` volume, so the next container with the same `-v unsloth-studio:/opt/unsloth-studio` resumes where this one left off, password included. Only what was written inside the container is lost. `docker volume rm unsloth-studio` discards the Studio data too, and cannot be undone.
+`docker rm -f` deletes the container, not your work: models stay in the Hugging Face cache, your files in the directory you mounted, and Unsloth Studio's accounts and chats on the `unsloth-studio` volume, so the next container with the same `-v unsloth-studio:/opt/unsloth-studio` resumes where this one left off, password included. Only what was written inside the container is lost. `docker volume rm unsloth-studio` discards the Unsloth Studio data too, and cannot be undone.
These take a container, not an image, so `docker stop unsloth/unsloth` fails with "No such container". Use the `NAMES` or `CONTAINER ID` column of `docker ps -a`; without `--name`, Docker assigns a random one.
@@ -88,7 +88,7 @@ docker run --rm --gpus all --ipc=host -v "$PWD":/workspace/host \
### CPU-only hosts
-`latest` starts without a GPU on its own. Studio chat with GGUF models, JupyterLab and the GGUF tooling work; training does not. `core` refuses to start without a GPU unless you opt in:
+`latest` starts without a GPU on its own. Unsloth Studio chat with GGUF models, JupyterLab and the GGUF tooling work; training does not. `core` refuses to start without a GPU unless you opt in:
```bash
docker run -d -p 8000:8000 -p 8888:8888 unsloth/unsloth
@@ -100,7 +100,7 @@ docker run --rm -e UNSLOTH_ALLOW_CPU=1 unsloth/unsloth:core python -c "import un
| Tag | Contents | Use it for |
|---|---|---|
| `latest`, `studio` | Unsloth Studio web UI + JupyterLab + notebooks + key-only SSH | Most users. Train and chat in the browser. |
-| `core` | Training stack + JupyterLab + notebooks, no Studio | Notebooks, scripts, CI, slimmer pulls. |
+| `core` | Training stack + JupyterLab + notebooks, no Unsloth Studio | Notebooks, scripts, CI, slimmer pulls. |
| `nightly-<YYYY.MM.DD>`, `core-nightly-<YYYY.MM.DD>` | The same two images, one immutable pin per daily rebuild, kept 60 days | Reproducible runs. |
| `<version>`, `core-<version>` | Release builds | Pin a release. |
@@ -110,7 +110,7 @@ docker run --rm -e UNSLOTH_ALLOW_CPU=1 unsloth/unsloth:core python -c "import un
Turing (T4, RTX 20) through Blackwell (B200, GB200, RTX 50, RTX PRO 6000) run precompiled SASS: `sm_70 sm_75 sm_80 sm_86 sm_90 sm_100 sm_120` on `linux/amd64`, `sm_80 sm_90 sm_90a sm_100 sm_100a sm_120 sm_120a` on `linux/arm64`. SASS is forward-compatible within a major version, so Ada runs the `sm_86` binaries, B300 and GB300 the `sm_100` ones, and GB10 (DGX Spark) the `sm_120` ones. A source build inside the image compiles for `7.5;8.0;8.6;8.9;9.0;10.0;12.0+PTX`.
-Drivers: 570.26 or newer for CUDA 12.8 on every GPU, 580 or newer for B300, GB300 and GB10. On `linux/arm64` the bundled llama.cpp is a CUDA 13 build, because upstream ships no CUDA 12 one there, so training works from 570 but GGUF export and Studio chat need 580.
+Drivers: 570.26 or newer for CUDA 12.8 on every GPU, 580 or newer for B300, GB300 and GB10. On `linux/arm64` the bundled llama.cpp is a CUDA 13 build, because upstream ships no CUDA 12 one there, so training works from 570 but GGUF export and Unsloth Studio chat need 580.
Turing has no bfloat16; Unsloth falls back to float16 there. These images are CUDA only; for AMD use [`unsloth/unsloth-rocm`](https://hub.docker.com/r/unsloth/unsloth-rocm), a ROCm build of the training stack.
@@ -126,13 +126,13 @@ Turing has no bfloat16; Unsloth falls back to float16 there. These images are CU
| Variable | Effect |
|---|---|
-| `UNSLOTH_STUDIO_PASSWORD` | Initial Studio password for user `unsloth`; ignored once one is stored. Unset: generated, printed in the logs, and Studio stops after an hour unless changed (`UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0` disables). |
+| `UNSLOTH_STUDIO_PASSWORD` | Initial Unsloth Studio password for user `unsloth`; ignored once one is stored. Unset: generated, printed in the logs, and Unsloth Studio stops after an hour unless changed (`UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0` disables). |
| `JUPYTER_PASSWORD` | JupyterLab password, read by the `latest` launcher. Unset: generated and printed in the logs. On `core` you start JupyterLab yourself. |
-| `JUPYTER_PORT` | JupyterLab's port inside the container. Default `8888`. Studio's port (`UNSLOTH_STUDIO_PORT`) is refused. |
-| `UNSLOTH_STUDIO_PORT` | Studio's port inside the container, for `--network host` when 8000 is taken. Default `8000`. |
+| `JUPYTER_PORT` | JupyterLab's port inside the container. Default `8888`. Unsloth Studio's port (`UNSLOTH_STUDIO_PORT`) is refused. |
+| `UNSLOTH_STUDIO_PORT` | Unsloth Studio's port inside the container, for `--network host` when 8000 is taken. Default `8000`. |
| `SSH_KEY` or `PUBLIC_KEY` | OpenSSH public key for root login; enables sshd on port 22. Password login is never enabled. |
| `UNSLOTH_ALLOW_CPU=1` | Allow starting without a GPU (`latest` already does). Dropped when a GPU is visible, where it would turn off Unsloth's training patches. |
-| `UNSLOTH_STUDIO_SECURE=1` | Serve Studio over a Cloudflare HTTPS link only, bound to loopback inside the container so no raw port is published. Fails closed if the tunnel does not come up. |
+| `UNSLOTH_STUDIO_SECURE=1` | Serve Unsloth Studio over a Cloudflare HTTPS link only, bound to loopback inside the container so no raw port is published. Fails closed if the tunnel does not come up. |
| `UNSLOTH_STUDIO_CLOUDFLARE=1` | The same HTTPS link, with the local port still served. Not valid together with `UNSLOTH_STUDIO_SECURE=1`. |
| `UNSLOTH_JUPYTER_CLOUDFLARE=1` | Publish JupyterLab through a Cloudflare quick tunnel and print the URL. |
| `UNSLOTH_SKIP_NOTEBOOK_REFRESH=1` | Do not refresh the notebooks from GitHub on start; the copy baked into the image is still used. |
@@ -150,29 +150,31 @@ The working directory is `/workspace`. Mount what you want to keep:
|---|---|
| `/workspace/host` | Your files. Mount your project directory here. |
| `/workspace/.cache/huggingface` | Model downloads. Mount your host HF cache to reuse it. |
-| `/opt/unsloth-studio` | Studio's accounts, chats, outputs, exports and runs (`latest`). Use a named volume: without one, `docker rm` loses them. |
+| `/opt/unsloth-studio` | Unsloth Studio's accounts, chats, outputs, exports and runs (`latest`). Use a named volume: without one, `docker rm` loses them. |
| `/workspace/.cache/triton` | Compiled kernels. Optional, speeds up restarts. |
| `/workspace/unsloth-notebooks` | The synced notebooks. Your edits are kept across refreshes. |
| `/workspace/Unsloth Notebooks` | The same notebooks grouped by topic, rebuilt on each start. |
The container runs as root by default. On `core`, `--user <uid>:<gid>` is supported and keeps files on your mounts owned by you. `latest` runs its services as root and does not start under `--user`.
-### Studio data and Studio code
+<a id="studio-data-and-studio-code"></a>
-Studio's code ships in the image under `/opt/unsloth-studio-app` and is linked into `/opt/unsloth-studio` at every start, so the volume holds only your data and every image runs its own code.
+### Unsloth Studio data and code
+
+Unsloth Studio's code ships in the image under `/opt/unsloth-studio-app` and is linked into `/opt/unsloth-studio` at every start, so the volume holds only your data and every image runs its own code.
- A volume from before that split holds an old image's code. The first start of a newer image moves it aside to `/opt/unsloth-studio/.unsloth-studio-legacy/` and links the new code in; nothing is deleted. Delete it once the new image works (`docker exec <c> rm -rf /opt/unsloth-studio/.unsloth-studio-legacy`), along with the 9 GB uv cache at `/opt/unsloth-studio/cache/uv`, or start with `-e UNSLOTH_STUDIO_KEEP_LEGACY=0` to skip keeping it. Going back to a pre-split image needs those entries moved back, or `docker run --rm -v unsloth-studio:/opt/unsloth-studio --entrypoint unsloth-studio-home <current image> --restore`.
-- Use a named volume, not a bind mount of a Windows or macOS host directory: the Studio home needs symlinks, and Docker Desktop's file sharing may refuse to create them, stopping the container at start. A Linux directory, including one inside a WSL 2 distribution, works.
