github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 08:02:14 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #10539] Allow "use_mmap" to be set at a global level using enviroment variables. #6935

New Issue
Closed
opened 2026-04-12 18:49:40 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-12 18:49:40 -05:00
Owner
Copy Link

Originally created by @Slymi on GitHub (May 2, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10539

Context:

I am running Ollama V0.6.7 using Docker on a system running TRUENAS Scale with 128 GB of RAM and 02x 16 GB RTX 4060 TI. The container running off a NVME drive.

Current Issue

After loading a model that does not use the new Ollama engine with "use_mmap" = true, the allocated RAM is not fully released when the model stopped. Not only that, loading times are worse with "use_mmap" = true.

Example with "use_mmap" = true (Qwen3:32b Q4_K_M @ 16384 Context Size):

Before loading (System RAM Usage on the right):
Image

Loaded:
Image

Stopped (450~ MB released?):
Image

Results:
Image

Logs:
Qwen3_MMAP_Logs.txt

Example with "use_mmap" = false (Qwen3:32b Q4_K_M @ 16384 Context Size):

Loaded:
Image

Stopped:
Image

Results:
Image

Logs:
Qwen3_NoMMAP_Logs.txt

Notes:

  • Restarting the container does not free up the RAM either, requiring a full system reboot to resolve.
  • In both examples, subsequent model loading times are reduced to 4-5 seconds as the model has been picked up by ZFS ARC (RAM Cache).
  • If the RAM limit of the container is reduced to below the model file size, "use_mmap" = true would significantly lengthen the time it takes to load the model. Example as shown below.

Examples of 8 GB container RAM limit and loading from ZFS ARC.

"use_mmap" = true:
Image

"use_mmap" = false:
Image

12x improvement with "use_mmap" = false as the model is already stored in the ZFS Cache.

Conclusion:

I hope the developers of Ollama will look into this and allow "use_mmap" to be set globally in the enviroment variables in order to resolve these kind of issues. As of now, I'm using OpenWebUI to set "use_mmap" = false for all models. However, some other frontends lack the option to do so. Thus, restricting the frontends that can be used that allow for a satisfactory experience.

References that I think is related:
#6854
#10076

Originally created by @Slymi on GitHub (May 2, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10539 ### Context: I am running Ollama V0.6.7 using Docker on a system running TRUENAS Scale with 128 GB of RAM and 02x 16 GB RTX 4060 TI. The container running off a NVME drive. ### Current Issue After loading a model that does not use the new Ollama engine with "use_mmap" = true, the allocated RAM is not fully released when the model stopped. Not only that, loading times are worse with "use_mmap" = true. ### Example with "use_mmap" = true (Qwen3:32b Q4_K_M @ 16384 Context Size): **Before loading (System RAM Usage on the right):** ![Image](https://github.com/user-attachments/assets/31b480e3-b069-44f6-9206-57bd456ed3f4) **Loaded:** ![Image](https://github.com/user-attachments/assets/a613c938-347d-4604-91ae-5d572601b274) **Stopped (450~ MB released?):** ![Image](https://github.com/user-attachments/assets/c01413aa-d322-45d0-99a9-96aeb76df50f) **Results:** ![Image](https://github.com/user-attachments/assets/8743c584-1227-4819-aaaf-b96dad595396) **Logs:** [Qwen3_MMAP_Logs.txt](https://github.com/user-attachments/files/20015338/Qwen3_MMAP_Logs.txt) ### Example with "use_mmap" = false (Qwen3:32b Q4_K_M @ 16384 Context Size): **Loaded:** ![Image](https://github.com/user-attachments/assets/8ff9f7da-27b9-4f83-b976-cb7e70068208) **Stopped:** ![Image](https://github.com/user-attachments/assets/ef076a2d-a809-4488-b1bb-86a1a7d92955) **Results:** ![Image](https://github.com/user-attachments/assets/bdd6e4de-b51c-4ed9-ac8d-e7ea963b6500) **Logs:** [Qwen3_NoMMAP_Logs.txt](https://github.com/user-attachments/files/20015446/Qwen3_NoMMAP_Logs.txt) **Notes:** - Restarting the container does not free up the RAM either, requiring a full system reboot to resolve. - In both examples, subsequent model loading times are reduced to 4-5 seconds as the model has been picked up by ZFS ARC (RAM Cache). - If the RAM limit of the container is reduced to below the model file size, "use_mmap" = true would significantly lengthen the time it takes to load the model. Example as shown below. ### Examples of 8 GB container RAM limit and loading from ZFS ARC. **"use_mmap" = true:** ![Image](https://github.com/user-attachments/assets/b1a2f18f-f111-4b0e-bc53-80a42e72e8f3) **"use_mmap" = false:** ![Image](https://github.com/user-attachments/assets/b0cdcc1e-0bb3-4242-b0ee-260b34034385) 12x improvement with "use_mmap" = false as the model is already stored in the ZFS Cache. ### Conclusion: I hope the developers of Ollama will look into this and allow "use_mmap" to be set globally in the enviroment variables in order to resolve these kind of issues. As of now, I'm using OpenWebUI to set "use_mmap" = false for all models. However, some other frontends lack the option to do so. Thus, restricting the frontends that can be used that allow for a satisfactory experience. **References that I think is related:** #6854 #10076
GiteaMirror added the feature request label 2026-04-12 18:49:40 -05:00
GiteaMirror commented 2026-04-12 18:49:42 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jun 6, 2025):

The new engine doesn't use mmap so use_mmap is ignored. For the old engine, #8895.

<!-- gh-comment-id:2948914926 --> @rick-github commented on GitHub (Jun 6, 2025): The new engine doesn't use mmap so `use_mmap` is ignored. For the old engine, #8895.
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 feature request
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#6935
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?