github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 16:40:08 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #11917] 2*4090---The highest GPU load rate can only reach 50% #7910

New Issue
Closed
opened 2026-04-12 20:04:34 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-12 20:04:34 -05:00
Owner
Copy Link

Originally created by @niboliang on GitHub (Aug 15, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/11917

What is the issue?

Configure 2 GTX4090 graphics cards on Ubuntu servers and use load balancing，Each graphics card

Image

GPU load rate can only reach 50%

Relevant log output

OS

Docker

GPU

Nvidia

CPU

Intel

Ollama version

0.9.2

Originally created by @niboliang on GitHub (Aug 15, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/11917 ### What is the issue? Configure 2 GTX4090 graphics cards on Ubuntu servers and use load balancing，Each graphics card <img width="528" height="233" alt="Image" src="https://github.com/user-attachments/assets/c987dc02-55f5-4cfc-8ee9-f8d979c08526" /> GPU load rate can only reach 50% ### Relevant log output ```shell ``` ### OS Docker ### GPU Nvidia ### CPU Intel ### Ollama version 0.9.2
GiteaMirror added the bug label 2026-04-12 20:04:34 -05:00
GiteaMirror commented 2026-04-12 20:04:36 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Aug 15, 2025):

Model layers are processed sequentially in ollama. GPU-0 gets one half of the layers. GPU-1 gets the other half. When a single inference starts, GPU-0 runs at 100% doing inference with the first half of the model layers, while GPU-1 is idle. When GPU-0 is done, the CPU transfers the intermediate results to GPU-1, which then runs at 100% over the second half of the layers while GPU-0 is idle. This means that on average for a single inference, each GPU is used half of the time. Generally, %usage = 100/number_of_devices.

<!-- gh-comment-id:3191215181 --> @rick-github commented on GitHub (Aug 15, 2025): Model layers are processed sequentially in ollama. GPU-0 gets one half of the layers. GPU-1 gets the other half. When a single inference starts, GPU-0 runs at 100% doing inference with the first half of the model layers, while GPU-1 is idle. When GPU-0 is done, the CPU transfers the intermediate results to GPU-1, which then runs at 100% over the second half of the layers while GPU-0 is idle. This means that on average for a single inference, each GPU is used half of the time. Generally, %usage = 100/number_of_devices.
GiteaMirror commented 2026-04-12 20:04:37 -05:00
Author
Owner
Copy Link

@niboliang commented on GitHub (Aug 16, 2025):

Model layers are processed sequentially in ollama. GPU-0 gets one half of the layers. GPU-1 gets the other half. When a single inference starts, GPU-0 runs at 100% doing inference with the first half of the model layers, while GPU-1 is idle. When GPU-0 is done, the CPU transfers the intermediate results to GPU-1, which then runs at 100% over the second half of the layers while GPU-0 is idle. This means that on average for a single inference, each GPU is used half of the time. Generally, %usage = 100/number_of_devices.

Does this mean that ollama does not support dual-card inference, so no matter how it is set up, the GPU usage rate will only be 50%? Thank you.

<!-- gh-comment-id:3193430746 --> @niboliang commented on GitHub (Aug 16, 2025): > Model layers are processed sequentially in ollama. GPU-0 gets one half of the layers. GPU-1 gets the other half. When a single inference starts, GPU-0 runs at 100% doing inference with the first half of the model layers, while GPU-1 is idle. When GPU-0 is done, the CPU transfers the intermediate results to GPU-1, which then runs at 100% over the second half of the layers while GPU-0 is idle. This means that on average for a single inference, each GPU is used half of the time. Generally, %usage = 100/number_of_devices. Does this mean that ollama does not support dual-card inference, so no matter how it is set up, the GPU usage rate will only be 50%? Thank you.
GiteaMirror commented 2026-04-12 20:04:38 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 1, 2025):

Does this mean that ollama does not support dual-card inference, so no matter how it is set up, the GPU usage rate will only be 50%?

For a single generation, yes. If the clients send multiple concurrent requests, GPU usage will be higher.

<!-- gh-comment-id:3243176457 --> @rick-github commented on GitHub (Sep 1, 2025): > Does this mean that ollama does not support dual-card inference, so no matter how it is set up, the GPU usage rate will only be 50%? For a single generation, yes. If the clients send multiple concurrent requests, GPU usage will be higher.
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 bug
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#7910
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?