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 #7531] Poor acceleration choices with mixed GPUs #4790

New Issue
Closed
opened 2026-04-12 15:45:30 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-12 15:45:30 -05:00
Owner
Copy Link

Originally created by @cobrafast on GitHub (Nov 6, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/7531

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

I've noticed is that Ollama makes poor decisions about acceleration in setups with heterogenous GPUs. I for example have a 16GB VRAM and a 3GB VRAM dGPU in my desktop PC and Ollama seems to only consider the smaller VRAM GPU, even if I set up CUDA_VISIBLE_DEVICES=0 to only let it compute on the bigger one.

time=2024-11-03T22:28:58.806+01:00 level=INFO source=gpu.go:221 msg="looking for compatible GPUs"
time=2024-11-03T22:28:58.806+01:00 level=INFO source=gpu_windows.go:167 msg=packages count=1
time=2024-11-03T22:28:58.806+01:00 level=INFO source=gpu_windows.go:214 msg="" package=0 cores=8 efficiency=0 threads=16
time=2024-11-03T22:28:59.090+01:00 level=INFO source=types.go:123 msg="inference compute" id=GPU-3392b891-9899-c4e1-5fff-f56fe0c463c5 library=cuda variant=v12 compute=8.9 driver=12.4 name="NVIDIA GeForce RTX 4080" total="16.0 GiB" available="14.7 GiB"
...
time=2024-11-06T21:16:40.820+01:00 level=INFO source=server.go:105 msg="system memory" total="127.7 GiB" free="85.7 GiB" free_swap="96.8 GiB"
time=2024-11-06T21:16:40.821+01:00 level=INFO source=memory.go:326 msg="offload to cuda" layers.requested=-1 layers.model=33 layers.offload=2 layers.split="" memory.available="[914.5 MiB]" memory.gpu_overhead="0 B" memory.required.full="3.4 GiB" memory.required.partial="839.3 MiB" memory.required.kv="768.0 MiB" memory.required.allocations="[839.3 MiB]" memory.weights.total="2.6 GiB" memory.weights.repeating="2.6 GiB" memory.weights.nonrepeating="77.1 MiB" memory.graph.full="128.0 MiB" memory.graph.partial="128.0 MiB"
...
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4080, compute capability 8.9, VMM: yes

If I'm reading this right, then Ollama thinks theres 839 MiB of available VRAM, which seems to be correct for the smaller GPU, but the bigger one should have some ~15 GiB available that don't seem to get considered at all.
This seems to make Ollama split the model between CPU and GPU, or run on CPU entirely.

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.3.14

Originally created by @cobrafast on GitHub (Nov 6, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/7531 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? I've noticed is that Ollama makes poor decisions about acceleration in setups with heterogenous GPUs. I for example have a 16GB VRAM and a 3GB VRAM dGPU in my desktop PC and Ollama seems to only consider the smaller VRAM GPU, even if I set up `CUDA_VISIBLE_DEVICES=0` to only let it compute on the bigger one. ``` time=2024-11-03T22:28:58.806+01:00 level=INFO source=gpu.go:221 msg="looking for compatible GPUs" time=2024-11-03T22:28:58.806+01:00 level=INFO source=gpu_windows.go:167 msg=packages count=1 time=2024-11-03T22:28:58.806+01:00 level=INFO source=gpu_windows.go:214 msg="" package=0 cores=8 efficiency=0 threads=16 time=2024-11-03T22:28:59.090+01:00 level=INFO source=types.go:123 msg="inference compute" id=GPU-3392b891-9899-c4e1-5fff-f56fe0c463c5 library=cuda variant=v12 compute=8.9 driver=12.4 name="NVIDIA GeForce RTX 4080" total="16.0 GiB" available="14.7 GiB" ... time=2024-11-06T21:16:40.820+01:00 level=INFO source=server.go:105 msg="system memory" total="127.7 GiB" free="85.7 GiB" free_swap="96.8 GiB" time=2024-11-06T21:16:40.821+01:00 level=INFO source=memory.go:326 msg="offload to cuda" layers.requested=-1 layers.model=33 layers.offload=2 layers.split="" memory.available="[914.5 MiB]" memory.gpu_overhead="0 B" memory.required.full="3.4 GiB" memory.required.partial="839.3 MiB" memory.required.kv="768.0 MiB" memory.required.allocations="[839.3 MiB]" memory.weights.total="2.6 GiB" memory.weights.repeating="2.6 GiB" memory.weights.nonrepeating="77.1 MiB" memory.graph.full="128.0 MiB" memory.graph.partial="128.0 MiB" ... ggml_cuda_init: found 1 CUDA devices: Device 0: NVIDIA GeForce RTX 4080, compute capability 8.9, VMM: yes ``` If I'm reading this right, then Ollama thinks theres 839 MiB of available VRAM, which seems to be correct for the smaller GPU, but the bigger one should have some ~15 GiB available that don't seem to get considered at all. This seems to make Ollama split the model between CPU and GPU, or run on CPU entirely. ### OS Windows ### GPU Nvidia ### CPU Intel ### Ollama version 0.3.14
GiteaMirror added the nvidiabug labels 2026-04-12 15:45:30 -05:00
GiteaMirror commented 2026-04-12 15:45:33 -05:00
Author
Owner
Copy Link

@cobrafast commented on GitHub (Nov 6, 2024):

Just updated to 0.4.0 and this seems fixed already. I'll keep an eye on it and report back or close accordingly.

<!-- gh-comment-id:2460898109 --> @cobrafast commented on GitHub (Nov 6, 2024): Just updated to `0.4.0` and this seems fixed already. I'll keep an eye on it and report back or close accordingly.
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 nvidia
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#4790
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?