+- Use a named volume, not a bind mount of a Windows or macOS host directory: the Unsloth Studio home needs symlinks, and Docker Desktop's file sharing may refuse to create them, stopping the container at start. A Linux directory, including one inside a WSL 2 distribution, works.
## Stopping while training
-On `docker stop` and `docker restart`, Studio stops a running training job at the next step and saves a checkpoint before it exits, so the run can be resumed from the Training page. Docker only waits 10 seconds by default, which is not enough for a large model, so give it the budget:
+On `docker stop` and `docker restart`, Unsloth Studio stops a running training job at the next step and saves a checkpoint before it exits, so the run can be resumed from the Training page. Docker only waits 10 seconds by default, which is not enough for a large model, so give it the budget:
```bash
docker stop -t 150 <container>
```
-or `stop_grace_period: 150s` in Compose. `docker/run.sh` sets this for you. `UNSLOTH_STUDIO_SHUTDOWN_STOP_TIMEOUT_S` (default 120) is the time Studio itself waits for the save; the container's own limit is 30 seconds above it.
+or `stop_grace_period: 150s` in Compose. `docker/run.sh` sets this for you. `UNSLOTH_STUDIO_SHUTDOWN_STOP_TIMEOUT_S` (default 120) is the time Unsloth Studio itself waits for the save; the container's own limit is 30 seconds above it.
A save is also bounded by the training stop watchdog, which force-terminates a worker that has not finished after `UNSLOTH_STUDIO_TRAINING_STOP_TIMEOUT_S` (default 600). Raising the shutdown budget past that does nothing on its own; raise both, and `docker/run.sh` forwards both.
@@ -182,7 +184,7 @@ A host shutdown is not covered: the daemon stops every container under its own `
On the `latest` image:
-- `unsloth-studio-update` upgrades Studio and Unsloth in place, leaving the torch and CUDA stack pinned. It survives `docker restart`, not `docker rm`.
+- `unsloth-studio-update` upgrades Unsloth Studio and Unsloth in place, leaving the torch and CUDA stack pinned. It survives `docker restart`, not `docker rm`.
- `unsloth-llama-update` fetches the newest prebuilt llama.cpp.
- `unsloth-jupyter-tunnel --force` opens a Cloudflare quick tunnel to JupyterLab. Without `--force` or `UNSLOTH_JUPYTER_CLOUDFLARE=1` it prints that it is disabled and exits.
diff --git a/docker/Dockerfile.studio b/docker/Dockerfile.studio
index c53c193e010..672b9bc354e 100644
--- a/docker/Dockerfile.studio
+++ b/docker/Dockerfile.studio
@@ -129,7 +129,7 @@ RUN set -eux \
# CUDA family) before the dedup symlinks their CUDA libs. Compare to the base's
# own torch (no hardcoded version); metadata only (QEMU arm64 can't import torch).
&& BASE_TORCH="$(/opt/unsloth-venv/bin/python -c "from importlib.metadata import version; print(version('torch'))")" \
- && "${UNSLOTH_STUDIO_HOME}/unsloth_studio/bin/python" -c "import sys; from importlib.metadata import version; assert sys.version_info[:2] == (3, 12), 'Studio venv python %d.%d is not 3.12 (UNSLOTH_PYTHON pin ignored) -- CUDA dedup below depends on it' % sys.version_info[:2]; v = version('torch'); assert v == '${BASE_TORCH}', 'Studio venv torch ' + v + ' does not match base venv torch ${BASE_TORCH} (CUDA dedup would link mismatched libs)'; print('Studio venv python %d.%d torch' % sys.version_info[:2], v, '== base', '${BASE_TORCH}')" \
+ && "${UNSLOTH_STUDIO_HOME}/unsloth_studio/bin/python" -c "import sys; from importlib.metadata import version; assert sys.version_info[:2] == (3, 12), 'Unsloth Studio venv python %d.%d is not 3.12 (UNSLOTH_PYTHON pin ignored) -- CUDA dedup below depends on it' % sys.version_info[:2]; v = version('torch'); assert v == '${BASE_TORCH}', 'Unsloth Studio venv torch ' + v + ' does not match base venv torch ${BASE_TORCH} (CUDA dedup would link mismatched libs)'; print('Unsloth Studio venv python %d.%d torch' % sys.version_info[:2], v, '== base', '${BASE_TORCH}')" \
# setup.sh may relink llama-quantize into build/bin; prove it still resolves its
# libraries. Content check, not rc: --help exits nonzero but prints usage.
&& { "${UNSLOTH_STUDIO_HOME}/llama.cpp/llama-quantize" --help 2>&1 || true; } | grep -q "usage" \
@@ -141,7 +141,7 @@ RUN set -eux \
exit 1; \
}; \
done \
- && echo "OK: Studio kept the base image's CUDA llama.cpp bundle" \
+ && echo "OK: Unsloth Studio kept the base image's CUDA llama.cpp bundle" \
&& rm -rf "${UNSLOTH_STUDIO_HOME}/src/.git" \
"${UNSLOTH_STUDIO_HOME}/src/studio/frontend/node_modules" \
"${UV_CACHE_DIR:-${UNSLOTH_STUDIO_HOME}/cache/uv}" \
@@ -186,7 +186,7 @@ RUN set -eux \
&& test ! -e "${UNSLOTH_STUDIO_HOME}/cache/uv" \
# through bash: the exec bit is set further down, with the other bin scripts
&& bash /usr/local/bin/unsloth-studio-home \
- && "${UNSLOTH_STUDIO_HOME}/unsloth_studio/bin/python" -c "import os, sys, studio; print('Studio venv', sys.prefix, 'imports studio from', studio.__file__); assert os.path.realpath(studio.__file__).startswith(os.path.realpath('${UNSLOTH_STUDIO_HOME}/src') + '/'), studio.__file__" \
+ && "${UNSLOTH_STUDIO_HOME}/unsloth_studio/bin/python" -c "import os, sys, studio; print('Unsloth Studio venv', sys.prefix, 'imports studio from', studio.__file__); assert os.path.realpath(studio.__file__).startswith(os.path.realpath('${UNSLOTH_STUDIO_HOME}/src') + '/'), studio.__file__" \
&& ls -la "${UNSLOTH_STUDIO_HOME}"
COPY supervisord.conf /etc/supervisor/supervisord.conf
diff --git a/docker/entrypoint.sh b/docker/entrypoint.sh
index 2a48144fb6c..2fd50203605 100755
--- a/docker/entrypoint.sh
+++ b/docker/entrypoint.sh
@@ -11,7 +11,7 @@ set -euo pipefail
# CUDA tool selection reads the venv. Fatal on failure (a half-linked home); no-op on the base image.
if [[ -x /usr/local/bin/unsloth-studio-home ]]; then
/usr/local/bin/unsloth-studio-home || {
- echo "ERROR: could not link Studio's code into ${UNSLOTH_STUDIO_HOME:-/opt/unsloth-studio}; see the messages above" >&2
+ echo "ERROR: could not link Unsloth Studio's code into ${UNSLOTH_STUDIO_HOME:-/opt/unsloth-studio}; see the messages above" >&2
exit 1
}
fi
@@ -100,7 +100,7 @@ fi
if [[ "${has_gpu}" == "0" && "${allow_cpu}" == "1" ]]; then
warn "UNSLOTH_ALLOW_CPU=1 and no GPU visible -- continuing on CPU."
- warn "CPU mode covers Jupyter, GGUF tooling and llama.cpp (GGUF) Studio chat."
+ warn "CPU mode covers Jupyter, GGUF tooling and llama.cpp (GGUF) Unsloth Studio chat."
warn "Training and loading Unsloth models (FastLanguageModel) still require an NVIDIA GPU."
sync_notebooks
exec "$@"
@@ -133,7 +133,7 @@ Likely causes (in order of frequency):
5. This host has no NVIDIA GPU at all (Docker Desktop on macOS, Windows
without WSL2 GPU support, CPU-only Linux). Training and loading Unsloth
- models need a GPU, but Jupyter, GGUF tooling and llama.cpp (GGUF) Studio
+ models need a GPU, but Jupyter, GGUF tooling and llama.cpp (GGUF) Unsloth Studio
chat work on CPU:
docker run -e UNSLOTH_ALLOW_CPU=1 ...
