github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 17:49:24 -05:00
Code Issues 14.6k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #1906] Ollama not respecting num_gpu to load entire model into VRAM for a model that I know should fit into 24GB. #47608

New Issue
Closed
opened 2026-04-28 04:30:02 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-28 04:30:02 -05:00
Owner
Copy Link

Originally created by @madsamjp on GitHub (Jan 10, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/1906

Originally assigned to: @jmorganca on GitHub.

Somewhat related to this issue: https://github.com/jmorganca/ollama/issues/1374

I have a model that I have configured to fit almost exactly into my 4090's VRAM. Prior to v 0.1.13, this model ran fine, and I could fit all layers into VRAM and fill the context. It would utilize a total of 23697MiB (which gave about 900MiB headroom). After 0.1.13, this model would cause me an OOM. I found a temporary solution to this by building ollama from source with the -DLLAMA_CUDA_FORCE_MMQ=on flag.

Now after the recent update to 0.1.19, there was a PR that alleged to fix this issue https://github.com/jmorganca/ollama/pull/1850.

I can now run that model without the OOM, however, Ollama never offloads more than 49 of the 63 layers to GPU. Even if I set the num_gpu parameter in interactive mode to 63 or higher, it still loads only 49 of the layers and only utilizes 21027MiB of a total of 24564MiB (only 86% of VRAM).

Here is the modelfile:

FROM deepseek-coder:33b-instruct-q5_K_S

PARAMETER num_gpu 63
PARAMETER num_ctx 2048

Is it possible to force Ollama to load the entire model into VRAM as it was before v 0.1.11? Do models now take up MORE VRAM than before?

VRAM is expensive and scarce at this time. I feel that giving users the flexibility to finely control models to maximize VRAM usage is imperative.

Originally created by @madsamjp on GitHub (Jan 10, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/1906 Originally assigned to: @jmorganca on GitHub. Somewhat related to this issue: https://github.com/jmorganca/ollama/issues/1374 I have a model that I have configured to fit almost exactly into my 4090's VRAM. Prior to v 0.1.13, this model ran fine, and I could fit all layers into VRAM and fill the context. It would utilize a total of 23697MiB (which gave about 900MiB headroom). After 0.1.13, this model would cause me an OOM. I found a temporary solution to this by building ollama from source with the `-DLLAMA_CUDA_FORCE_MMQ=on` flag. Now after the recent update to 0.1.19, there was a PR that alleged to fix this issue https://github.com/jmorganca/ollama/pull/1850. I can now run that model without the OOM, however, Ollama never offloads more than 49 of the 63 layers to GPU. Even if I set the num_gpu parameter in interactive mode to 63 or higher, it still loads only 49 of the layers and only utilizes 21027MiB of a total of 24564MiB (only 86% of VRAM). Here is the modelfile: ``` FROM deepseek-coder:33b-instruct-q5_K_S PARAMETER num_gpu 63 PARAMETER num_ctx 2048 ``` Is it possible to force Ollama to load the entire model into VRAM as it was before v 0.1.11? Do models now take up MORE VRAM than before? VRAM is expensive and scarce at this time. I feel that giving users the flexibility to finely control models to maximize VRAM usage is imperative.
GiteaMirror added the bug label 2026-04-28 04:30:02 -05:00
GiteaMirror commented 2026-04-28 04:30:04 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (May 10, 2024):

Hi there, this should be improved now as memory utilization estimates are more accurate – let me know if you're still seeing this issue and I can re-open.

<!-- gh-comment-id:2103634207 --> @jmorganca commented on GitHub (May 10, 2024): Hi there, this should be improved now as memory utilization estimates are more accurate – let me know if you're still seeing this issue and I can re-open.
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#47608
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?