search

NVIDIA-NIM-Agent-Blueprints / generative-virtual-screening

NIM Agent Blueprint for generative AI-based virtual screening
https://build.nvidia.com/nvidia/generative-virtual-screening-for-drug-discovery
Apache License 2.0
8 stars 4 forks source link

AF2 dataset download, docker login, and minimum hardware requirements #1

Open xinyu-dev opened 3 weeks ago

xinyu-dev commented 3 weeks ago

This is very well-written workflow with clear instructions. Great job!

I noticed a couple of things:

  1. docker compose up does not seem to download the AF2 dataset properly. This results in dataset not found issue when trying to run health check. Similar health checks passed for Diffdock and MolMIM, but not for AlphaFold.
  2. To solve the issue above I launched the AlphaFold2 NIM standalone, like here: https://docs.nvidia.com/nim/bionemo/alphafold2/latest/quickstart-guide.html . The data folder should have a size of ~500GB if correctly downloaded.
  3. I tested in SageMaker Notebook instance. Before running docker compose up, I had to do docker login with username as oauthtoken and password as NGC API key. Otherwise it seems to generate a not authorized error even though I have both NGC_API_KEY and NGC_CLI_API_KEY set up as env variables.
  4. The https://build.nvidia.com/nvidia/generative-virtual-screening-for-drug-discovery/blueprintcard says the minimum requirement is 1xA6000, A100, or H100. I was able to run the notebook on 1xA10 (g5.16xlarge AWS) with 24GB GPU memory. Just FYI.

Thanks very again for the effort of putting this together!

jtcasablanca commented 2 weeks ago

Minimum GPU requirements should be in the readme in this repo in case users get here without first passing through the build.nvidia.com page.