github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 00:51:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #6914] Work done by CPU instead of GPU #50885

New Issue
Closed
opened 2026-04-28 17:19:55 -05:00 by GiteaMirror · 7 comments
GiteaMirror commented 2026-04-28 17:19:55 -05:00
Owner
Copy Link

Originally created by @Iliceth on GitHub (Sep 23, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/6914

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

I'm aware this might not be a bug, but I'm trying to understand and figure out if I can and/or should change something. When I try Reflection 70b.

CPU: 8 cores 100% utilization
RAM: 23 of 32 GB in use
GPU: on average 5% utilization
VRAM: 23 of 24 GB in use

I assume the model does not fit in VRAM and therefor the spread across VRAM and RAM, I'm fine with that. But the fact that the CPU seems to be expected to do almost all of the lifting seems odd to me, although it might be normal.

OS

Linux

GPU

Nvidia

CPU

Intel

Ollama version

0.3.11

Originally created by @Iliceth on GitHub (Sep 23, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/6914 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? I'm aware this might not be a bug, but I'm trying to understand and figure out if I can and/or should change something. When I try Reflection 70b. CPU: 8 cores 100% utilization RAM: 23 of 32 GB in use GPU: on average 5% utilization VRAM: 23 of 24 GB in use I assume the model does not fit in VRAM and therefor the spread across VRAM and RAM, I'm fine with that. But the fact that the CPU seems to be expected to do almost all of the lifting seems odd to me, although it might be normal. ### OS Linux ### GPU Nvidia ### CPU Intel ### Ollama version 0.3.11
GiteaMirror added the question label 2026-04-28 17:19:55 -05:00
GiteaMirror commented 2026-04-28 17:19:58 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 23, 2024):

The CPU is much slower at doing inference than the GPU, so more time is spent in computation in the CPU than the GPU.

<!-- gh-comment-id:2368027755 --> @rick-github commented on GitHub (Sep 23, 2024): The CPU is much slower at doing inference than the GPU, so more time is spent in computation in the CPU than the GPU.
GiteaMirror commented 2026-04-28 17:19:58 -05:00
Author
Owner
Copy Link

@Iliceth commented on GitHub (Sep 23, 2024):

That I get, but can I force to only use GPU or does it have to do part of the inference on the CPU because system RAM is used additionally to the GPU's VRAM?

<!-- gh-comment-id:2368300110 --> @Iliceth commented on GitHub (Sep 23, 2024): That I get, but can I force to only use GPU or does it have to do part of the inference on the CPU because system RAM is used additionally to the GPU's VRAM?
GiteaMirror commented 2026-04-28 17:19:58 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 23, 2024):

Yes, it's because part of the model resides in system RAM. While it's technically possible for the GPU to access system RAM to do inference, it is still more efficient to have the CPU do inference on the weights in system RAM and have the GPU do inference on the weights in VRAM. If you want to experiment, you can set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 in the server environment and then use num_gpu in an API call or /set parameter num_gpu in the ollama cli to override the number of layers that ollama offloads to the GPU. I found that on my system (linux/4070) having the GPU do all inference across RAM and VRAM was extremely slow compared to the hybrid approach.

<!-- gh-comment-id:2368359345 --> @rick-github commented on GitHub (Sep 23, 2024): Yes, it's because part of the model resides in system RAM. While it's technically possible for the GPU to access system RAM to do inference, it is still more efficient to have the CPU do inference on the weights in system RAM and have the GPU do inference on the weights in VRAM. If you want to experiment, you can set `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` in the server environment and then use `num_gpu` in an API call or `/set parameter num_gpu` in the ollama cli to override the number of layers that ollama offloads to the GPU. I found that on my system (linux/4070) having the GPU do all inference across RAM and VRAM was extremely slow compared to the hybrid approach.
GiteaMirror commented 2026-04-28 17:19:59 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Sep 25, 2024):

@Iliceth if you have further questions, let us know. In short, running a smaller model that fits 100% in your GPU will yield the best performance.

<!-- gh-comment-id:2372636839 --> @dhiltgen commented on GitHub (Sep 25, 2024): @Iliceth if you have further questions, let us know. In short, running a smaller model that fits 100% in your GPU will yield the best performance.
GiteaMirror commented 2026-04-28 17:19:59 -05:00
Author
Owner
Copy Link

@sarjil77 commented on GitHub (Sep 25, 2024):

i nvidia-rtx A6000 but when i am running my ollama model and on checking with ps command it is showing me that it is running on 100% cpu. infact there is no process running on the system, this issue i was getting before also but at that time when i uninstalled it and on reinstalling it worked, but currently i am facing the same issue but not able to solve it.

please respond ASAP

Thank you in advance

<!-- gh-comment-id:2373171212 --> @sarjil77 commented on GitHub (Sep 25, 2024): i nvidia-rtx A6000 but when i am running my ollama model and on checking with ps command it is showing me that it is running on 100% cpu. infact there is no process running on the system, this issue i was getting before also but at that time when i uninstalled it and on reinstalling it worked, but currently i am facing the same issue but not able to solve it. please respond ASAP Thank you in advance
GiteaMirror commented 2026-04-28 17:20:00 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 25, 2024):

Server logs will help in debugging.

<!-- gh-comment-id:2373183064 --> @rick-github commented on GitHub (Sep 25, 2024): [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) will help in debugging.
GiteaMirror commented 2026-04-28 17:20:00 -05:00
Author
Owner
Copy Link

@Iliceth commented on GitHub (Sep 25, 2024):

@rick-github @dhiltgen I will try out your suggestion, thanks. I am very happy with how ollama performs with smaller models, it's just tempting to play with the big one's... :-p

<!-- gh-comment-id:2373475225 --> @Iliceth commented on GitHub (Sep 25, 2024): @rick-github @dhiltgen I will try out your suggestion, thanks. I am very happy with how ollama performs with smaller models, it's just tempting to play with the big one's... :-p
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label question
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#50885
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?