Hi, I am currently evaluating HF Spaces with docker and FastAPI for my projects. I ran into the problem that some of the models were to large for the free tier and I am trying to figure if the PPRO subscription adds memory and if so how much and if e.g. meta-llama/Meta-Llama-3.1-8B would fit after the subscription. There is not much info about this, or I overlooked it. Many thanks
The free space is clearly marked as having 16GB of RAM, but I don’t think the Zero GPU space available in Pro has that marked.
I happen to know this because I’ve been trying various things for one reason or another, but it appears from python that the amount of RAM itself, including page files, has not changed between the free space and the Zero GPU space.
However, I am not sure if it is a benefit of having a GPU, or actually slightly more CPU and RAM, or why for sure, but there is a phenomenon where programs unrelated to the GPU do not run in CPU space, but do run in Zero GPU space.
However, as long as it is not explicitly stated, it would be wise to consider it as an “extra”.
Of course, the difference from free space when using GPU and VRAM is obvious: Quota is too tight to use casually, but it is blazing fast during the 30 to 120 seconds when GPU is available.
Hi,
See Hugging Face – The AI community building the future. for details, a Pro subscription provides access to ZeroGPU (which are serverless GPUs - namely A100s which have 80GB of RAM).
