Milabench Size

[Delaunay]

• llama3 - 70B is big (132G)
• We could generate it on the fly #250
  • One possible issue is that it generate 8 checkpoints for 8 GPUs
  • Can we load the model on 4 GPUs ?

2.5M    ./cache
325M    ./data/processed
263G    ./data/llama3_70B        # <= Duplicated weights
252M    ./data/raw
15G     ./data/llama3_8B
25G     ./data/FakeImageNet
304G    ./data
2.8M    ./runs
5.7G    ./venv/gnn                    # <= gnn need their own pytorch
6.4G    ./venv/torch           
13G     ./venv
316G    .