@@ -214,7 +214,7 @@ if [ "$(uname -m)" = "aarch64" ]; then
if [ "$_drv_major" -lt 580 ]; then
echo "WARNING: this arm64 image bakes a CUDA 13 llama.cpp (upstream ships no CUDA 12 arm64 build)." >&2
echo " Host driver $_drv is < 580, which cannot load CUDA 13 binaries:" >&2
- echo " training (torch cu128) works, but GGUF export / Studio chat will fail" >&2
+ echo " training (torch cu128) works, but GGUF export / Unsloth Studio chat will fail" >&2
echo " until the host driver is upgraded to >= 580." >&2
fi
;;
diff --git a/docker/jupyter/install_sloth_stickers.py b/docker/jupyter/install_sloth_stickers.py
index 97bfb0bad3f..a2a241228bb 100644
--- a/docker/jupyter/install_sloth_stickers.py
+++ b/docker/jupyter/install_sloth_stickers.py
@@ -47,7 +47,7 @@ CURATED = [
def main() -> int:
parser = argparse.ArgumentParser(description = __doc__)
- parser.add_argument("--src", required = True, help = "Studio 'Sloth emojis' dir")
+ parser.add_argument("--src", required = True, help = "Unsloth 'Sloth emojis' dir")
parser.add_argument("--dest", required = True, help = "output dir (static/sloth)")
args = parser.parse_args()
diff --git a/docker/studio_home.sh b/docker/studio_home.sh
index 4d88292e5dc..42f2b5768e5 100755
--- a/docker/studio_home.sh
+++ b/docker/studio_home.sh
@@ -32,7 +32,7 @@ if [ "${1:-}" = "--restore" ]; then
if [ ! -d "$legacy" ]; then
# No kept copy (first used after the split, or deleted): a pre-split image finds its
# own tree hidden under the mount and links to nothing, so give it real copies.
- [ -d "$APP" ] || die "no $legacy to restore and no $APP on this image: run --restore under an image that has the Studio code in $APP (the one that last ran this volume), it copies that code into the home for an older image"
+ [ -d "$APP" ] || die "no $legacy to restore and no $APP on this image: run --restore under an image that has the Unsloth Studio code in $APP (the one that last ran this volume), it copies that code into the home for an older image"
copied=()
for target in "$HOME_DIR"/*; do
[ -L "$target" ] || continue
@@ -172,11 +172,11 @@ done
if [ -s "$APP/.src-update.rollback" ]; then
updater="${UNSLOTH_STUDIO_UPDATER:-/usr/local/bin/unsloth-studio-update}"
if [ -x "$updater" ]; then
- log "an update was interrupted (record at $APP/.src-update.rollback); putting the previous install back before Studio starts"
+ log "an update was interrupted (record at $APP/.src-update.rollback); putting the previous install back before Unsloth Studio starts"
if UNSLOTH_STUDIO_HOME="$HOME_DIR" "$updater" --recover; then
log "the previous install is back"
else
- log "WARNING: the previous install could not be put back (see above); Studio may fail to start"
+ log "WARNING: the previous install could not be put back (see above); Unsloth Studio may fail to start"
log " finish it by hand: docker exec <container> unsloth-studio-update --recover"
fi
else
diff --git a/docker/studio_launch.sh b/docker/studio_launch.sh
index 8db9fefd0da..61fe9a41313 100644
--- a/docker/studio_launch.sh
+++ b/docker/studio_launch.sh
@@ -161,7 +161,7 @@ elif [[ -n "${UNSLOTH_STUDIO_PASSWORD:-}" ]]; then
STUDIO_NOTE="user unsloth, password from UNSLOTH_STUDIO_PASSWORD env"
UNSLOTH_STUDIO_PASSWORD_STATE=initial
else
- STUDIO_NOTE="user unsloth, generated password printed below once Studio is up"
+ STUDIO_NOTE="user unsloth, generated password printed below once Unsloth Studio is up"
UNSLOTH_STUDIO_PASSWORD_STATE=generated
fi
unset UNSLOTH_STUDIO_PASSWORD
diff --git a/docker/studio_password.sh b/docker/studio_password.sh
index ea96a1dcc76..7fb8578151e 100755
--- a/docker/studio_password.sh
+++ b/docker/studio_password.sh
@@ -59,7 +59,7 @@ case "$STATE" in
# replaces. Wait for the row, and read the file after it.
while [[ ! -s "$FILE" ]] || ! unsloth-studio-run --initialized; do
if (( $(date +%s) >= deadline )); then
- STUDIO_LINE="the first-boot password did not appear in ${WAIT}s; check the studio log above"
+ STUDIO_LINE="the first-boot password did not appear in ${WAIT}s; check the Unsloth Studio log above"
break
fi
sleep 1
@@ -69,7 +69,7 @@ case "$STATE" in
password="$(tr -d '\r\n' < "$FILE")"
note=""
if (( TIMEOUT > 0 )); then
- note=" (change it on first sign-in: Studio stops after $(duration "$TIMEOUT") with the default password; UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0 disables)"
+ note=" (change it on first sign-in: Unsloth Studio stops after $(duration "$TIMEOUT") with the default password; UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0 disables)"
fi
STUDIO_LINE="username: unsloth password: ${password}${note}"
fi ;;
@@ -113,7 +113,7 @@ fi
rule="$(printf '=%.0s' $(seq 1 72))"
printf '%s\n' "${_g}${rule}${_r}"
printf ' %s\n' "${_g}${title}${_r}"
-printf ' Studio %s %s\n' "${_b}http://localhost:${UNSLOTH_STUDIO_PORT:-8000}${_r}" "${_y}${studio_text}${_r}"
+printf ' Unsloth %s %s\n' "${_b}http://localhost:${UNSLOTH_STUDIO_PORT:-8000}${_r}" "${_y}${studio_text}${_r}"
printf ' JupyterLab %s %s\n' "${_b}http://localhost:${JUPYTER_PORT}${_r}" "${_y}${jupyter_text}${_r}"
printf ' %s\n' "Ports are the container's: use the host side of your -p flags, or an SSH tunnel to a remote host."
printf '%s\n' "${_g}${rule}${_r}"
diff --git a/docker/unsloth_studio_update.sh b/docker/unsloth_studio_update.sh
index 547294f212c..6f6f49079e7 100755
--- a/docker/unsloth_studio_update.sh
+++ b/docker/unsloth_studio_update.sh
@@ -65,7 +65,7 @@ if [ -z "$PY" ] && [ -L "$STUDIO_HOME/bin/unsloth" ]; then
venv_bin="$(dirname "$(readlink -f "$STUDIO_HOME/bin/unsloth")")"
[ -x "$venv_bin/python" ] && PY="$venv_bin/python"
fi
-[ -n "$PY" ] || { echo "unsloth-studio-update: could not find the Studio venv under $STUDIO_HOME" >&2; exit 1; }
+[ -n "$PY" ] || { echo "unsloth-studio-update: could not find the Unsloth Studio venv under $STUDIO_HOME" >&2; exit 1; }
# The editable tree may sit behind a symlink from the home into the image's copy: work
# on the real path, but keep pip pointed at the home path (SRC_INSTALL) so the recorded
@@ -134,7 +134,7 @@ find_supctl() {
[ -x "$SUPCTL" ] && { "$SUPCTL" status studio >/dev/null 2>&1 || _st=$?; } || _st=4
}
-log "Studio venv: $PY"
+log "Unsloth Studio venv: $PY"
log "before: unsloth $(version_of)"
# One updater at a time: two would take each other's trees for leftovers. Held on an
@@ -189,7 +189,7 @@ if [ -s "$KEEP_ROLLBACK" ]; then
fi
fi
else
- log "--no-restart: Studio may still be running the interrupted update; restart it to load the restored install"
+ log "--no-restart: Unsloth Studio may still be running the interrupted update; restart it to load the restored install"
fi
else
echo "unsloth-studio-update: could not put the previous packages back (see the CRITICAL lines above); fix the cause and run this again. Nothing else was changed." >&2
@@ -428,7 +428,7 @@ if [ -n "$REF" ]; then
fi
fi
if [ ! -d "$SRC" ]; then
- echo "unsloth-studio-update: no source tree at $SRC (a wheel install has none, and a broken src link points at nothing); --ref needs the image's Studio checkout. Use a plain update instead; nothing was changed." >&2
+ echo "unsloth-studio-update: no source tree at $SRC (a wheel install has none, and a broken src link points at nothing); --ref needs the image's Unsloth Studio checkout. Use a plain update instead; nothing was changed." >&2
exit 1
fi
log "installing from git: unsloth @${REF}, unsloth-zoo @${_zoo_ref}"
@@ -492,9 +492,9 @@ log "after: unsloth $(version_of)"
# parks supervisord's program in FATAL, so check first and put the old one back.
if ! studio_tree_ok; then
if restore && studio_tree_ok; then
- log "the update could not start Studio, so the previous install was restored and Studio was not restarted."
+ log "the update could not start Unsloth Studio, so the previous install was restored and Unsloth Studio was not restarted."
else
- log "CRITICAL: the previous install could not be restored cleanly; Studio will fail to start until this is fixed."
