#MLOps with Hugging Face Spaces and Dagger

1 messages · Page 1 of 1 (latest)

tropic radish
#

https://www.youtube.com/watch?v=CoHjpBQLp98 by @tight talon

YouTube
Dagger
Community Call Demo: MLOps with Hugging Face Spaces and Dagger

In this demo, Sam Alba, co-founder of Dagger, demonstrates how to build machine learning apps using Hugging Face and Dagger. The app showcased is a text summarization tool that utilizes a popular pre-trained model from Hugging Face's model hub. Sam explains the code and pipelines involved, including testing the model's performance and deploying ...

â–¶ Play video
#

@median lance for all the MLOps fans in the house 😉

vivid hollow
#

great, any followup about GPU access?

tight talon
# vivid hollow great, any followup about GPU access?

Yes, some discussions on the API for allocating memory on a given GPU. The idea would be to start by exposing a simple container.WithGPU(max_memory=XXX). There are ideas for dagger to help with the multi-tenancy of GPU(s). Do you have a opinions there?

vivid hollow
#

well, our use case only require 1 GPU at the moment. but it's worth noting that this may be more complex than just adjusting the memory.
according to this, https://cloud.google.com/kubernetes-engine/docs/concepts/gpus#features, "By default, Kubernetes only supports assigning GPUs as whole units to containers". So someone / something needs to be aware of the chosen strategy (time sharing vs multi-instance). Sadly I don't have experience outside of the GCP/GKE arena

Google Cloud
About GPUs in GKE  |  Google Kubernetes Engine (GKE)  |  Google Clo...

Learn about using hardware accelerators in your GKE clusters

tropic radish
vivid hollow
#

I signed up already 🙂

#

Though I might not be able to make it

#

Do you know the time?

tropic radish
#

It is at 9 am PT. We will have it recorded though, so you will get a link to the recording via email.

vivid hollow
#

nvm, found it !