Is this what my approval should look like? Perhaps it takes time to… register?

Thanks again for helping me out on this. I’ve been stuck on trying to provision a GPU for weeks and feel closer then ever lol

This is my compute environment

Oh shoot I thought the 4 referred to instances, but stands for vCPUs. This would make sense why it’s not working, since the minimum vCPUs is set to 8 (and the minimum vCPU is out of the three compute options is 8 Let me request another increase.

For reference, I requested 256 and they approved it as fast as when they approve 4 on a new account. Then you don't need to play this game anymore, and since it is batch workloads there isn't a big risk of leaving many / expensive GPUs on for a long time. Just make sure to check the billing console to catch any stray GPUs!

I have noticed something interesting. Given that I currently only have the 4 vCPU limit, I decided to try to use the p2.xlarge instance, since this is the only p instance with 4 vCPUs. Now the job gets started and finishes, however I get “GPU is not available :(“. I tried using TensorFlow with the appropriate TensorFlow GPU image from Docker and I still got 0 GPUs being used. I have confirmed the correct job queue and compute env.

Could possibly be due to MinvCPUBatch being set to 8, when the p2.xlarge only has 4 vCPUs, I’ll keep tinkering. Let me know if you’ve run into this.

The pytorch one in the repo had the same issue

Update: They approved me for 8 vCPUs, which is just good enough to provision a p3.2xlarge instance. This worked! Perhaps the p2 instances aren’t CUDA enabled? There’s no mention of CUDA in the description of P2. Anyways, thanks for all of your help!