github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-16 13:19:40 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #10960] Ollama appears to corrupt deepseek-r1:7b model #53734

New Issue
Closed
opened 2026-04-29 04:36:57 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-04-29 04:36:57 -05:00
Owner
Copy Link

Originally created by @RetroDev256 on GitHub (Jun 3, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10960

What is the issue?

Certain trivial inputs cause deepseek-r1:7b to enter an infinite thinking loop, and the AI output appears mangled after several minutes. Attached is a script recording. Here is an asciinema link: https://asciinema.org/a/721755

Relevant log output

(not embedded because it's too large)
deepseek_trash.txt

OS

Linux (Gentoo)

GPU

NVIDIA GeForce RTX 4060 Laptop GPU

CPU

Intel(R) Core(TM) i9-14900HX

Ollama version

ollama version is 0.9.0

Originally created by @RetroDev256 on GitHub (Jun 3, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10960 ### What is the issue? Certain trivial inputs cause deepseek-r1:7b to enter an infinite thinking loop, and the AI output appears mangled after several minutes. Attached is a `script` recording. Here is an asciinema link: https://asciinema.org/a/721755 ### Relevant log output (not embedded because it's too large) [deepseek_trash.txt](https://github.com/user-attachments/files/20579938/deepseek_trash.txt) ### OS Linux (Gentoo) ### GPU NVIDIA GeForce RTX 4060 Laptop GPU ### CPU Intel(R) Core(TM) i9-14900HX ### Ollama version ollama version is 0.9.0
GiteaMirror added the bug label 2026-04-29 04:36:57 -05:00
GiteaMirror commented 2026-04-29 04:36:59 -05:00
Author
Owner
Copy Link

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

The model might be spending so many tokens thinking that it runs out of space, causing it to lose cohesion. Increase the context length and see if the output improves. Server logs may aid in debugging.

<!-- gh-comment-id:2937057705 --> @rick-github commented on GitHub (Jun 3, 2025): The model might be spending so many tokens thinking that it runs out of space, causing it to lose cohesion. Increase the [context length](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-specify-the-context-window-size) and see if the output improves. [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) may aid in debugging.
GiteaMirror commented 2026-04-29 04:37:00 -05:00
Author
Owner
Copy Link

@RetroDev256 commented on GitHub (Jun 3, 2025):

Thanks! That makes sense. After boosting the context length, it returns an answer.
Image
EDIT: server logs don't show anything unique for shorter context length.

<!-- gh-comment-id:2937317190 --> @RetroDev256 commented on GitHub (Jun 3, 2025): Thanks! That makes sense. After boosting the context length, it returns an answer. ![Image](https://github.com/user-attachments/assets/230c2745-c13e-4461-a7d9-f3a87fbc6872) EDIT: server logs don't show anything unique for shorter context length.
GiteaMirror commented 2026-04-29 04:37:00 -05:00
Author
Owner
Copy Link

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

I think if you enabled debugging (OLLAMA_DEBUG=1 in the server environment) you will see messages about shifting, which indicates a context buffer overflow.

<!-- gh-comment-id:2937343442 --> @rick-github commented on GitHub (Jun 3, 2025): I think if you enabled debugging (`OLLAMA_DEBUG=1` in the server environment) you will see messages about shifting, which indicates a context buffer overflow.
GiteaMirror commented 2026-04-29 04:37:01 -05:00
Author
Owner
Copy Link

@RetroDev256 commented on GitHub (Jun 3, 2025):

Yep, it's there alright.
Image
Is this an issue that ollama would be expected to fix? Or is the behavior expected? Wondering if the issue should be closed.

<!-- gh-comment-id:2937425181 --> @RetroDev256 commented on GitHub (Jun 3, 2025): Yep, it's there alright. ![Image](https://github.com/user-attachments/assets/366ab962-4fe1-4b9b-8a67-073daa879929) Is this an issue that ollama would be expected to fix? Or is the behavior expected? Wondering if the issue should be closed.
GiteaMirror commented 2026-04-29 04:37:02 -05:00
Author
Owner
Copy Link

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

It comes down to a configuration issue. Thinking models are token hungry, and ollama recently increased the default length partly in response, but some models are just thinkier than others. There are downsides to having ollama guess at the amount required, so it's up to the user to set limits. Hand-in-hand with num_ctx is num_predict, which limits the number of tokens that can be generated.

<!-- gh-comment-id:2937451820 --> @rick-github commented on GitHub (Jun 3, 2025): It comes down to a configuration issue. Thinking models are token hungry, and ollama recently increased the default length partly in response, but some models are just thinkier than others. There are downsides to having ollama guess at the amount required, so it's up to the user to set limits. Hand-in-hand with `num_ctx` is `num_predict`, which limits the number of tokens that can be generated.
GiteaMirror commented 2026-04-29 04:37:05 -05:00
Author
Owner
Copy Link

@RetroDev256 commented on GitHub (Jun 3, 2025):

Thanks for the info! I solved my issue by creating a modelfile for deepseek-r1:7b which increased the number of tokens for the context, and I'll be sure to also set num_predict.

<!-- gh-comment-id:2937455752 --> @RetroDev256 commented on GitHub (Jun 3, 2025): Thanks for the info! I solved my issue by creating a modelfile for `deepseek-r1:7b` which increased the number of tokens for the context, and I'll be sure to also set `num_predict`.
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#53734
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?