[GH-ISSUE #5773] Favor idle GPUs that fit over largest free memory GPUs when scheduling #29355

Open
opened 2026-04-22 08:08:34 -05:00 by GiteaMirror · 0 comments
Owner

Originally created by @dhiltgen on GitHub (Jul 18, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/5773

Originally assigned to: @dhiltgen on GitHub.

The scheduler today tries to find a single GPU to run a model based on the largest amount of free VRAM, but on multi-GPU setups where one GPU is significantly larger than others, this can lead to smaller models clumping on the largest GPU. The algorithm should be a bit smarter to try to find an idle GPU first to avoid this clumping behavior.

Originally created by @dhiltgen on GitHub (Jul 18, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/5773 Originally assigned to: @dhiltgen on GitHub. The scheduler today tries to find a single GPU to run a model based on the [largest amount of free VRAM](https://github.com/ollama/ollama/blob/main/server/sched.go#L690-L693), but on multi-GPU setups where one GPU is significantly larger than others, this can lead to smaller models clumping on the largest GPU. The algorithm should be a bit smarter to try to find an idle GPU first to avoid this clumping behavior.
GiteaMirror added the feature request label 2026-04-22 08:08:34 -05:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#29355