+ log "CRITICAL: the previous install could not be restored cleanly; Unsloth Studio will fail to start until this is fixed."
fi
log "If a new dependency is missing, re-run with --with-deps."
exit 1
@@ -566,21 +566,21 @@ if [ "$RESTART" = "1" ]; then
done
if [ "$_up" = "1" ]; then
commit_update
- log "Studio is answering on port ${_port}"
+ log "Unsloth Studio is answering on port ${_port}"
else
- back_out "Studio did not answer on port ${_port} within ${_wait}s (see docker logs)"
+ back_out "Unsloth Studio did not answer on port ${_port} within ${_wait}s (see docker logs)"
fi
fi
else
# nothing here can start Studio, so the import and frontend checks above are
# the whole validation
commit_update
- log "supervisor not managing 'studio' here; restart Studio yourself"
+ log "supervisor not managing 'studio' here; restart Unsloth Studio yourself"
log " (e.g. 'docker restart <container>')"
fi
else
commit_update
- log "--no-restart: restart Studio to load the update"
+ log "--no-restart: restart Unsloth Studio to load the update"
log " docker exec <container> supervisorctl restart studio"
fi
diff --git a/install.ps1 b/install.ps1
index 92074425c5f..479b4c146fa 100644
--- a/install.ps1
+++ b/install.ps1
@@ -3211,7 +3211,7 @@ exit 1
if (-not (Test-StudioUvNoCache)) {
Write-StudioUvCacheMarker -StudioRoot $StudioRoot -Cache $studioCache
}
- step "uv cache" "forced Studio cache isolation ($studioCache); already-cached packages may download again" "Yellow"
+ step "uv cache" "forced Unsloth Studio cache isolation ($studioCache); already-cached packages may download again" "Yellow"
return
}
@@ -3338,19 +3338,19 @@ exit 1
}
"studio" {
if ($chosenCache) {
- step "uv cache" "reusing this install's Studio cache ($selectedCache)"
+ step "uv cache" "reusing this install's Unsloth Studio cache ($selectedCache)"
# Never about the directory we are falling back TO: the Studio cache is itself
# a candidate now, so it can be the one refused, and naming it claims a fallback
# that did not happen.
} elseif ($scanBlocked -and -not [string]::IsNullOrWhiteSpace([string]$blockedCache) -and $blockedCache -ne $selectedCache) {
- step "uv cache" "using new Studio-owned cache ($selectedCache); part of $blockedCache could not be read, so cached packages may download again" "Yellow"
+ step "uv cache" "using new Unsloth Studio-owned cache ($selectedCache); part of $blockedCache could not be read, so cached packages may download again" "Yellow"
} elseif ($scanBlocked -and [string]::IsNullOrWhiteSpace([string]$blockedCache)) {
- step "uv cache" "using new Studio-owned cache ($selectedCache); the existing uv cache could not be inspected, so cached packages may download again" "Yellow"
+ step "uv cache" "using new Unsloth Studio-owned cache ($selectedCache); the existing uv cache could not be inspected, so cached packages may download again" "Yellow"
# Warm and still here means the write probe refused it.
} elseif ($warnCache -and $warnCache -ne $selectedCache) {
- step "uv cache" "using new Studio-owned cache ($selectedCache); $warnCache is populated but not writable, so cached packages may download again" "Yellow"
+ step "uv cache" "using new Unsloth Studio-owned cache ($selectedCache); $warnCache is populated but not writable, so cached packages may download again" "Yellow"
} else {
- step "uv cache" "using new Studio-owned cache ($selectedCache)"
+ step "uv cache" "using new Unsloth Studio-owned cache ($selectedCache)"
}
}
}
diff --git a/install.sh b/install.sh
index cc3e7bfeb97..989bfe79138 100755
--- a/install.sh
+++ b/install.sh
@@ -771,7 +771,7 @@ _configure_uv_cache() {
_UV_CACHE_MODE=isolated
export UV_CACHE_DIR
_uv_no_cache_requested || _record_uv_cache_choice
- step "uv cache" "forced Studio cache isolation ($UV_CACHE_DIR); already-cached packages may download again" "$C_WARN"
+ step "uv cache" "forced Unsloth Studio cache isolation ($UV_CACHE_DIR); already-cached packages may download again" "$C_WARN"
return 0
fi
@@ -953,17 +953,17 @@ _configure_uv_cache() {
;;
studio)
if [ -n "$_uv_chosen_cache" ]; then
- step "uv cache" "reusing this install's Studio cache ($UV_CACHE_DIR)"
+ step "uv cache" "reusing this install's Unsloth Studio cache ($UV_CACHE_DIR)"
# Never about the directory we are falling back TO: the Studio cache is itself a
# candidate now, so it can be the one refused, and naming it claims a fallback
# that did not happen.
elif [ "$_uv_scan_blocked" = true ] && [ "$_uv_blocked_cache" != "$UV_CACHE_DIR" ]; then
- step "uv cache" "using new Studio-owned cache ($UV_CACHE_DIR); part of $_uv_blocked_cache could not be read, so cached packages may download again" "$C_WARN"
+ step "uv cache" "using new Unsloth Studio-owned cache ($UV_CACHE_DIR); part of $_uv_blocked_cache could not be read, so cached packages may download again" "$C_WARN"
# Warm and still here means the write probe refused it.
elif [ -n "$_uv_warn_cache" ] && [ "$_uv_warn_cache" != "$UV_CACHE_DIR" ]; then
- step "uv cache" "using new Studio-owned cache ($UV_CACHE_DIR); $_uv_warn_cache is populated but not writable, so cached packages may download again" "$C_WARN"
+ step "uv cache" "using new Unsloth Studio-owned cache ($UV_CACHE_DIR); $_uv_warn_cache is populated but not writable, so cached packages may download again" "$C_WARN"
else
- step "uv cache" "using new Studio-owned cache ($UV_CACHE_DIR)"
+ step "uv cache" "using new Unsloth Studio-owned cache ($UV_CACHE_DIR)"
fi
;;
esac
diff --git a/scripts/p2p_integrity_probe.py b/scripts/p2p_integrity_probe.py
index e5e1e7de0a6..a8ffc943058 100644
--- a/scripts/p2p_integrity_probe.py
+++ b/scripts/p2p_integrity_probe.py
@@ -77,7 +77,7 @@ def _print_nvml_agreement() -> None:
sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "studio" / "backend"))
from core.inference.llama_cpp import LlamaCppBackend as backend
except Exception as e: # noqa: BLE001 -- diagnostic, never fatal
- print(f"could not import the Studio backend to cross-check NVML: {e}\n")
+ print(f"could not import the Unsloth backend to cross-check NVML: {e}\n")
return
try:
nvml = backend._probe_nvml_nvlink_topology()
@@ -87,7 +87,7 @@ def _print_nvml_agreement() -> None:
return
print("=== NVML vs nvidia-smi topo -m ===")
if nvml is None:
- print("NVML gave no verdict here; Studio would use `nvidia-smi topo -m`.\n")
+ print("NVML gave no verdict here; Unsloth would use `nvidia-smi topo -m`.\n")
return
if topo is None:
print("`nvidia-smi topo -m` gave no verdict; nothing to compare against.\n")
@@ -109,7 +109,7 @@ def _print_nvml_agreement() -> None:
print(
f"{sorted(topo_yes - nvml_yes)} are NVLink to topo -m but not to NVML.\n"
"This is the conservative direction: by default NVML answers first, so "
- "Studio\nwould not enable P2P for them. Setting "
+ "Unsloth\nwould not enable P2P for them. Setting "
"UNSLOTH_P2P_TOPO_CROSSCHECK=1 makes topo -m\nwin any disagreement, which "
"would enable them instead."
)
diff --git a/studio/MCP.md b/studio/MCP.md
index d082b767588..d9e8238e481 100644
--- a/studio/MCP.md
+++ b/studio/MCP.md
@@ -2,11 +2,11 @@
## Connect Blender MCP
-Blender MCP is **disabled by default**. Studio downloads a pinned, checksum-verified
+Blender MCP is **disabled by default**. Unsloth Studio downloads a pinned, checksum-verified
runtime on first enable/test and caches it on the backend machine. No commands,
Git or pip installs are needed. Subsequent starts use the cache without internet.
The Blender add-on is installed separately. No MCP archive or source is shipped
-in Studio's Python package or desktop build.
+in Unsloth Studio's Python package or desktop build.
1. Open **Manage MCP servers → Blender**, approve the execution warning and choose
**Enable Blender MCP**. Use a tool-capable model with MCP enabled for the chat.
@@ -22,8 +22,8 @@ A green dot means Blender is connected; amber means only the MCP server is conne
Setup help stays in the same Blender entry. Advanced settings configure the bridge
port (default `9876`) and optional Blender executable. This port is not an HTTP URL.
-The bridge uses loopback on the **Studio backend machine**, not a remote browser.
-Studio does not install or launch Blender during setup. Approved tools can run
+The bridge uses loopback on the **Unsloth Studio backend machine**, not a remote browser.
+Unsloth Studio does not install or launch Blender during setup. Approved tools can run
Python, write files and launch background Blender. Existing tool permissions apply;
external model providers receive tool results. Keep the unauthenticated bridge local.
@@ -33,9 +33,11 @@ https://projects.blender.org/lab/blender_mcp (GPL-3.0-or-later).
The pinned revision and SHA-256 are in `backend/integrations/blender/runtime.py`.
Downloads are staged and verified before activation; failures leave the server
disabled and can be retried with **Enable Blender MCP**. Merely opening the dialog
-or launching Studio does not download anything.
+or launching Unsloth Studio does not download anything.
-## Studio's own MCP server
+<a id="studios-own-mcp-server"></a>
+
+## Unsloth Studio's own MCP server
Unsloth can expose a local MCP server so an MCP client can inspect models and
GPU state, validate recipes, start or stop training, inspect recipe output, and
diff --git a/studio/NVLINK_P2P.md b/studio/NVLINK_P2P.md
index e444ac33e36..e05f2ff4d28 100644
--- a/studio/NVLINK_P2P.md
+++ b/studio/NVLINK_P2P.md
@@ -51,7 +51,7 @@ A product name cannot tell you whether the box has a bridge.
asking it again is not verification. vLLM reached the same conclusion and performs a real
data-integrity check (`can_actually_p2p`, from vllm#2728).
