Failure when deploying a custom serving endpoint LLLM

[ismaelhenzel]

I'm currently experimenting with vector search using Databricks. Everything runs smoothly when I load the model deployed in Unity Catalog into a notebook session and ask questions using Python. However, when I attempt to serve it, I encounter a generic error.

The container builds successfully. However, upon running the code, I encounter an error. Debugging this serve endpoint is challenging because the machine is not available for an interactive session. I've observed that the error occurs when the code attempts to retrieve the index or endpoint, as shown below:

• vsc.get_endpoint(name=vector_search_endpoint_name) returns  An error occurred while loading the model. Expecting value: line 1 column 1 (char 0).
• index = vsc.get_index(vector_search_endpoint_name, vs_index_fullname) returns  An error occurred while loading the model. Expecting value: line 1 column 1 (char 0)

** my workspace is fully private

** i'm basing my code in this databricks example https://notebooks.databricks.com/demos/llm-rag-chatbot/index.html#

[ismaelhenzel]

i'm using openai text-embedding-3-large as embedding model, dbrx as chat model, and databricks as vectorstore, everything deployed and working fine in the workspace. But for some reason, error trying to serve the model, my unity catalog is aready in a public storage for this poc because the serve didn't support the firewall of a private storage.

[srikanth009]

Even I am face similar challenges of debugging this serve endpoint due to non interactive session. Looking for any alternate solutions. It was running perfectly when logging and loading the model in databricks, but shows errors after creating an endpoint and while querying it.

[brycejune]

Ensure your vector_search_endpoint_name and vs_index_fullname match the deployment setup. Check model deployment logs for detailed errors and confirm your workspace's network settings allow access to Unity Catalog and model serving endpoints in a private workspace.
Regards,

Bryce June

[Usmanr000]

"Deploying a custom serving endpoint for LLMs can be challenging, especially when handling model dependencies and scaling issues. Has anyone found a reliable workaround for deployment failures? Also, for those looking for updates on government assistance programs in Pakistan, you can check 8171  for the latest BISP and Ehsaas Program details."

[Goiderearsur]

Get your Ehsaas Program 8171 online registration and CNIC check online now. Enter your National ID card number & get Rs. 25000/12000 grant.8171 web portal