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 #10092] Using split memory (RAM+VRAM) should never happen #32376

New Issue
Closed
opened 2026-04-22 13:34:56 -05:00 by GiteaMirror · 9 comments
GiteaMirror commented 2026-04-22 13:34:56 -05:00
Owner
Copy Link

Originally created by @Mugane on GitHub (Apr 2, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10092

What is the issue?

Ollama will split models that are too big for the VRAM into VRAM+RAM. This should basically never happen if there is enough system memory since the bottleneck is ALWAYS the transfer between VRAM & RAM. This results in model execution that is an order of magnitude slower than RAM alone. It cannot be interrupted without killing the Ollama server, so it is a really big deal to never do this.

Relevant log output

OS

Linux

GPU

Nvidia

CPU

Intel

Ollama version

0.5.11

Originally created by @Mugane on GitHub (Apr 2, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10092 ### What is the issue? Ollama will split models that are too big for the VRAM into VRAM+RAM. This should basically never happen if there is enough system memory since the bottleneck is ALWAYS the transfer between VRAM & RAM. This results in model execution that is an order of magnitude slower than RAM alone. It cannot be interrupted without killing the Ollama server, so it is a really big deal to never do this. ### Relevant log output ```shell ``` ### OS Linux ### GPU Nvidia ### CPU Intel ### Ollama version 0.5.11
GiteaMirror added the bug label 2026-04-22 13:34:56 -05:00
GiteaMirror commented 2026-04-22 13:35:00 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 2, 2025):

Can you provide more information on the performance issue you are seeing? I did a quick check with qwen2.5:7b and found that model execution was slowest at 0% VRAM usage (all RAM), increased as the RAM+VRAM split favoured VRAM, and was maximum at 100% VRAM usage.

Image

<!-- gh-comment-id:2773395738 --> @rick-github commented on GitHub (Apr 2, 2025): Can you provide more information on the performance issue you are seeing? I did a quick check with qwen2.5:7b and found that model execution was slowest at 0% VRAM usage (all RAM), increased as the RAM+VRAM split favoured VRAM, and was maximum at 100% VRAM usage. ![Image](https://github.com/user-attachments/assets/cfef8a64-db8b-4f2e-957d-4b1364689220)
GiteaMirror commented 2026-04-22 13:35:01 -05:00
Author
Owner
Copy Link

@Mugane commented on GitHub (Apr 2, 2025):

Do it with a model that doesn't fit in VRAM (compare 100% RAM vs split) for example llama3.3:70b-instruct-q8_0). Also it might be affected by your CPU - I ran on an Intel Core i7-10850H (2.7GHz) with 12 virtual processors & 128GB DDR4 SODIMM@2133MT/s. Video card is NVidia Quadro RTX 5000/Max-Q 16GB VRAM

<!-- gh-comment-id:2773407966 --> @Mugane commented on GitHub (Apr 2, 2025): Do it with a model that doesn't fit in VRAM (compare 100% RAM vs split) for example llama3.3:70b-instruct-q8_0). Also it might be affected by your CPU - I ran on an Intel Core i7-10850H (2.7GHz) with 12 virtual processors & 128GB DDR4 SODIMM@2133MT/s. Video card is NVidia Quadro RTX 5000/Max-Q 16GB VRAM
GiteaMirror commented 2026-04-22 13:35:02 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 3, 2025):

llama3.3:70b-instruct-q8_0, i7-13700, 96G DDR5 @ 5200, RTX 4070 12G

Image

<!-- gh-comment-id:2775193822 --> @rick-github commented on GitHub (Apr 3, 2025): llama3.3:70b-instruct-q8_0, i7-13700, 96G DDR5 @ 5200, RTX 4070 12G ![Image](https://github.com/user-attachments/assets/2aba0f55-89bc-456c-9e45-0b220c8ce1c8)
GiteaMirror commented 2026-04-22 13:35:03 -05:00
Author
Owner
Copy Link

@Mugane commented on GitHub (Apr 8, 2025):

Thanks, may I ask what you're using to run the test? I'll see if I can duplicate locally in case there's a real-world vs test setup issue or it's just me or some local issue.