-Studio reads the topology itself and requires a confirmed NVLink between every selected pair.
+Unsloth Studio reads the topology itself and requires a confirmed NVLink between every selected pair.
NVLink traffic does not traverse the PCIe root complex, so the IOMMU fault class above
cannot apply to it. The probe fails closed: a missing `nvidia-smi`, a non-zero exit, a
timeout, an unparsable matrix, or a device mask that cannot be mapped to PCI indices all
@@ -144,6 +144,6 @@ bool use_peer_access = getenv("GGML_CUDA_P2P") != nullptr;
```
So `GGML_CUDA_P2P=0` turns peer access **on**. The variable must be unset entirely.
-Studio deletes it from the llama-server environment when its inherited value reads as off
+Unsloth Studio deletes it from the llama-server environment when its inherited value reads as off
(`0`, `false`, `no`, `off`, empty), on every backend, so the intuitive spelling of the
opt-out does what the user meant. Set it to `1` and it is honoured as a deliberate request.
diff --git a/studio/PDF_OCR.md b/studio/PDF_OCR.md
index c8d2fbc598f..ac2c26edff1 100644
--- a/studio/PDF_OCR.md
+++ b/studio/PDF_OCR.md
@@ -1,6 +1,6 @@
# Scanned PDF uploads
-Studio can extract text from scanned PDFs in Chat, project sources, and Data Recipes.
+Unsloth Studio can extract text from scanned PDFs in Chat, project sources, and Data Recipes.
Existing selectable text is preserved. OCR runs on image-only pages, including scans
with a selectable header or footer. Small logos do not require OCR when selectable
text is present. Large images with little selectable body text are treated as
@@ -16,8 +16,8 @@ other requests.
## Local OCR setup
PyMuPDF includes the OCR integration. Install Tesseract language data on the machine
-running the Studio backend, and set TESSDATA_PREFIX to the folder containing the
-.traineddata files before starting Studio. No data is downloaded automatically.
+running the Unsloth Studio backend, and set TESSDATA_PREFIX to the folder containing the
+.traineddata files before starting Unsloth Studio. No data is downloaded automatically.
For example, install your operating system's Tesseract package and English language
data, then point TESSDATA_PREFIX to its tessdata directory. You can also obtain
@@ -35,7 +35,7 @@ setting overrides this for chat/project uploads. Data Recipes follows the backen
default.
RAG_OCR_MAX_PAGES defaults to 20 scanned pages per PDF. Raise it before starting
-Studio when longer scans are expected. Local OCR uses RAG_OCR_DPI (default 150).
+Unsloth Studio when longer scans are expected. Local OCR uses RAG_OCR_DPI (default 150).
If scanned pages remain unreadable, the upload fails with their page numbers rather
than silently accepting an incomplete document. Configure the OCR language data,
diff --git a/studio/backend/core/export/orchestrator.py b/studio/backend/core/export/orchestrator.py
index 97953a42f3c..3695de545a9 100644
--- a/studio/backend/core/export/orchestrator.py
+++ b/studio/backend/core/export/orchestrator.py
@@ -235,7 +235,7 @@ class ExportOrchestrator:
# sweep has taken its snapshot, and the worker adopted below would then outlive
# Studio holding the model in memory.
if is_process_shutting_down():
- raise RuntimeError("Studio is shutting down; not starting an export subprocess")
+ raise RuntimeError("Unsloth is shutting down; not starting an export subprocess")
with (
child_environment_for_spawn(cache_env),
@@ -295,7 +295,7 @@ class ExportOrchestrator:
"export worker (pid %s) survived the reap; leaving it adopted",
_spawned_proc.pid,
)
- raise RuntimeError("Studio is shutting down; not starting an export subprocess")
+ raise RuntimeError("Unsloth is shutting down; not starting an export subprocess")
logger.info("Export subprocess started (pid=%s)", _spawned_proc.pid)
def _shutdown_subprocess(self, timeout: float = 10.0) -> bool:
diff --git a/studio/backend/core/inference/chat_generation_runs.py b/studio/backend/core/inference/chat_generation_runs.py
index ffd240adacf..33272c6a919 100644
--- a/studio/backend/core/inference/chat_generation_runs.py
+++ b/studio/backend/core/inference/chat_generation_runs.py
@@ -675,7 +675,7 @@ class ChatGenerationSupervisor:
worker_token = worker_token,
status = "failed" if shutting_down else "cancelled",
finish_reason = "interrupted" if shutting_down else "cancelled",
- error = "Studio shut down during generation" if shutting_down else None,
+ error = "Unsloth shut down during generation" if shutting_down else None,
)
return
# Spans the lifecycle gate as well as preparation: a run waiting on the gate is still queued, so its lease
@@ -690,7 +690,7 @@ class ChatGenerationSupervisor:
worker_token = worker_token,
status = "failed" if shutting_down else "cancelled",
finish_reason = "interrupted" if shutting_down else "cancelled",
- error = "Studio shut down during generation" if shutting_down else None,
+ error = "Unsloth shut down during generation" if shutting_down else None,
)
return
if not await asyncio.to_thread(db.mark_running, run_id, worker_token):
@@ -809,7 +809,7 @@ class ChatGenerationSupervisor:
if run_id in self._shutdown_runs:
status = "failed"
finish_reason = "interrupted"
- error = "Studio shut down during generation"
+ error = "Unsloth shut down during generation"
elif current["cancelRequested"] or (cancel_event.is_set() and error is None):
# A bare event is not proof of a user stop: the streaming paths set this same event from their cleanup
# after emitting an in-band error, so a parsed failure outranks it. An explicit cancelRequested still
@@ -864,7 +864,7 @@ class ChatGenerationSupervisor:
error = (
None
if cancelled
- else "Studio shut down during generation"
+ else "Unsloth shut down during generation"
if shutting_down
else str(exc)[:1000]
),
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index c42b8bc45b4..b3949bc1a30 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -25353,7 +25353,7 @@ class LlamaCppBackend:
# _detected_gpus, not gpus: Manual modes empty gpus on purpose.
# gpu_ids is pinned for the child below; the arch gate warns better.
logger.warning(
- "Studio could not enumerate any GPU, so this load was planned "
+ "Unsloth could not enumerate any GPU, so this load was planned "
"without device information: %s",
LlamaCppBackend._explain_empty_gpu_probe(binary),
)
diff --git a/studio/backend/core/inference/native_audio.py b/studio/backend/core/inference/native_audio.py
index e955099ec97..f324a877a1b 100644
--- a/studio/backend/core/inference/native_audio.py
+++ b/studio/backend/core/inference/native_audio.py
@@ -674,7 +674,7 @@ def native_audio_download_plan(model_name: str, hf_token: Optional[str] = None)
audio_type = _native_audio_type(normalized)
if audio_type in PYTHON310_AUDIO_TYPES and sys.version_info < (3, 10):
family = "Higgs TTS" if audio_type.startswith("higgs_") else "MiniMax Music 3"
- raise ValueError(f"{family} requires Python 3.10 or newer in Studio.")
+ raise ValueError(f"{family} requires Python 3.10 or newer in Unsloth.")
if local_checkpoint and audio_type is None:
return {
"entries": [],
@@ -947,7 +947,7 @@ class NativeAudioBackend:
"its official local runtime does not support AMD ROCm."
)
if audio_type == "minimax_music3" and sys.version_info < (3, 10):
- raise RuntimeError("MiniMax Music 3 requires Python 3.10 or newer in Studio.")
+ raise RuntimeError("MiniMax Music 3 requires Python 3.10 or newer in Unsloth.")
if model_name in self.models:
self.active_model_name = model_name
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index a795698e592..fb819812572 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -557,7 +557,7 @@ class InferenceOrchestrator:
from utils.process_lifetime import is_process_shutting_down
if is_process_shutting_down():
- raise RuntimeError("Studio is shutting down; not starting an inference subprocess")
+ raise RuntimeError("Unsloth is shutting down; not starting an inference subprocess")
from utils.native_path_leases import (
native_path_secret_removed_for_child_start,
run_without_native_path_secret,
@@ -655,7 +655,7 @@ class InferenceOrchestrator:
)
except Exception as exc:
logger.debug("Could not reap the raced inference worker: %s", exc)
- raise RuntimeError("Studio is shutting down; not starting an inference subprocess")
+ raise RuntimeError("Unsloth is shutting down; not starting an inference subprocess")
logger.info("Inference subprocess started (pid=%s)", _spawned_proc.pid)
def _cancel_generation(self) -> None:
diff --git a/studio/backend/core/inference/sd_cpp_engine.py b/studio/backend/core/inference/sd_cpp_engine.py
index b18ab6cb1a8..aa2a0fd7e57 100644
--- a/studio/backend/core/inference/sd_cpp_engine.py
+++ b/studio/backend/core/inference/sd_cpp_engine.py
@@ -836,7 +836,7 @@ class SdCppEngine:
# shutdown sweep has taken its snapshot, and sd-cli would then keep running with
# nothing left to reap it, holding VRAM past the app.
if is_process_shutting_down():
- raise SdCppCancelled("Studio is shutting down; not starting sd-cli.")
