Thanks. Actually, I start discussing this idea around and supporting Jupiter notebook is the first step towards serverless for AI. I did a post on Linkedin and I found enthusiastic support about the idea. Serverless fits very well with AI, apparently.
The second step would be to be able to run action using a GPU for executing models. This requires that we can "tag" some nodes to have a GPU attached and some runtimes to be executed only in those nodes. Looks like in the community some work to tag actions to be run on GPU nodes had been already done so I asked them to submit the patch. Hopefully, it will be out soon otherwise I will try to do work on that too... -- Michele Sciabarra mich...@sciabarra.com ----- Original message ----- From: Dascalita Dragos <ddrag...@gmail.com> To: dev@openwhisk.apache.org Subject: Re: Proposal: Adding Support for Juypter Notebooks Date: Monday, October 26, 2020 5:30 AM Michele, this is awesome. I'm so appreciative for building on the existing "pythonaiaction". Thanks for updating it to the latest Tensorflow. I think there's a lot of value in running AI Actions in a serverless fashion to quickly build AI APIs (inference mainly; training is a separate topic that's worth investigating; I mean training batches in parallel, using AI actions). For inference however teams can experiment with many different AI models in a very (cost) efficient way. Paying only for what you use is what makes the serverless execution model so attractive for AI. Kudos for moving this ahead ! A BIG thank you ! On Sat, Oct 24, 2020 at 2:21 AM Michele Sciabarra <mich...@sciabarra.com> wrote: > I implemented support for Jupyter Notebooks for a Python 3.8 runtime (that > I temporarily published as actionloop/action-python-v3.8) as you can see > here: > > https://imgur.com/dTjj6UJ > > It was easy as I leveraged the "compilation" script of Actionloop that is > written in Python and also the nbconvert library that is also Python-based. > > However since I believe this is a useful feature, and since Jupyter > Notebooks are available for all the languages I was thinking to make it a > standard feature of the Go Proxy (aka AcionLoop) so it is inherited by all > the runtimes based on ActionLoop (Python, PHP, Rust, Go, and the flavors of > Java and Node that uses it). > > So what the community thinks? Is widespread support for Jupyter Notebook > as a "source" format, that I detect and handle the init level and to > extract the code from it, a worthy addition? > > -- > Michele Sciabarra > mich...@sciabarra.com >