<!-- gh-comment-id:2785686325 --> @Mugane commented on GitHub (Apr 8, 2025): Thanks, may I ask what you're using to run the test? I'll see if I can duplicate locally in case there's a real-world vs test setup issue or it's just me or some local issue.
GiteaMirror commented 2026-04-22 13:35:04 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 8, 2025):

model=llama3.3:70b-instruct-q8_0 ; for i in {0..11} ; do r="{\"layers\":$i}" ; tps=$(for m in default true false ; do mmap="" ; [ "$m" == "true" ] && mmap='"use_mmap":true,' ; [ "$m" == "false" ] && mmap='"use_mmap":false,' ; tps=$(curl -s localhost:11434/api/generate -d '{"model":"'$model'","options":{'$mmap'"num_gpu":'$i',"num_predict":100},"stream":false,"prompt":"why is the sky blue","seed":42}' | jq '.eval_count/(.eval_duration/1000000000)') ; echo -n "+{\"tps_mmap_$m\":$tps}"; done) ; jq -cn "$r$tps+{\"vram\":$(nvidia-smi -x -q | yq -p=xml -o=json  | jq -r '.nvidia_smi_log.gpu.processes.process_info.used_memory // "0"|gsub(" MiB";"")')}" ; done
<!-- gh-comment-id:2785824015 --> @rick-github commented on GitHub (Apr 8, 2025): ``` model=llama3.3:70b-instruct-q8_0 ; for i in {0..11} ; do r="{\"layers\":$i}" ; tps=$(for m in default true false ; do mmap="" ; [ "$m" == "true" ] && mmap='"use_mmap":true,' ; [ "$m" == "false" ] && mmap='"use_mmap":false,' ; tps=$(curl -s localhost:11434/api/generate -d '{"model":"'$model'","options":{'$mmap'"num_gpu":'$i',"num_predict":100},"stream":false,"prompt":"why is the sky blue","seed":42}' | jq '.eval_count/(.eval_duration/1000000000)') ; echo -n "+{\"tps_mmap_$m\":$tps}"; done) ; jq -cn "$r$tps+{\"vram\":$(nvidia-smi -x -q | yq -p=xml -o=json | jq -r '.nvidia_smi_log.gpu.processes.process_info.used_memory // "0"|gsub(" MiB";"")')}" ; done ```
GiteaMirror commented 2026-04-22 13:35:04 -05:00
Author
Owner
Copy Link

@Mugane commented on GitHub (Apr 13, 2025):

@rick-github Did you actually run that?

<!-- gh-comment-id:2799998004 --> @Mugane commented on GitHub (Apr 13, 2025): @rick-github Did you actually run that?
GiteaMirror commented 2026-04-22 13:35:05 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 13, 2025):

Yes, that's how I got the numbers to create the graph in https://github.com/ollama/ollama/issues/10092#issuecomment-2775193822

<!-- gh-comment-id:2800004712 --> @rick-github commented on GitHub (Apr 13, 2025): Yes, that's how I got the numbers to create the graph in https://github.com/ollama/ollama/issues/10092#issuecomment-2775193822
GiteaMirror commented 2026-04-22 13:35:06 -05:00
Author
Owner
Copy Link

@Mugane commented on GitHub (Apr 30, 2025):

This is not completed. I have not had time to debug your script yet. The issue persists.

<!-- gh-comment-id:2842983271 --> @Mugane commented on GitHub (Apr 30, 2025): This is not completed. I have not had time to debug your script yet. The issue persists.
GiteaMirror commented 2026-04-22 13:35:06 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 30, 2025):

What problem are you having with the script?

<!-- gh-comment-id:2843013721 --> @rick-github commented on GitHub (Apr 30, 2025): What problem are you having with the script?
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#32376
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?