Some interesting examples, eg. The sentiment analysis in combination with email routing.
I'm curious about the models and GPU support too. I'm guessing this is currently beta and they are awaiting popularity of this.
The simple ( & cost-efficient) method would be to support CPU-first and then up the scale accordingly. No harm in that, but it could be made more clear regarding the roadmap.
Edit: during private beta there is support for <10 mb. Models.
7 comments
[ 2.6 ms ] story [ 29.3 ms ] threadI'm curious about the models and GPU support too. I'm guessing this is currently beta and they are awaiting popularity of this.
The simple ( & cost-efficient) method would be to support CPU-first and then up the scale accordingly. No harm in that, but it could be made more clear regarding the roadmap.
Edit: during private beta there is support for <10 mb. Models.
https://developers.cloudflare.com/constellation/platform/dat...
Discussions @ https://discord.gg/cloudflaredev
I don't see how this makes Cloudflare interesting for AI. I suspect the same problems will exist.
This isn't so different than running onnx in a worker without constellation, and using KV directly?
The negative is it's such a closed ecosystem.
You're thinking of years ago when the CPU time limit was 50ms. These days the limit is 30 seconds, plenty of time for these tasks.
I hit the CPU time limit a few months ago. I was paying for unbound and my account default was on. I guess I also hit the memory limit (128MB).
Perhaps we should reply on: https://community.cloudflare.com/t/long-running-webcrypto-ap... and https://community.cloudflare.com/t/options-for-password-hash.... Though I'm not sure if it will work, certain algorithms might take more than 128MB.
Thanks :)
Edit: I just found https://blog.cloudflare.com/d1-turning-it-up-to-11/ so I'll read that first