github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #1288] Ollama with multiple GPUs #47175

New Issue
Closed
opened 2026-04-28 03:24:33 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-28 03:24:33 -05:00
Owner
Copy Link

Originally created by @technovangelist on GitHub (Nov 27, 2023).
Original GitHub issue: https://github.com/ollama/ollama/issues/1288

Originally assigned to: @dhiltgen on GitHub.

If you are running ollama on a machine with multiple GPUs, inference will be slower than the same machine with one gpu but it will still be faster than the same machine with no gpu. The benefit of multiple GPUs is access to more video memory, allowing for larger models or more of the model to be processed by the GPU.

BUT if you have enough video memory on the first gpu, we should use only the one gpu, to ensure that perf is as fast as possible. Otherwise it is slower for no good reason.

And if possible, it would be great to identify the faster gpu and use that first.

Originally created by @technovangelist on GitHub (Nov 27, 2023). Original GitHub issue: https://github.com/ollama/ollama/issues/1288 Originally assigned to: @dhiltgen on GitHub. If you are running ollama on a machine with multiple GPUs, inference will be slower than the same machine with one gpu but it will still be faster than the same machine with no gpu. The benefit of multiple GPUs is access to more video memory, allowing for larger models or more of the model to be processed by the GPU. BUT if you have enough video memory on the first gpu, we should use only the one gpu, to ensure that perf is as fast as possible. Otherwise it is slower for no good reason. And if possible, it would be great to identify the faster gpu and use that first.
GiteaMirror commented 2026-04-28 03:24:34 -05:00
Author
Owner
Copy Link

@peteygao commented on GitHub (Nov 28, 2023):

This is because the model checkpoint synchronisation is dependent on the slowest GPU running in the cluster.

So you're correct, you can utilise increased VRAM distributed across all the GPUs, but the inference speed will be bottlenecked by the speed of the slowest GPU.

This should be a separate feature request: Specifying which GPUs to use when there are multiple GPUs attached to the machine.

<!-- gh-comment-id:1829054243 --> @peteygao commented on GitHub (Nov 28, 2023): This is because the model checkpoint synchronisation is dependent on the slowest GPU running in the cluster. So you're correct, you can utilise increased VRAM distributed across all the GPUs, but the inference speed will be bottlenecked by the speed of the _slowest_ GPU. This should be a separate feature request: Specifying which GPUs to use when there are multiple GPUs attached to the machine.
GiteaMirror commented 2026-04-28 03:24:35 -05:00
Author
Owner
Copy Link

@technovangelist commented on GitHub (Nov 28, 2023):

I hadn't really thought through about a machine with different kinds of gpus, though that is what my last sentence is about. But mostly i was thinking about when i have 2-4 t4's or a bunch of a100's attached that are all identical. In the discord however it was suggested that someone wanted to dedicate one gpu to work projects and another to Ollama.

<!-- gh-comment-id:1830006005 --> @technovangelist commented on GitHub (Nov 28, 2023): I hadn't really thought through about a machine with different kinds of gpus, though that is what my last sentence is about. But mostly i was thinking about when i have 2-4 t4's or a bunch of a100's attached that are all identical. In the discord however it was suggested that someone wanted to dedicate one gpu to work projects and another to Ollama.
GiteaMirror commented 2026-04-28 03:24:36 -05:00
Author
Owner
Copy Link

@peteygao commented on GitHub (Dec 6, 2023):

I was assuming you were using a heterogenous GPU setup given your last sentence about the "identify the faster GPU".

If your model can fit inside a single GPU, that will yield the maximum performance since there's zero latency penalty from synchronising with other GPUs.

Multi-GPU setup is only useful in two scenarios:

  1. Increasing throughput by having parallel inferences, 1 inference per GPU (assuming the model fits into the VRAM entirely)
  2. Ability to use larger parameter models by splitting the tensors across the GPUs--you'll have less throughput compared to a single "large" GPU, but at least you can run larger models. You lose less throughput if the GPUs are utilising NVLink rather than the PCI-e bus, but I guess you are using a cloud provider and have no way of controlling the physical hardware.
<!-- gh-comment-id:1842421536 --> @peteygao commented on GitHub (Dec 6, 2023): I was assuming you were using a heterogenous GPU setup given your last sentence about the "identify the faster GPU". If your model can fit inside a single GPU, that will yield the maximum performance since there's zero latency penalty from synchronising with other GPUs. Multi-GPU setup is only useful in two scenarios: 1) Increasing throughput by having parallel inferences, 1 inference per GPU (assuming the model fits into the VRAM entirely) 2) Ability to use larger parameter models by splitting the tensors across the GPUs--you'll have less throughput compared to a single "large" GPU, but at least you can run larger models. You lose less throughput if the GPUs are utilising NVLink rather than the PCI-e bus, but I guess you are using a cloud provider and have no way of controlling the physical hardware.
GiteaMirror commented 2026-04-28 03:24:37 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Mar 12, 2024):

Let's track this optimization with #1656

<!-- gh-comment-id:1992072945 --> @dhiltgen commented on GitHub (Mar 12, 2024): Let's track this optimization with #1656
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
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#47175
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?