GenAI - howto 

Feel free to add notes/ pointers/ advice for others at HWU !

CoPilot

 https://copilot.microsoft.com/ 

This is Microsoft Copilot which is running in our tenancy so that your data are stored securely. Thanks to this offering, you should be able to access the most recent OpenAI LLM as well as other image generation models  (Dall-E) that can be used to generate images from the same interface. To access it, complete the login with your HWU Microsoft credentials.  

See also :  https://heriotwatt.sharepoint.com/sites/artificialintelligence/SitePages/Getting-started-with-AI.aspx

Ollama

 Alessandro recommends  students in the NLP course to use Ollama: https://ollama.com/ . If you have a MacBook it is blazing fast and works really well locally. This contains both LLMs and VLMs (including LLava). You can find models of different sizes which are quantised to run very fast. You can then use the following API from Python: https://github.com/ollama/ollama-python 

Lightning AI

 https://lightning.ai/sign-up  

LMStudio

https://lmstudio.ai/ 

Phil says: You can share processing between CPU / GPU, it mimics the OpenAI API so code can easily be switched from running against a local LLM to OpenAI's cloud processing if you want, and it integrates really well to the Hugging Face Hub to browse and download different models/quantisations inclucing LVMs (eg LLava). I'm running it on Windows at home, and the beta on Linux at work.

Access to Open AI models via Azure 

Please raise a ticket with IS for each request providing some information on the intended use and include the following details:

1. Business Unit

2. Business Unit Code

3. Cost Centre Code / Grant Job Code

4. Project Name ( or if Teaching the Subject name)

5. Project Code (or if Teaching the Subject Code)

6. Requisitioner Name

7. Budget allocated in £

8. Budget Contact (name and email of someone to be alerted when reaching the budget allocated. )

9. Language Model required

10. Data Sovereignty requirements - please specify if data is to stay in the UK, EU or NONE if it can be anywhere in the world.

        * please note not all Language Models are present in all regions

        * for GPT-4o having data placed regionally can come with a higher cost than having it global.

11. We will also require the names of those who will require access, and the level of access they would require. Contributor or User

ssh Gateway

 https://www.macs.hw.ac.uk/~imcc/sshgw.docx