N/A

I honestly don't see much of an issue here with only the screenshots you have provided, yes there is a spike, but it does settle back down to a lower level than what it spiked to, for something that does music source separation with deep learning likely completely in memory that is not a crazy amount of memory at all

@Brody I tried do memory profiling with guppy in the RQ worker, but I don't see anything unusual here either

@Brody any other ideas on how I could make requests more cost-effective? $1.60 per request does not seem very scalable to me, but I don't know what the industry standard is either, so maybe I'm being ignorant

I mean, I didn't think deep learning was really cost effective to begin with?

I think maybe that the initial memory spike is actually caused by the RQ worker downloading Demucs at runtime

I'm just baking the model into my Docker image

sounds possible

@Brody that was the issue 😅

just a friendly reminder that pinging is against the rules

oh sorry

thx for answering my questions

no problem