github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 08:02:14 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #4995] Ollama GPU not loding properly #3159

New Issue
Closed
opened 2026-04-12 13:38:38 -05:00 by GiteaMirror · 8 comments
GiteaMirror commented 2026-04-12 13:38:38 -05:00
Owner
Copy Link

Originally created by @tankvpython on GitHub (Jun 12, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/4995

Originally assigned to: @jmorganca on GitHub.

What is the issue?

I am facing an issue with the Ollama service. I have an RTX 4090 GPU with 80GB of RAM and 24GB of VRAM. When I run the Llama 3 70B model and ask it a question, it initially loads on the GPU, but after 5-10 seconds, it shifts entirely to the CPU. This causes the response time to be slow. Please provide me with a solution for this. Thank you in advance.

Note:- GPU load is 6-12 % and CPU load is 70% .

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

v0.1.43

Originally created by @tankvpython on GitHub (Jun 12, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/4995 Originally assigned to: @jmorganca on GitHub. ### What is the issue? I am facing an issue with the Ollama service. I have an **RTX 4090 GPU** with 80GB of RAM and 24GB of VRAM. When I run the Llama 3 70B model and ask it a question, it initially loads on the GPU, but after 5-10 seconds, it shifts entirely to the CPU. This causes the response time to be slow. Please provide me with a solution for this. Thank you in advance. Note:- GPU load is 6-12 % and CPU load is 70% . ### OS Windows ### GPU Nvidia ### CPU Intel ### Ollama version v0.1.43
GiteaMirror added the questionperformance labels 2026-04-12 13:38:39 -05:00
GiteaMirror commented 2026-04-12 13:38:41 -05:00
Author
Owner
Copy Link

@phong-phuong commented on GitHub (Jun 12, 2024):

Your GPU has too little VRAM to run a 70 billion parameter model entirely on the GPU. Try a smaller model like a 22 billion parameter model.

<!-- gh-comment-id:2162542743 --> @phong-phuong commented on GitHub (Jun 12, 2024): Your GPU has too little VRAM to run a 70 billion parameter model entirely on the GPU. Try a smaller model like a 22 billion parameter model.
GiteaMirror commented 2026-04-12 13:38:42 -05:00
Author
Owner
Copy Link

@tankvpython commented on GitHub (Jun 12, 2024):

Your GPU has too little VRAM to run a 70 billion parameter model entirely on the GPU. Try a smaller model like a 22 billion parameter model.

if you have any idea what is configuration needed ?

<!-- gh-comment-id:2162573179 --> @tankvpython commented on GitHub (Jun 12, 2024): > Your GPU has too little VRAM to run a 70 billion parameter model entirely on the GPU. Try a smaller model like a 22 billion parameter model. if you have any idea what is configuration needed ?
GiteaMirror commented 2026-04-12 13:38:43 -05:00
Author
Owner
Copy Link

@phong-phuong commented on GitHub (Jun 12, 2024):

See https://huggingface.co/TheBloke/Llama-2-70B-Chat-GPTQ/discussions/2
A user used a 48GB card to run a 70b model.

<!-- gh-comment-id:2162589151 --> @phong-phuong commented on GitHub (Jun 12, 2024): See https://huggingface.co/TheBloke/Llama-2-70B-Chat-GPTQ/discussions/2 A user used a 48GB card to run a 70b model.
GiteaMirror commented 2026-04-12 13:38:44 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Jun 12, 2024):

Hi @tankvpython sorry about this – I have a similar system and it should definitely be partially loading the 70B model to GPU – would it be possible to check with ollama ps after loading a model?

<!-- gh-comment-id:2163367364 --> @jmorganca commented on GitHub (Jun 12, 2024): Hi @tankvpython sorry about this – I have a similar system and it should definitely be partially loading the 70B model to GPU – would it be possible to check with `ollama ps` after loading a model?
GiteaMirror commented 2026-04-12 13:38:44 -05:00
Author
Owner
Copy Link

@tankvpython commented on GitHub (Jun 13, 2024):

Hi @tankvpython sorry about this – I have a similar system and it should definitely be partially loading the 70B model to GPU – would it be possible to check with ollama ps after loading a model?

Yes i am able to do this i will check it

<!-- gh-comment-id:2164668881 --> @tankvpython commented on GitHub (Jun 13, 2024): > Hi @tankvpython sorry about this – I have a similar system and it should definitely be partially loading the 70B model to GPU – would it be possible to check with `ollama ps` after loading a model? Yes i am able to do this i will check it
GiteaMirror commented 2026-04-12 13:38:45 -05:00
Author
Owner
Copy Link

@phong-phuong commented on GitHub (Jun 13, 2024):

If you are using the standard ollama3:70b version, that one is 4bit quant and uses ~33GB VRAM which is a lot more than your GPU can handle. When that happens, the rest of the model gets offloaded to the CPU which is much slower.
If you want faster interference times, you can try a lower quant like the 2bit one which uses ~17GB, much more suitable for a 24GB card : ollama run llama3:70b-instruct-q2_K.
Just bear in mind, that lowering the quant size may degrade the quality of the answers (think of it like compressing an image or video).

<!-- gh-comment-id:2166299807 --> @phong-phuong commented on GitHub (Jun 13, 2024): If you are using the standard ollama3:70b version, that one is 4bit quant and uses ~33GB VRAM which is a lot more than your GPU can handle. When that happens, the rest of the model gets offloaded to the CPU which is much slower. If you want faster interference times, you can try a lower quant like the 2bit one which uses ~17GB, much more suitable for a 24GB card : `ollama run llama3:70b-instruct-q2_K`. Just bear in mind, that lowering the quant size may degrade the quality of the answers (think of it like compressing an image or video).
GiteaMirror commented 2026-04-12 13:38:45 -05:00
Author
Owner
Copy Link

@tankvpython commented on GitHub (Jun 13, 2024):

If you are using the standard ollama3:70b version, that one is 4bit quant and uses ~33GB VRAM which is a lot more than your GPU can handle. When that happens, the rest of the model gets offloaded to the CPU which is much slower. If you want faster interference times, you can try a lower quant like the 2bit one which uses ~17GB, much more suitable for a 24GB card : ollama run llama3:70b-instruct-q2_K. Just bear in mind, that lowering the quant size may degrade the quality of the answers (think of it like compressing an image or video).

thanks for help i don't need light weight model because of i need more accuracy

<!-- gh-comment-id:2166550159 --> @tankvpython commented on GitHub (Jun 13, 2024): > If you are using the standard ollama3:70b version, that one is 4bit quant and uses ~33GB VRAM which is a lot more than your GPU can handle. When that happens, the rest of the model gets offloaded to the CPU which is much slower. If you want faster interference times, you can try a lower quant like the 2bit one which uses ~17GB, much more suitable for a 24GB card : `ollama run llama3:70b-instruct-q2_K`. Just bear in mind, that lowering the quant size may degrade the quality of the answers (think of it like compressing an image or video). thanks for help i don't need light weight model because of i need more accuracy
GiteaMirror commented 2026-04-12 13:38:45 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Sep 5, 2024):

I think we can close this one as it seems the system is behaving as intended. ollama ps will show the model is loading split between CPU and GPU, which will result in slower performance, bottlenecked by the CPU.

<!-- gh-comment-id:2332718308 --> @dhiltgen commented on GitHub (Sep 5, 2024): I think we can close this one as it seems the system is behaving as intended. `ollama ps` will show the model is loading split between CPU and GPU, which will result in slower performance, bottlenecked by the CPU.
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 performance 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#3159
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?