github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-15 20:54:54 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #12265] ollama v0.11.10 vs v0.11.11-rc0 - regression #54667

New Issue
Closed
opened 2026-04-29 06:47:57 -05:00 by GiteaMirror · 9 comments
GiteaMirror commented 2026-04-29 06:47:57 -05:00
Owner
Copy Link

Originally created by @eXt73 on GitHub (Sep 12, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12265

What is the issue?

Under settings:
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="OLLAMA_KV_CACHE_TYPE=q4_0"
Environment="OLLAMA_NEW_ESTIMATES=1"
Environment="OLLAMA_NEW_ENGINE=1"

Under Ollama v0.11.10, models take up at least 30% less VRAM than under v0.11.11-rc0. After activating the new Ollama engine and new memory estimates, Ollama v0.11.10 dramatically reduced VRAM requirements – by about 30% on the RTX 5090 and even several times on the GH200 – when run with the Environment="OLLAMA_NUM_PARALLEL=30" variable. The test version 11-rc0 again results in a significant increase in VRAM usage - despite the use of variables:

Environment="OLLAMA_NEW_ESTIMATES=1"
Environment="OLLAMA_NEW_ENGINE=1"

Tested on models:

Mistral-Small-3.2-24B
Qwen3-30B-A3B-Instruct-2507
embeddinggemma-300m
etc.

Moreover, under v0.11.10 models start even several times faster [with new engine & new estimates].

Image

Relevant log output

OS

Linux

GPU

Nvidia

CPU

AMD

Ollama version

v0.11.11-rc0

Originally created by @eXt73 on GitHub (Sep 12, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12265 ### What is the issue? Under settings: Environment="OLLAMA_FLASH_ATTENTION=1" Environment="OLLAMA_KV_CACHE_TYPE=q4_0" Environment="OLLAMA_NEW_ESTIMATES=1" Environment="OLLAMA_NEW_ENGINE=1" Under Ollama v0.11.10, models take up at least 30% less VRAM than under v0.11.11-rc0. After activating the new Ollama engine and new memory estimates, Ollama v0.11.10 dramatically reduced VRAM requirements – by about 30% on the RTX 5090 and even several times on the GH200 – when run with the Environment="OLLAMA_NUM_PARALLEL=30" variable. The test version 11-rc0 again results in a significant increase in VRAM usage - despite the use of variables: Environment="OLLAMA_NEW_ESTIMATES=1" Environment="OLLAMA_NEW_ENGINE=1" Tested on models: Mistral-Small-3.2-24B Qwen3-30B-A3B-Instruct-2507 embeddinggemma-300m etc. Moreover, under v0.11.10 models start even several times faster [with new engine & new estimates]. <img width="3840" height="2160" alt="Image" src="https://github.com/user-attachments/assets/422feea1-bbe7-4aef-9beb-4a9ee345fa3e" /> ### Relevant log output ```shell ``` ### OS Linux ### GPU Nvidia ### CPU AMD ### Ollama version v0.11.11-rc0
GiteaMirror added the bug label 2026-04-29 06:47:57 -05:00
GiteaMirror commented 2026-04-29 06:48:00 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 12, 2025):

Please post text, not screenshots.

<!-- gh-comment-id:3285077045 --> @rick-github commented on GitHub (Sep 12, 2025): Please post text, not screenshots.
GiteaMirror commented 2026-04-29 06:48:01 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 12, 2025):

Mistral-Small-3.2-24B-Instruct-2506-GGUF:latest is split GGUF model with a vision component, even though it doesn't work. Previously the model was loaded with the new ollama engine because of OLLAMA_NEW_ENGINE=1. However, with this change ollama no longer loads split models into the new engine (because the vision component doesn't work). So in 0.11.11-rc0, the model is being loaded into the old engine, which has less accurate memory estimation.

