← all posts site ↗

photo ai and a voice assistant sharing four gigs

the speech model holds most of a four gig card, and the photo manager also wants the gpu for machine learning. i assumed they would fight and planned a big carve out. then i checked the real state and the conflict was much smaller than it looked.

the photo ai holds zero vram when idle, its container had been up for days with no restarts and no oom kill, and its nightly job only touches the card for about ninety seconds in the middle of the night. that is the only window where both systems want the gpu, and i am asleep then.

still tuned it: pin only the model search actually uses so every search is warm instead of paying a one to two second cold load after five idle minutes, leave the arena off because it doubles the video memory buffers, and leave the nightly job alone. added a floor alarm so a collision is loud instead of a mystery slowdown. the escape hatch stays: flip the photo ai machine learning over to cpu if it ever fills the card.