github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-15 12:48:05 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #6769] OLLAMA_FLASH_ATTENTION regression on 0.3.10? #4265

New Issue
Closed
opened 2026-04-12 15:11:58 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-12 15:11:58 -05:00
Owner
Copy Link

Originally created by @coodoo on GitHub (Sep 12, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/6769

Originally assigned to: @jmorganca on GitHub.

What is the issue?

After upgrading to the latest version 0.3.10, with OLLAMA_FLASH_ATTENTION=1 set in env, seemed the tokens per second were halved, in my experiment, same code used to have tps around 23 and now it's only 11.

Wondering is there any known regression with regard to FLASH_ATTENTION?

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.3.10

Originally created by @coodoo on GitHub (Sep 12, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/6769 Originally assigned to: @jmorganca on GitHub. ### What is the issue? After upgrading to the latest version `0.3.10`, with `OLLAMA_FLASH_ATTENTION=1` set in env, seemed the tokens per second were halved, in my experiment, same code used to have tps around 23 and now it's only 11. Wondering is there any known regression with regard to FLASH_ATTENTION? ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.3.10
GiteaMirror added the bug label 2026-04-12 15:11:58 -05:00
GiteaMirror commented 2026-04-12 15:11:59 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Sep 12, 2024):

@coodoo is this on Apple Silicon? I'm still seeing a ~10% speedup between OLLAMA_FLASH_ATTENTION=1 and OLLAMA_FLASH_ATTENTION=0.

Also, which model are you using?

Sorry you hit this, and thanks for creating an issue!

<!-- gh-comment-id:2345316285 --> @jmorganca commented on GitHub (Sep 12, 2024): @coodoo is this on Apple Silicon? I'm still seeing a ~10% speedup between `OLLAMA_FLASH_ATTENTION=1` and `OLLAMA_FLASH_ATTENTION=0`. Also, which model are you using? Sorry you hit this, and thanks for creating an issue!
GiteaMirror commented 2026-04-12 15:11:59 -05:00
Author
Owner
Copy Link

@coodoo commented on GitHub (Sep 12, 2024):

Yup same M1 mac book pro, same model (gemma2-9B), same code, same test data.

I'm currently running more benchmarks to identify the root cause, hopefully something will come out! 🤞

<!-- gh-comment-id:2345321953 --> @coodoo commented on GitHub (Sep 12, 2024): Yup same M1 mac book pro, same model (`gemma2-9B`), same code, same test data. I'm currently running more benchmarks to identify the root cause, hopefully something will come out! 🤞
GiteaMirror commented 2026-04-12 15:11:59 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Jul 23, 2025):

This should be improved but let me know if it's not

<!-- gh-comment-id:3109679707 --> @jmorganca commented on GitHub (Jul 23, 2025): This should be improved but let me know if it's not
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#4265
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?