You can get back to the previous behaviour by creating a model that doesn't have the vision GGUF in it:

ollama show --modelfile hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:latest | grep -v d6af684a > Modelfile
ollama create hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-novision-2506-GGUF:latest

$ ollama -v
ollama version is 0.11.11-rc0
$ ollama ps
NAME                                                                      ID              SIZE     PROCESSOR    CONTEXT    UNTIL   
hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-novision-2506-GGUF:latest    b396eb94fd04    17 GB    100% GPU     65536      Forever    
hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:latest             39224c6559c9    22 GB    100% GPU     65536      Forever
<!-- gh-comment-id:3285833827 --> @rick-github commented on GitHub (Sep 12, 2025): `Mistral-Small-3.2-24B-Instruct-2506-GGUF:latest` is split GGUF model with a vision component, even though it doesn't work. Previously the model was loaded with the new ollama engine because of `OLLAMA_NEW_ENGINE=1`. However, with [this change](https://github.com/ollama/ollama/pull/12241) ollama no longer loads split models into the new engine (because the vision component doesn't work). So in 0.11.11-rc0, the model is being loaded into the old engine, which has less accurate memory estimation. You can get back to the previous behaviour by creating a model that doesn't have the vision GGUF in it: ``` ollama show --modelfile hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:latest | grep -v d6af684a > Modelfile ollama create hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-novision-2506-GGUF:latest ``` ```console $ ollama -v ollama version is 0.11.11-rc0 $ ollama ps NAME ID SIZE PROCESSOR CONTEXT UNTIL hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-novision-2506-GGUF:latest b396eb94fd04 17 GB 100% GPU 65536 Forever hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:latest 39224c6559c9 22 GB 100% GPU 65536 Forever ```
GiteaMirror commented 2026-04-29 06:48:02 -05:00
Author
Owner
Copy Link

@eXt73 commented on GitHub (Sep 12, 2025):

Thanks a lot for the explanation - now it makes sense :) btw. as for the image - it was only for 'illustrative purposes' - I included the key information in the text ;)

<!-- gh-comment-id:3285898787 --> @eXt73 commented on GitHub (Sep 12, 2025): Thanks a lot for the explanation - now it makes sense :) btw. as for the image - it was only for 'illustrative purposes' - I included the key information in the text ;)
GiteaMirror commented 2026-04-29 06:48:02 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 12, 2025):

Except for the full model name.

<!-- gh-comment-id:3285901931 --> @rick-github commented on GitHub (Sep 12, 2025): Except for the full model name.
GiteaMirror commented 2026-04-29 06:48:02 -05:00
Author
Owner
Copy Link

@eXt73 commented on GitHub (Sep 12, 2025):

ok You're right :)

<!-- gh-comment-id:3285943092 --> @eXt73 commented on GitHub (Sep 12, 2025): ok You're right :)
GiteaMirror commented 2026-04-29 06:48:03 -05:00
Author
Owner
Copy Link

@eXt73 commented on GitHub (Sep 12, 2025):

Will it be the same in Devstral ?

<!-- gh-comment-id:3285960532 --> @eXt73 commented on GitHub (Sep 12, 2025): Will it be the same in Devstral ?
GiteaMirror commented 2026-04-29 06:48:04 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 12, 2025):

Which devstral?

<!-- gh-comment-id:3285979193 --> @rick-github commented on GitHub (Sep 12, 2025): Which devstral?
GiteaMirror commented 2026-04-29 06:48:04 -05:00
Author
Owner
Copy Link

@eXt73 commented on GitHub (Sep 12, 2025):

hf.co/mistralai/Devstral-Small-2507_gguf:Q8_0

<!-- gh-comment-id:3285982344 --> @eXt73 commented on GitHub (Sep 12, 2025): hf.co/mistralai/Devstral-Small-2507_gguf:Q8_0
GiteaMirror commented 2026-04-29 06:48:05 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 12, 2025):

devstral is not a split GGUF model so should be unaffected.

<!-- gh-comment-id:3286286909 --> @rick-github commented on GitHub (Sep 12, 2025): devstral is not a split GGUF model so should be unaffected.
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#54667
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?