[GH-ISSUE #10351] Incorrect memory allocation #53311

Closed
opened 2026-04-29 02:35:56 -05:00 by GiteaMirror · 2 comments
Owner

Originally created by @bitcandy on GitHub (Apr 20, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10351

What is the issue?

Incorrect memory allocation: only 50% of the video card memory is being utilized, which is especially noticeable on the latest model gemma3:27b-it-qat. On other models, around 1-3 gigabyte of memory often remains unused, leading ollama to start using RAM

gemma3:27b-it-qat 29eb0b9aeda3 27 GB 10%/90% CPU/GPU 59 minutes from now

|=========================================+========================+======================|
| 0% 58C P2 83W / 195W | 5267MiB / 8192MiB | 0% Default |
+-----------------------------------------+------------------------+----------------------+
| 0% 61C P2 75W / 280W | 5275MiB / 11264MiB | 0% Default |
+-----------------------------------------+------------------------+----------------------+
| 0% 58C P2 38W / 195W | 4981MiB / 8192MiB | 0% Default |
+-----------------------------------------+------------------------+----------------------+

On other start it actually utilize less then that....

Relevant log output


OS

Linux

GPU

Nvidia

CPU

No response

Ollama version

0.6.5

Originally created by @bitcandy on GitHub (Apr 20, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10351 ### What is the issue? Incorrect memory allocation: only 50% of the video card memory is being utilized, which is especially noticeable on the latest model gemma3:27b-it-qat. On other models, around 1-3 gigabyte of memory often remains unused, leading ollama to start using RAM gemma3:27b-it-qat 29eb0b9aeda3 27 GB 10%/90% CPU/GPU 59 minutes from now |=========================================+========================+======================| | 0% 58C P2 83W / 195W | 5267MiB / 8192MiB | 0% Default | +-----------------------------------------+------------------------+----------------------+ | 0% 61C P2 75W / 280W | 5275MiB / 11264MiB | 0% Default | +-----------------------------------------+------------------------+----------------------+ | 0% 58C P2 38W / 195W | 4981MiB / 8192MiB | 0% Default | +-----------------------------------------+------------------------+----------------------+ On other start it actually utilize less then that.... ### Relevant log output ```shell ``` ### OS Linux ### GPU Nvidia ### CPU _No response_ ### Ollama version 0.6.5
GiteaMirror added the bug label 2026-04-29 02:35:56 -05:00
Author
Owner

@rick-github commented on GitHub (Apr 20, 2025):

https://github.com/ollama/ollama/issues/10041#issuecomment-2816399723

<!-- gh-comment-id:2817318662 --> @rick-github commented on GitHub (Apr 20, 2025): https://github.com/ollama/ollama/issues/10041#issuecomment-2816399723
Author
Owner

@bitcandy commented on GitHub (Apr 20, 2025):

at Ollama 0.6.6 it is little bit better, it use 5803MiB

<!-- gh-comment-id:2817324821 --> @bitcandy commented on GitHub (Apr 20, 2025): at Ollama 0.6.6 it is little bit better, it use 5803MiB
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#53311