Needed RAM for the full bloomz model?

#38
by luxianos - opened

Hi, how much RAM memory would be needed to load the full bloomz model? I'm trying different aws instance types configurations but at least 768GB still is not enough.

@luxianos did you find a solution to this? I'm trying to deploy it on AWS as well and keep getting errors.

BigScience Workshop org

I generally use 8 * 80GB A100s, which works fine with 8-way pipeline parallelism, so 768GB should be enough, maybe you're not loading it in bfloat16?

@luxianos did you find a solution to this? I'm trying to deploy it on AWS as well and keep getting errors.

I found around 800-850GB was enough

Sign up or log in to comment