Hacker News Reader

Comment by mark_l_watson | Hacker News Reader

I manage a ML team and the article has several ‘borrowable’ ideas, the primary idea supporting model builders working on their local Linux device and simplifying the process of transferring locally built models to a scalable platform. If we were starting today to implement our in house system I might just use Ubber’s system as a checklist for features to implement.

Not too off topic: after mostly building and training models on remote compute (both at work and for my at home learning projects), I recently bought a fast laptop for home use that has a 1070 GPU and the experience of using a local device with a GPU makes me regret many years of SSHing to remote compute or using Jupyter remotely for machine learning development. I mention this as a lesson learned that may be helpful.

View on HN

The local dev painpoint is real. I was working on a library that can intercept and transparently offload keras fit() calls from ones local machine to a faster remote machine. Unfortunately, it seems most people just use GPU workstations and don’t need the speed improvements locally.

I recently bought a fast laptop for home use that has a 1070 GPU and the experience of using a local device with a GPU makes me regret many years of SSHing to remote compute

Most interesting problems in my experience are constrained by I/O bandwidth, and being able to run C/GPUs for extended periods without thermal issues. You really need something in a DC for that. I chose not to bother with a GPU laptop myself because there just seemed to be no point, for ML at least. I guess just for playing, it’s a nice luxury, but I can’t imagine relying on it for real work.

Gaming desktop PCs exist for a single reason - "being able to run C/GPUs for extended periods without thermal issues." If you were doing this for actual work, you'd obviously need to step upwards to a real workstation, which would actually be built quite similarly to a bog-standard gaming desktop save for using higher reliability components, e.g. FirePro- or Quadro-class GPU hardware. But to suggest that you need "something in a DC" to explore interesting problems seems quite wrong to me.

Uber's software doesn't exist in the public realm, and so there is nothing to learn from it. It is imaginary fluff. I have posted a list of the ones that do actually exist in another comment. And you should've at least got a 1080.

I wanted a portable laptop and just saw 1070s on what was available. What brand of GPU laptop do you suggest?

I got some ideas from reading the article, but yes, I agree with you that the article should have a link to a public git repo for at least some of their tooling.

EDIT: I wanted the multi-GPU Lambda Labs box, but with my lifestyle I require some mobility. I agree that the 1070 is a bit weak, but is still better than no GPU and I like The System76 laptop I bought even when I am not using the GPU.

You will get a lot more ideas from other software that actually do exist. I can easily write a fake blog post with fake thoughts giving you fake ideas, but that's not what you want; you deserve ideas that work.

System76 does sell laptops with the 1080.

It actually gives a lot of idea and validation for concepts that I was thinking about for my project https://github.com/polyaxon/polyaxon, namely for going from trained model to serving those models on other platforms in a seamless way. I also think that high-level articles like this one are really good for people who are trying to build internal tools around ML/AI ops, people can get more or less ideas about what and how other companies are managing their ML pipelines and how they impact the productivity of their teams.

You should be getting ideas from actual open source projects that actually exist, not from imaginary/fake ones. There are numerous serving packages that actually exist and offer ideas.

Reply on news.ycombinator.com