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 #9575] Endless generations in QwQ quants #6245

New Issue
Open
opened 2026-04-12 17:40:43 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-04-12 17:40:43 -05:00
Owner
Copy Link

Originally created by @wsbagnsv1 on GitHub (Mar 7, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9575

What is the issue?

Unsloth discovered that the ordering of the samplers somehow breaks the model, you can read up on it here:
https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-without-bugs

This causes endless generations and since you cant change the order of the samplers, to my knowledge at least, QwQ is broken. I would think that a rather simple line in the Modelfile that enables to pass "--samplers "top_k;top_p;min_p;temperature;dry;typ_p;xtc" " to llama.cpp should fix the issue.

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @wsbagnsv1 on GitHub (Mar 7, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9575 ### What is the issue? Unsloth discovered that the ordering of the samplers somehow breaks the model, you can read up on it here: https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-without-bugs This causes endless generations and since you cant change the order of the samplers, to my knowledge at least, QwQ is broken. I would think that a rather simple line in the Modelfile that enables to pass "--samplers "top_k;top_p;min_p;temperature;dry;typ_p;xtc" " to llama.cpp should fix the issue. ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-12 17:40:43 -05:00
GiteaMirror commented 2026-04-12 17:40:45 -05:00
Author
Owner
Copy Link

@crazyi commented on GitHub (Mar 7, 2025):

I find the same issue, how can I fix it in Ollama?

<!-- gh-comment-id:2706492873 --> @crazyi commented on GitHub (Mar 7, 2025): I find the same issue, how can I fix it in Ollama?
GiteaMirror commented 2026-04-12 17:40:46 -05:00
Author
Owner
Copy Link

@sultanqasim commented on GitHub (Mar 7, 2025):

You can run it as suggested by Qwen developers with a repetition penalty of 1.0 which avoids the sampler ordering issue. To use a higher repetition penalty, you will need the ordering modification they described.

<!-- gh-comment-id:2706572324 --> @sultanqasim commented on GitHub (Mar 7, 2025): You can run it as suggested by Qwen developers with a repetition penalty of 1.0 which avoids the sampler ordering issue. To use a higher repetition penalty, you will need the ordering modification they described.
GiteaMirror commented 2026-04-12 17:40:46 -05:00
Author
Owner
Copy Link

@yorktownting commented on GitHub (Mar 8, 2025):

@crazyi @wsbagnsv1

Change the Modelfile can solve this problem. This is my version, it works well:

FROM ./QwQ-32B-Q5_K_M.gguf
PARAMETER top_k 40
PARAMETER top_p 0.95
PARAMETER min_p 0.1
PARAMETER temperature 0.6
PARAMETER num_predict 32768
PARAMETER repeat_penalty 1.1
TEMPLATE """<|im_start|>user\n{{ .Prompt }}<|im_end|>\n<|im_start|>assistant\n<think>\n"""
<!-- gh-comment-id:2707875151 --> @yorktownting commented on GitHub (Mar 8, 2025): @crazyi @wsbagnsv1 Change the Modelfile can solve this problem. This is my version, it works well: ``` FROM ./QwQ-32B-Q5_K_M.gguf PARAMETER top_k 40 PARAMETER top_p 0.95 PARAMETER min_p 0.1 PARAMETER temperature 0.6 PARAMETER num_predict 32768 PARAMETER repeat_penalty 1.1 TEMPLATE """<|im_start|>user\n{{ .Prompt }}<|im_end|>\n<|im_start|>assistant\n<think>\n""" ```
GiteaMirror commented 2026-04-12 17:40:46 -05:00
Author
Owner
Copy Link

@sultanqasim commented on GitHub (Mar 8, 2025):

@yorktownting While those settings would give usable results, that doesn't fix the sampler ordering issue that makes chains of thought worse than they need to be. See https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively.

<!-- gh-comment-id:2707910793 --> @sultanqasim commented on GitHub (Mar 8, 2025): @yorktownting While those settings would give usable results, that doesn't fix the sampler ordering issue that makes chains of thought worse than they need to be. See <https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively>.
GiteaMirror commented 2026-04-12 17:40:47 -05:00
Author
Owner
Copy Link

@yorktownting commented on GitHub (Mar 10, 2025):

@yorktownting While those settings would give usable results, that doesn't fix the sampler ordering issue that makes chains of thought worse than they need to be. See https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively.虽然这些设置可以提供可用的结果，但这并不能解决使连贯思维比需要的更差的采样顺序问题。请参阅 https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively。

Thank you, I change the sampler order in my modelfile as follows, it do perform better:

FROM ./QwQ-32B-Q5_K_M.gguf
PARAMETER temperature 0.6
PARAMETER min_p 0.0
PARAMETER repeat_penalty 1.0
PARAMETER top_k 40
PARAMETER top_p 0.95
PARAMETER num_predict 32768
TEMPLATE """<|im_start|>user\n{{ .Prompt }}<|im_end|>\n<|im_start|>assistant\n<think>\n"""
<!-- gh-comment-id:2709194580 --> @yorktownting commented on GitHub (Mar 10, 2025): > [@yorktownting](https://github.com/yorktownting) While those settings would give usable results, that doesn't fix the sampler ordering issue that makes chains of thought worse than they need to be. See https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively.虽然这些设置可以提供可用的结果，但这并不能解决使连贯思维比需要的更差的采样顺序问题。请参阅 https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively。 Thank you, I change the sampler order in my modelfile as follows, it do perform better: ``` FROM ./QwQ-32B-Q5_K_M.gguf PARAMETER temperature 0.6 PARAMETER min_p 0.0 PARAMETER repeat_penalty 1.0 PARAMETER top_k 40 PARAMETER top_p 0.95 PARAMETER num_predict 32768 TEMPLATE """<|im_start|>user\n{{ .Prompt }}<|im_end|>\n<|im_start|>assistant\n<think>\n""" ```
GiteaMirror commented 2026-04-12 17:40:47 -05:00
Author
Owner
Copy Link

@Nukepayload2 commented on GitHub (Mar 17, 2025):

I have a question after reading https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively .

--samplers "top_k;top_p;min_p;temperature;dry;typ_p;xtc"

How can I set --samplers in modelfile?

I couldn't find it in https://github.com/ollama/ollama/blob/main/docs/modelfile.md#valid-parameters-and-values

<!-- gh-comment-id:2729307077 --> @Nukepayload2 commented on GitHub (Mar 17, 2025): I have a question after reading https://docs.unsloth.ai/basics/tutorial-how-to-run-qwq-32b-effectively . > --samplers "top_k;top_p;min_p;temperature;dry;typ_p;xtc" How can I set `--samplers` in `modelfile`? I couldn't find it in https://github.com/ollama/ollama/blob/main/docs/modelfile.md#valid-parameters-and-values
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#6245
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?