+ raise SdCppCancelled("Unsloth is shutting down; not starting sd-cli.")
proc = subprocess.Popen(
cmd,
stdout = subprocess.PIPE,
@@ -860,7 +860,7 @@ class SdCppEngine:
if is_process_shutting_down():
logger.info("shutdown began during the spawn; killing the new sd-cli")
_terminate(proc)
- raise SdCppCancelled("Studio is shutting down; not starting sd-cli.")
+ raise SdCppCancelled("Unsloth is shutting down; not starting sd-cli.")
# Drain stdout on a reader thread so the timeout holds even when the child hangs WITHOUT printing (a plain `for
# line in proc.stdout` blocks until EOF). Lines, then a None sentinel, go to a queue the main loop polls against
# a wall-clock deadline. iter_sd_cpp_records also splits sd-cli's in-place progress redraws, which carry no
diff --git a/studio/backend/core/inference/sd_cpp_server.py b/studio/backend/core/inference/sd_cpp_server.py
index b32623f3d80..5076c29abba 100644
--- a/studio/backend/core/inference/sd_cpp_server.py
+++ b/studio/backend/core/inference/sd_cpp_server.py
@@ -270,7 +270,7 @@ class SdCppServer:
# sd-server after the step-7 sweep has taken its snapshot.
if is_process_shutting_down():
self._spawn_error = RuntimeError(
- "Studio is shutting down; not starting sd-server"
+ "Unsloth is shutting down; not starting sd-server"
)
spawned.set()
return
@@ -316,7 +316,7 @@ class SdCppServer:
)
self._process = None
self._spawn_error = RuntimeError(
- "Studio is shutting down; not starting sd-server"
+ "Unsloth is shutting down; not starting sd-server"
)
spawned.set()
return
diff --git a/studio/backend/core/inference/stt_ggml_sidecar.py b/studio/backend/core/inference/stt_ggml_sidecar.py
index 9432d9de395..960f64115f5 100644
--- a/studio/backend/core/inference/stt_ggml_sidecar.py
+++ b/studio/backend/core/inference/stt_ggml_sidecar.py
@@ -983,7 +983,7 @@ class GgmlSttSidecar:
# would otherwise start whisper-server after the sweep had run.
if is_process_shutting_down():
raise SttLoadCancelledError(
- "Studio is shutting down; not starting whisper-server."
+ "Unsloth is shutting down; not starting whisper-server."
)
process = subprocess.Popen(
command,
@@ -1010,7 +1010,7 @@ class GgmlSttSidecar:
process.wait(timeout = 10)
forget_pid(process.pid)
raise SttLoadCancelledError(
- "Studio is shutting down; not starting whisper-server."
+ "Unsloth is shutting down; not starting whisper-server."
)
try:
self._wait_for_server(process, port, cancel_event)
diff --git a/studio/backend/core/inference/stt_mtmd_sidecar.py b/studio/backend/core/inference/stt_mtmd_sidecar.py
index 63416d5007c..b0683fa27d9 100644
--- a/studio/backend/core/inference/stt_mtmd_sidecar.py
+++ b/studio/backend/core/inference/stt_mtmd_sidecar.py
@@ -938,7 +938,7 @@ class MtmdSttSidecar:
# step-7 sweep has already passed by.
if is_process_shutting_down():
raise SttLoadCancelledError(
- "Studio is shutting down; not starting the MTMD server."
+ "Unsloth is shutting down; not starting the MTMD server."
)
process = subprocess.Popen(
cmd,
@@ -963,7 +963,7 @@ class MtmdSttSidecar:
if is_process_shutting_down():
_reap(process)
raise SttLoadCancelledError(
- "Studio is shutting down; not starting the MTMD server."
+ "Unsloth is shutting down; not starting the MTMD server."
)
if not self._wait_for_server(process, port, cancel_event):
# Reap it here: _process was never assigned, so unload() cannot reach a child that ignores SIGTERM and
diff --git a/studio/backend/core/inference/stt_transformers_worker.py b/studio/backend/core/inference/stt_transformers_worker.py
index 0af7e340241..61f15662dec 100644
--- a/studio/backend/core/inference/stt_transformers_worker.py
+++ b/studio/backend/core/inference/stt_transformers_worker.py
@@ -337,7 +337,9 @@ class WhisperWorker:
# cancel_pending_loads only reaches the chat /load attempts, so without this a
# quit during an STT load starts a worker the step-7 sweep has already passed.
if is_process_shutting_down():
- raise SttWorkerSpawnError("Studio is shutting down; not starting the dictation worker.")
+ raise SttWorkerSpawnError(
+ "Unsloth is shutting down; not starting the dictation worker."
+ )
cache_env = get_hf_cache_paths().child_env({})
try:
with (
@@ -390,7 +392,9 @@ class WhisperWorker:
pass
self._process = None
self._close_queues()
- raise SttWorkerSpawnError("Studio is shutting down; not starting the dictation worker.")
+ raise SttWorkerSpawnError(
+ "Unsloth is shutting down; not starting the dictation worker."
+ )
logger.info(
"STT worker started (pid=%s) for %s on %s", _spawned_proc.pid, snapshot_path, device
)
diff --git a/studio/backend/core/rag/embed_llama_server.py b/studio/backend/core/rag/embed_llama_server.py
index 4e3862e17b5..d0d560b7baf 100644
--- a/studio/backend/core/rag/embed_llama_server.py
+++ b/studio/backend/core/rag/embed_llama_server.py
@@ -868,7 +868,7 @@ class LlamaServerBackend:
# encode still resolving or downloading its model as the app quits would
# otherwise Popen a server after terminate_all had taken its snapshot.
if is_process_shutting_down():
- raise RuntimeError("Studio is shutting down; not starting the embed server")
+ raise RuntimeError("Unsloth is shutting down; not starting the embed server")
proc = subprocess.Popen(
cmd,
stdout = subprocess.PIPE,
@@ -891,7 +891,7 @@ class LlamaServerBackend:
if is_process_shutting_down():
logger.info("shutdown began during the spawn; killing the new embed server")
self._kill_process()
- raise RuntimeError("Studio is shutting down; not starting the embed server")
+ raise RuntimeError("Unsloth is shutting down; not starting the embed server")
self._port = port
self._stdout_thread = account_thread(
target = self._drain_stdout,
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 5d17c2877fe..e2e75a848bf 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -1824,7 +1824,7 @@ class TrainingBackend:
# the worker would then train on past it holding the GPU.
if is_process_shutting_down():
logger.info(
- "Studio is shutting down; not starting training worker for %s",
+ "Unsloth is shutting down; not starting training worker for %s",
start_request_id,
)
return False
@@ -1841,7 +1841,7 @@ class TrainingBackend:
# adoption ran first, so the worker is in the sweep record for as
# long as it exists.
if is_process_shutting_down():
- raise RuntimeError("Studio is shutting down")
+ raise RuntimeError("Unsloth is shutting down")
except Exception:
logger.error(
"Could not keep the training subprocess; terminating it",
@@ -2579,7 +2579,7 @@ class TrainingBackend:
# this respawn after the shutdown sweep has taken its snapshot.
if is_process_shutting_down():
raise RuntimeError(
- "Studio is shutting down; not respawning the training worker"
+ "Unsloth is shutting down; not respawning the training worker"
)
new_proc.start()
adopt_pid(new_proc.pid)
@@ -2602,7 +2602,7 @@ class TrainingBackend:
"could not reap the new training worker", exc_info = True
)
raise RuntimeError(
- "Studio is shutting down; not respawning the training worker"
+ "Unsloth is shutting down; not respawning the training worker"
)
except Exception:
logger.error("Failed to respawn training subprocess", exc_info = True)
diff --git a/studio/backend/hub/services/models/folder_browser.py b/studio/backend/hub/services/models/folder_browser.py
index 4b38152dcb4..2e17a6659ad 100644
--- a/studio/backend/hub/services/models/folder_browser.py
+++ b/studio/backend/hub/services/models/folder_browser.py
@@ -88,7 +88,7 @@ def _build_browse_allowlist(
_add(outputs_root())
_add(exports_root())
except Exception as exc: # noqa: BLE001 -- best-effort
- logger.debug("browse-folders: studio roots unavailable: %s", exc)
+ logger.debug("browse-folders: Unsloth roots unavailable: %s", exc)
try:
for folder in list_scan_folders():
p = folder.get("path")
diff --git a/studio/backend/integrations/blender/launcher.py b/studio/backend/integrations/blender/launcher.py
index e14ade066c7..de4a3d390a3 100644
--- a/studio/backend/integrations/blender/launcher.py
+++ b/studio/backend/integrations/blender/launcher.py
@@ -17,7 +17,7 @@ def main() -> int:
return 2
runtime = Path(sys.argv[1])
if not (runtime / ".ready").is_file():
- print("Open Blender MCP in Studio and enable it to download the runtime.", file = sys.stderr)
+ print("Open Blender MCP in Unsloth and enable it to download the runtime.", file = sys.stderr)
return 1
sys.path.insert(0, str(runtime))
from blmcp import main as upstream_main
diff --git a/studio/backend/integrations/blender/service.py b/studio/backend/integrations/blender/service.py
index b60aa856d70..d6664c56ff5 100644
--- a/studio/backend/integrations/blender/service.py
+++ b/studio/backend/integrations/blender/service.py
@@ -111,7 +111,7 @@ async def probe(
except Exception:
return McpServerProbeResult(
ok = False,
- error = "Could not connect to Blender MCP. Check the Studio backend installation and retry.",
+ error = "Could not connect to Blender MCP. Check the Unsloth backend installation and retry.",
)
result = McpServerProbeResult(ok = True, tool_count = len(tools))
if check_bridge:
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index fd780cf33ce..cb3c04a14ff 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -2308,7 +2308,7 @@ class ChatCompletionRequest(BaseModel):
description = (
"[x-unsloth] How a local GGUF chat compacts once context_overflow is "
"truncate_oldest. 'checkpoint' resets to the latest turn plus standing "
- "instructions (Studio default). 'rolling' drops oldest complete turns. "
+ "instructions (Unsloth default). 'rolling' drops oldest complete turns. "
"Unset uses UNSLOTH_CONTEXT_POLICY."
