Measured runtime memory usage (pyTorch)

#6
by HexGrimm - opened

The model card lists the 37.53 MB weight footprint but not runtime memory, so I measured
the process. Numbers below are from Windows 11, RTX 4080 SUPER, Python 3.13,
torch 2.10.0+cu128, device="cuda".

Stage Process RSS VRAM peak-alloc VRAM reserved
import torch only 479 MB 0 0
model loaded 715 MB 39 MB 46 MB
3 s clip 1344 MB 87 MB 110 MB
16 s clip 1391 MB 260 MB 380 MB
125 s clip 1425 MB 262 MB 558 MB

Live VRAM after each synthesis settles at ~44 MB. CPU-only (device="cpu"): 480 MB after
import, 636 MB after load, 916 MB at 16 s, 1089 MB at 125 s.

Sign up or log in to comment