Skip to content

nkityd09/cml_hf_models

Repository files navigation

CML HuggingFace Models

Choose any 7B or 13B LLM from HuggingFace and deploy as a CML Model. Cloudera Machine Learning models expose an Inference endpoint for users to access and communicate with. The AMP creates a Gradio App UI which can be used to interact with the deployed CML Model.

Key Features

  • Launch any 7B or 13B HuggingFace Model(Can be configured for larger models after deployment)
  • Deploys CML Model with 1 replica
  • Deploys CML Gradio Application which interacts with deployed CML Model

Prerequisite

Increase Ephemeral Storage Limit by navigating to CML Workspace -> Site Administration -> Settings -> Ephemeral Storage (in GB) and set it to a value >= 50

When a CML model is created, the model is loaded in the scratch space of a pod, LLM models are larger than the default 10 GB which causes issues during deployment.

Resource Requirements

The AMP Model has been configured to use the following

  • 4 CPU
  • 32 GB RAM
  • 1 GPUs

Steps to Configure CML AMP

  1. Navigate to CML Workspace -> Site Administration -> AMPs Tab

  2. Under AMP Catalog Sources section, We will "Add a new source By" selecting "Catalog File URL"

  3. Provide the following URL and click "Add Source"

https://raw.githubusercontent.com/nkityd09/cml_catalog/main/catalog.yaml
  1. Once added, We will be able to see the CML HuggingFace Models in the AMP section and deploy it from there.

  2. Click on the AMP and "Configure Project"

    • Add HuggingFace Model Name, defaults to HuggingFaceH4/zephyr-7b-alpha
    • If accessing a gated model, add HuggingFace token. Can be left blank for non gated models Configuring AMP
  3. Once the AMP steps are completed, We can access the Gradio UI via the Applications page.

CML Application in Action

The CML Application serves a Gradio UI to interact with an LLM deployed as CML Model

The Gradio UI provides a Question Textbox which can be used to send questions to the CML Model. The submit button sends the request, its response is displayed in the Answer Text Box. Application UI

About

Deploy HF models on CML

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published