Feature Request: Minimal Installation Option for Open WebUI
Problem Statement
Currently, when installing Open WebUI via pip (pip install open-webui), the installation includes a full set of dependencies including NVIDIA CUDA drivers, PyTorch, and various ML libraries that total approximately 7.7GB. However, many users only need the frontend UI and backend server to connect to a remote OpenAI, Anthropic or other instances running on a separate machine, or in the cloud, making these heavy dependencies unnecessary.
Specifically, for VMPilot, the Open WebUI frontend and backend are the only components needed, as the ML model is in the cloud.
Current Situation
When installing via pip, the following large packages are included even when they won't be used:
The Docker implementation already has an option to run with a remote Ollama instance, but the pip installation doesn't offer a lightweight alternative.
Feature Request
Create a minimal installation option for Open WebUI that:
Includes both frontend UI and backend server components
Removes unnecessary ML/CUDA or Ollama dependencies
Maintains core functionality like chat history, user management, and other backend services
Significantly reduces the installation size (potentially from 7.7GB to <1GB)
Proposed Implementation Options
Create optional dependency groups in pyproject.toml:
[project.optional-dependencies]minimal=[# Core dependencies for frontend and backend without ML/CUDA]full=[# All dependencies including ML/CUDA]
Users could then install with: pip install open-webui[minimal]
Create an installation script based on the Dockerfile's conditional logic:
The Dockerfile already has conditional logic using build arguments (USE_CUDA, USE_OLLAMA, etc.)
This could be translated into an installation script with command-line options
Users could specify whether they want CUDA support, which embedding model to use, etc.
Provide a separate pip package like open-webui-minimal that only includes the necessary components for connecting to a remote Ollama instance.
Benefits
Reduced disk usage requirements
Faster installation times
Better suited for low-resource environments
More flexible deployment options
Improved user experience for those who run Ollama on a separate machine
Technical Considerations
User data, chat history, and authentication should remain fully functional
The frontend UI should be identical to the full installation
Documentation should clearly explain the differences between installation options
Originally created by @dror-llm on GitHub (Mar 20, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/11901
## Feature Request: Minimal Installation Option for Open WebUI
### Problem Statement
Currently, when installing Open WebUI via pip (`pip install open-webui`), the installation includes a full set of dependencies including NVIDIA CUDA drivers, PyTorch, and various ML libraries that total approximately 7.7GB. However, many users only need the frontend UI and backend server to connect to a remote OpenAI, Anthropic or other instances running on a separate machine, or in the cloud, making these heavy dependencies unnecessary.
Specifically, for VMPilot, the Open WebUI frontend and backend are the only components needed, as the ML model is in the cloud.
### Current Situation
When installing via pip, the following large packages are included even when they won't be used:
```
2807552 /opt/venv/lib/python3.12/site-packages/nvidia
1583956 /opt/venv/lib/python3.12/site-packages/torch
701508 /opt/venv/lib/python3.12/site-packages/triton
381248 /opt/venv/lib/python3.12/site-packages/open_webui
207676 /opt/venv/lib/python3.12/site-packages/cusparselt
144852 /opt/venv/lib/python3.12/site-packages/milvus_lite
140704 /opt/venv/lib/python3.12/site-packages/pyarrow
130848 /opt/venv/lib/python3.12/site-packages/playwright
118796 /opt/venv/lib/python3.12/site-packages/scipy
99296 /opt/venv/lib/python3.12/site-packages/transformers
92284 /opt/venv/lib/python3.12/site-packages/opencv_python.libs
...
```
The Docker implementation already has an option to run with a remote Ollama instance, but the pip installation doesn't offer a lightweight alternative.
### Feature Request
Create a minimal installation option for Open WebUI that:
1. Includes both frontend UI and backend server components
2. Removes unnecessary ML/CUDA or Ollama dependencies
3. Maintains core functionality like chat history, user management, and other backend services
4. Significantly reduces the installation size (potentially from 7.7GB to <1GB)
### Proposed Implementation Options
1. **Create optional dependency groups in pyproject.toml**:
```toml
[project.optional-dependencies]
minimal = [
# Core dependencies for frontend and backend without ML/CUDA
]
full = [
# All dependencies including ML/CUDA
]
```
Users could then install with: `pip install open-webui[minimal]`
2. **Create an installation script** based on the Dockerfile's conditional logic:
- The Dockerfile already has conditional logic using build arguments (`USE_CUDA`, `USE_OLLAMA`, etc.)
- This could be translated into an installation script with command-line options
- Users could specify whether they want CUDA support, which embedding model to use, etc.
3. **Provide a separate pip package** like `open-webui-minimal` that only includes the necessary components for connecting to a remote Ollama instance.
### Benefits
- Reduced disk usage requirements
- Faster installation times
- Better suited for low-resource environments
- More flexible deployment options
- Improved user experience for those who run Ollama on a separate machine
### Technical Considerations
- User data, chat history, and authentication should remain fully functional
- The frontend UI should be identical to the full installation
- Documentation should clearly explain the differences between installation options
Sorry again. Didn't realize that github automatically creates the ticket in the repo you're in.
This is where it is supposed to go https://github.com/drorm/vmpilot/issues/42
<!-- gh-comment-id:2739493797 -->
@dror-llm commented on GitHub (Mar 20, 2025):
Sorry again. Didn't realize that github automatically creates the ticket in the repo you're in.
This is where it is supposed to go
https://github.com/drorm/vmpilot/issues/42
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Originally created by @dror-llm on GitHub (Mar 20, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/11901
Feature Request: Minimal Installation Option for Open WebUI
Problem Statement
Currently, when installing Open WebUI via pip (
pip install open-webui), the installation includes a full set of dependencies including NVIDIA CUDA drivers, PyTorch, and various ML libraries that total approximately 7.7GB. However, many users only need the frontend UI and backend server to connect to a remote OpenAI, Anthropic or other instances running on a separate machine, or in the cloud, making these heavy dependencies unnecessary.Specifically, for VMPilot, the Open WebUI frontend and backend are the only components needed, as the ML model is in the cloud.
Current Situation
When installing via pip, the following large packages are included even when they won't be used:
The Docker implementation already has an option to run with a remote Ollama instance, but the pip installation doesn't offer a lightweight alternative.
Feature Request
Create a minimal installation option for Open WebUI that:
Proposed Implementation Options
Create optional dependency groups in pyproject.toml:
Users could then install with:
pip install open-webui[minimal]Create an installation script based on the Dockerfile's conditional logic:
USE_CUDA,USE_OLLAMA, etc.)Provide a separate pip package like
open-webui-minimalthat only includes the necessary components for connecting to a remote Ollama instance.Benefits
Technical Considerations
@dror-llm commented on GitHub (Mar 20, 2025):
Sorry again. Didn't realize that github automatically creates the ticket in the repo you're in.
This is where it is supposed to go
https://github.com/drorm/vmpilot/issues/42