github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 08:59:13 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #5390] deepseek-coder-v2-lite flash attention not enabled #3371

New Issue
Closed
opened 2026-04-12 14:00:02 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-12 14:00:02 -05:00
Owner
Copy Link

Originally created by @reddev-aroy on GitHub (Jun 30, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/5390

What is the issue?

As soon as the context length limit is reached for deepseek-coder-v2-lite the models are just repeating previous answers and keeps looping itself even after asking for something else in its response. This issue is resolved in LM Studio as after enabling flash attention this seems to be resolved, however this issue still exists in the latest ollama 0.1.48
I am suspecting it to be issue with the model itself but flash attention seems to resolve it in LM Studio. Need help resolving this issue in Ollama as running this model in cpu only seems to be faster for me attest in ollama than LM Studio with num_gpu 0

Ollama version - 0.1.48
Model used - deepseek-coder-v2-lite-instruct-Q5_K_M

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.1.48

Originally created by @reddev-aroy on GitHub (Jun 30, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/5390 ### What is the issue? As soon as the context length limit is reached for deepseek-coder-v2-lite the models are just repeating previous answers and keeps looping itself even after asking for something else in its response. This issue is resolved in LM Studio as after enabling flash attention this seems to be resolved, however this issue still exists in the latest ollama 0.1.48 I am suspecting it to be issue with the model itself but flash attention seems to resolve it in LM Studio. Need help resolving this issue in Ollama as running this model in cpu only seems to be faster for me attest in ollama than LM Studio with num_gpu 0 Ollama version - 0.1.48 Model used - deepseek-coder-v2-lite-instruct-Q5_K_M ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.1.48
GiteaMirror added the bug label 2026-04-12 14:00:02 -05:00
GiteaMirror commented 2026-04-12 14:00:05 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jun 30, 2024):

You can enable flash attention in ollama by setting OLLAMA_FLASH_ATTENTION=1 in the environment.

<!-- gh-comment-id:2198505803 --> @rick-github commented on GitHub (Jun 30, 2024): You can enable flash attention in ollama by setting OLLAMA_FLASH_ATTENTION=1 in the environment.
GiteaMirror commented 2026-04-12 14:00:05 -05:00
Author
Owner
Copy Link

@reddev-aroy commented on GitHub (Jun 30, 2024):

You can enable flash attention in ollama by setting OLLAMA_FLASH_ATTENTION=1 in the environment.

Tried below and restarting ollama, doesn't seem to work. Maybe deepseek-coder-v2 in ollama is turning off flash_attn automatically for this model architecture? Not sure.

launchctl setenv OLLAMA_FLASH_ATTENTION 1

<!-- gh-comment-id:2198511152 --> @reddev-aroy commented on GitHub (Jun 30, 2024): > You can enable flash attention in ollama by setting OLLAMA_FLASH_ATTENTION=1 in the environment. Tried below and restarting ollama, doesn't seem to work. Maybe deepseek-coder-v2 in ollama is turning off flash_attn automatically for this model architecture? Not sure. launchctl setenv OLLAMA_FLASH_ATTENTION 1
GiteaMirror commented 2026-04-12 14:00:06 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jun 30, 2024):

It's an issue with llama.cpp, it turns off flash attention when the K and V heads are different.

<!-- gh-comment-id:2198528627 --> @rick-github commented on GitHub (Jun 30, 2024): It's an issue with llama.cpp, it [turns off flash attention](https://github.com/ggerganov/llama.cpp/blob/1c5eba6f8e628fb0a98afb27d8aaeb3b0e136451/src/llama.cpp#L17412) when the K and V heads are different.
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#3371
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?