),
)
@@ -2326,7 +2326,7 @@ class ChatCompletionRequest(BaseModel):
studio_tool_history: Optional[bool] = Field(
None,
description = (
- "[x-unsloth] The replayed tool calls were produced by Studio's local "
+ "[x-unsloth] The replayed tool calls were produced by Unsloth's local "
"tool loop rather than by an OpenAI-compatible client tool contract."
),
)
@@ -2715,7 +2715,7 @@ class ChatCountTokensRequest(ReasoningControlsRequest):
None,
description = (
"[x-unsloth] Mirrors ChatCompletionRequest: the replayed tool calls came from "
- "Studio's local tool loop, so _takes_tool_passthrough routes the count the way "
+ "Unsloth's local tool loop, so _takes_tool_passthrough routes the count the way "
"it routes the completion. Declared rather than left to extra='allow', which "
"coerces nothing and would read the string 'false' as a claim of ownership."
),
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index f22fcf49b1c..eff38bb86f4 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -5075,7 +5075,7 @@ def _skill_tool_tip(*, can_create: bool, compact: bool = False) -> str:
"message mentions an enabled skill as @skill-name, call read_skill for that named skill "
"before answering."
+ create_tip
- + " Skill allowed-tools metadata never overrides Studio tool permissions.\n"
+ + " Skill allowed-tools metadata never overrides Unsloth tool permissions.\n"
+ catalog
)
@@ -9659,7 +9659,7 @@ async def _maybe_auto_switch_model(
raise HTTPException(
status_code = 400,
detail = openai_error_body(
- "The requested text-to-speech model requires Python 3.10 or newer in Studio.",
+ "The requested text-to-speech model requires Python 3.10 or newer in Unsloth.",
status = 400,
code = "unsupported_runtime",
param = "model",
@@ -13844,7 +13844,7 @@ async def _preflight_native_audio_placement(
if audio_type in ("higgs_tts2", "higgs_tts3") and sys.version_info < (3, 10):
raise HTTPException(
status_code = 400,
- detail = "Higgs TTS requires Python 3.10 or newer in Studio.",
+ detail = "Higgs TTS requires Python 3.10 or newer in Unsloth.",
)
# Before every VRAM question: sizing a CPU load refuses it on a full GPU.
@@ -13880,7 +13880,7 @@ async def _preflight_native_audio_placement(
"MiniMax Music 3 requires an NVIDIA CUDA GPU in its official local runtime."
)
if sys.version_info < (3, 10):
- raise ValueError("MiniMax Music 3 requires Python 3.10 or newer in Studio.")
+ raise ValueError("MiniMax Music 3 requires Python 3.10 or newer in Unsloth.")
if device not in (hardware.DeviceType.CUDA, hardware.DeviceType.XPU):
if placement.requested_gpu_ids:
raise ValueError(
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 6862fe0683a..a6c0cdf73a9 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -1531,7 +1531,7 @@ def _build_browse_allowlist(
_add(outputs_root())
_add(exports_root())
except Exception as exc: # noqa: BLE001 -- best-effort
- logger.debug("browse-folders: studio roots unavailable: %s", exc)
+ logger.debug("browse-folders: Unsloth roots unavailable: %s", exc)
try:
for folder in list_scan_folders():
p = folder.get("path")
diff --git a/studio/backend/storage/chat_generation_runs_db.py b/studio/backend/storage/chat_generation_runs_db.py
index c05c267f10c..5dd0f37c2dc 100644
--- a/studio/backend/storage/chat_generation_runs_db.py
+++ b/studio/backend/storage/chat_generation_runs_db.py
@@ -749,7 +749,7 @@ def wait_for_events(
def reconcile_runs(
- *, error: str = "Studio restarted during generation", stale_after_ms: int | None = None
+ *, error: str = "Unsloth restarted during generation", stale_after_ms: int | None = None
) -> list[str]:
"""Settle active runs, returning the ids settled. ``stale_after_ms`` is what makes this safe to run
while Studio is serving: with it, only runs whose progress lease has not moved for that long are
@@ -816,5 +816,5 @@ def reconcile_runs(
conn.close()
-def reconcile_orphaned_runs(error: str = "Studio restarted during generation") -> int:
+def reconcile_orphaned_runs(error: str = "Unsloth restarted during generation") -> int:
return len(reconcile_runs(error = error))
diff --git a/studio/backend/tests/test_chat_generation_runs.py b/studio/backend/tests/test_chat_generation_runs.py
index 99f4ebe143f..b61b3c6eff0 100644
--- a/studio/backend/tests/test_chat_generation_runs.py
+++ b/studio/backend/tests/test_chat_generation_runs.py
@@ -456,7 +456,7 @@ def test_an_uncancelled_run_still_reconciles_as_interrupted(chat_home):
run = runs_db.get_run("run-1", "alice")
assert (run["status"], run["finishReason"]) == ("failed", "interrupted")
- assert run["error"] == "Studio restarted during generation"
+ assert run["error"] == "Unsloth restarted during generation"
assert runs_db.list_events("run-1")[-1]["payload"]["interrupted"] is True
diff --git a/studio/backend/tests/test_chat_generation_supervisor.py b/studio/backend/tests/test_chat_generation_supervisor.py
index 062323d597a..7f9671b4579 100644
--- a/studio/backend/tests/test_chat_generation_supervisor.py
+++ b/studio/backend/tests/test_chat_generation_supervisor.py
@@ -579,7 +579,7 @@ async def test_graceful_supervisor_shutdown_is_interrupted(durable_run, monkeypa
await supervisor.stop()
run = runs_db.get_run("run-1", "alice")
assert (run["status"], run["finishReason"]) == ("failed", "interrupted")
- assert run["error"] == "Studio shut down during generation"
+ assert run["error"] == "Unsloth shut down during generation"
def test_thread_delete_captures_durable_run_before_cascade(durable_run):
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 67a68855c7e..2120b84a6f9 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -10869,9 +10869,9 @@ def install_python_stack() -> int:
# )
# 8. Unsloth dependencies
- if not _skip_step(REQ_ROOT / "studio.txt", "studio deps", no_deps = False):
+ if not _skip_step(REQ_ROOT / "studio.txt", "Unsloth Studio deps", no_deps = False):
pip_install(
- "Installing studio dependencies",
+ "Installing Unsloth Studio dependencies",
"--no-cache-dir",
req = REQ_ROOT / "studio.txt",
)
diff --git a/tests/python/test_cross_platform_parity.py b/tests/python/test_cross_platform_parity.py
index 985edde47e1..06515d8ded8 100644
--- a/tests/python/test_cross_platform_parity.py
+++ b/tests/python/test_cross_platform_parity.py
@@ -1139,7 +1139,7 @@ class TestInstallUvCacheRootParity:
"preserving custom UV_CACHE_DIR",
"reusing existing shared cache",
"avoid duplicate Torch/CUDA downloads",
- "using new Studio-owned cache",
+ "using new Unsloth Studio-owned cache",
"already-cached packages may download again",
"so cached packages may download again",
):
diff --git a/tests/python/test_docker_studio_password_banner.py b/tests/python/test_docker_studio_password_banner.py
index 8c5b9793479..b589e14d6c0 100644
--- a/tests/python/test_docker_studio_password_banner.py
+++ b/tests/python/test_docker_studio_password_banner.py
@@ -131,7 +131,7 @@ def test_the_generated_password_is_printed_once_studio_writes_it(tmp_path: Path)
assert "60 minutes" in res.stdout, "the change-it-or-shut-down window is not explained"
assert "Unsloth container ready" in res.stdout
assert (
- "Studio http://localhost:8000 username: unsloth password: s3cret pass"
+ "Unsloth http://localhost:8000 username: unsloth password: s3cret pass"
in res.stdout
)
diff --git a/tests/python/test_docker_studio_port.py b/tests/python/test_docker_studio_port.py
index 574f5b91453..5d95389bdfd 100644
--- a/tests/python/test_docker_studio_port.py
+++ b/tests/python/test_docker_studio_port.py
@@ -126,7 +126,7 @@ def test_the_ready_summary_and_its_health_probe_use_the_port(tmp_path: Path):
)
res = subprocess.run(["bash", str(PASSWORD)], capture_output = True, text = True, env = e, timeout = 60)
assert res.returncode == 0, res.stderr
- assert "Studio http://localhost:9000" in res.stdout, res.stdout
+ assert "Unsloth http://localhost:9000" in res.stdout, res.stdout
assert "http://127.0.0.1:9000/api/health" in log.read_text(encoding = "utf-8")
diff --git a/tests/python/test_windows_python_venv_hardening.py b/tests/python/test_windows_python_venv_hardening.py
index 17d675c37d8..854dca01d0e 100644
--- a/tests/python/test_windows_python_venv_hardening.py
+++ b/tests/python/test_windows_python_venv_hardening.py
@@ -764,13 +764,13 @@ try {{
"use --isolated-uv-cache to isolate"
),
"studio": (
- f"using new Studio-owned cache ({studio}); part of {shared} could not be "
+ f"using new Unsloth Studio-owned cache ({studio}); part of {shared} could not be "
"read, so cached packages may download again"
if default_state == "denied-bucket"
- else f"using new Studio-owned cache ({studio})"
+ else f"using new Unsloth Studio-owned cache ({studio})"
),
"isolated": (
- f"forced Studio cache isolation ({studio}); already-cached packages may download again"
+ f"forced Unsloth Studio cache isolation ({studio}); already-cached packages may download again"
),
}[expected_mode]
diff --git a/tests/sh/test_install_uv_cache_root.sh b/tests/sh/test_install_uv_cache_root.sh
index 18a807405c7..1ee44d44339 100644
--- a/tests/sh/test_install_uv_cache_root.sh
+++ b/tests/sh/test_install_uv_cache_root.sh
@@ -121,14 +121,14 @@ for shell in sh bash; do
run_case "$shell" "missing default selects Studio" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$HOME_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
mkdir -p "$HOME_CACHE/CACHEDIR.TAG/inside"
: > "$HOME_CACHE/CACHEDIR.TAG/inside/payload"
: > "$HOME_CACHE/.gitignore"
run_case "$shell" "marker-only default stays Studio and is not traversed" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$HOME_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
if [ -f "$HOME_CACHE/CACHEDIR.TAG/inside/payload" ] && [ -f "$HOME_CACHE/.gitignore" ]; then
ok "$shell: marker-only probe is non-destructive"
else
@@ -142,7 +142,7 @@ for shell in sh bash; do
: > "$HOME_CACHE/.lock"
run_case "$shell" "uv venv scaffolding stays Studio-owned" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$HOME_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
mkdir -p "$HOME_CACHE/archive-v0/package"
: > "$HOME_CACHE/archive-v0/package/payload.py"
@@ -168,7 +168,7 @@ for shell in sh bash; do
mkdir -p "$EMPTY_XDG_CACHE"
run_case "$shell" "empty XDG does not fall back to populated HOME" unset "" false \
"$HOME_DIR" value "$EMPTY_XDG" "$ROOT" "$EMPTY_XDG_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
CONFIG_CACHE="$CASE/uv.toml cache"
mkdir -p "$CONFIG_CACHE/wheels-v5/package"
@@ -188,11 +188,11 @@ for shell in sh bash; do
run_case "$shell" "forced isolation uses Studio despite populated default" unset "" true \
"$HOME_DIR" value "$XDG_DIR" "$ROOT" "$XDG_CACHE" "$STUDIO_CACHE" isolated \
- "forced Studio cache isolation ($STUDIO_CACHE); already-cached packages may download again" "$STUDIO_CACHE"
+ "forced Unsloth Studio cache isolation ($STUDIO_CACHE); already-cached packages may download again" "$STUDIO_CACHE"
run_case "$shell" "unresolvable default safely selects Studio" unset "" false \
"" unset "" "$ROOT" "" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
# wheels-v* is .msgpack/.http only on uv 0.10: one `--dry-run` leaves a file.
META_CACHE="$CASE/metadata only/uv"
@@ -205,7 +205,7 @@ for shell in sh bash; do
: > "$META_CACHE/sdists-v9/pypi/pkg/download.lock"
run_case "$shell" "metadata-only default is not a warm cache" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$META_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
mkdir -p "$META_CACHE/archive-v0/hash/torch"
: > "$META_CACHE/archive-v0/hash/torch/_C.so"
@@ -242,7 +242,7 @@ for shell in sh bash; do
if [ "$(id -u 2>/dev/null || echo 0)" != 0 ] && chmod 000 "$DENIED_CACHE/archive-v0" 2>/dev/null; then
run_case "$shell" "unreadable bucket says why it fell back" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$DENIED_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE); part of $DENIED_CACHE could not be read, so cached packages may download again" \
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE); part of $DENIED_CACHE could not be read, so cached packages may download again" \
"$STUDIO_CACHE"
chmod 755 "$DENIED_CACHE/archive-v0" 2>/dev/null || true
fi
@@ -274,7 +274,7 @@ for shell in sh bash; do
EMPTY_CACHE="$CASE/unrecorded/uv"
run_case "$shell" "studio mode still selects the Studio cache" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$EMPTY_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
check_marker "studio mode records the Studio cache" "$STUDIO_CACHE"
rm -f "$MARKER"
@@ -291,11 +291,11 @@ for shell in sh bash; do
_PRESET_MARKER="$STUDIO_CACHE"
run_case "$shell" "a warm recorded Studio cache outranks a warm default" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$BUILDS_CACHE" "$STUDIO_CACHE" studio \
- "reusing this install's Studio cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "reusing this install's Unsloth Studio cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
# ...and it does NOT outrank an explicit isolation request.
run_case "$shell" "isolation still wins over a warm marker" unset "" true \
"$HOME_DIR" unset "" "$ROOT" "$BUILDS_CACHE" "$STUDIO_CACHE" isolated \
- "forced Studio cache isolation ($STUDIO_CACHE); already-cached packages may download again" \
+ "forced Unsloth Studio cache isolation ($STUDIO_CACHE); already-cached packages may download again" \
"$STUDIO_CACHE"
# A marker naming a cache that EXISTS and is no longer warm loses to uv's default, so a
# machine that really has moved on still reaches the shared cache.
@@ -311,7 +311,7 @@ for shell in sh bash; do
_PRESET_MARKER="$CASE/gone/uv"
run_case "$shell" "a marker naming a deleted cache keeps Studio" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$STUDIO_CACHE" "$STUDIO_CACHE" studio \
- "reusing this install's Studio cache ($STUDIO_CACHE)" \
+ "reusing this install's Unsloth Studio cache ($STUDIO_CACHE)" \
"$STUDIO_CACHE"
_PRESET_MARKER=""
rm -rf "$STUDIO_CACHE"
@@ -321,7 +321,7 @@ for shell in sh bash; do
rm -f "$MARKER"
run_case "$shell" "isolation still forces the Studio cache" unset "" true \
"$HOME_DIR" unset "" "$ROOT" "$BUILDS_CACHE" "$STUDIO_CACHE" isolated \
- "forced Studio cache isolation ($STUDIO_CACHE); already-cached packages may download again" \
+ "forced Unsloth Studio cache isolation ($STUDIO_CACHE); already-cached packages may download again" \
"$STUDIO_CACHE"
check_marker "isolated mode records the Studio cache" "$STUDIO_CACHE"
@@ -340,7 +340,7 @@ for shell in sh bash; do
ln -s "$VICTIM" "$ROOT/cache/uv-cache-dir" 2>/dev/null && {
run_case "$shell" "a symlinked marker is replaced, not followed" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$EMPTY_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
if [ "$(cat "$VICTIM")" = "do not clobber" ] && [ ! -L "$ROOT/cache/uv-cache-dir" ]; then
ok "$shell: the symlink target was left alone"
else
@@ -354,7 +354,7 @@ for shell in sh bash; do
if [ "$(id -u 2>/dev/null || echo 0)" != 0 ] && chmod 200 "$MARKER" 2>/dev/null; then
run_case "$shell" "an unreadable marker does not fail the install" unset "" false \
"$HOME_DIR" unset "" "$ROOT" "$EMPTY_CACHE" "$STUDIO_CACHE" studio \
- "using new Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
+ "using new Unsloth Studio-owned cache ($STUDIO_CACHE)" "$STUDIO_CACHE"
chmod 600 "$MARKER" 2>/dev/null || true
if [ "$(cat "$MARKER")" = "/previous/install/cache" ]; then
ok "$shell: an unreadable marker is left as it was"
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 706e413df47..fd9feadb75e 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -543,7 +543,7 @@ def _load_run_module():
spec = importlib.util.spec_from_file_location("studio.backend.run", run_py)
if spec is None or spec.loader is None:
- raise ImportError(f"Could not load studio backend from {run_py}")
+ raise ImportError(f"Could not load Unsloth backend from {run_py}")
module = importlib.util.module_from_spec(spec)
sys.modules["studio.backend.run"] = module
try: