Qwen3 0.6B loops with default settings

[iAdanos]

Android app, Qwen 0.6B loops 100 persent cases for me

[Juude]

is your sampler settings greedy？ the model has changed， if you downloaded the model earlier， the sampler settings may not be right， redownload the model or change sampler settings to mixed will fix the problem

[iAdanos]

No, sampler was "mixed" already.

Re-downloading model and setting options like the ones you provided didn't help.

I also tried playing with Min-P and Top-K, but it didn't help either.

I am testing the model with default (empty ?) system promt and an input Как варить пельмени?.

[Juude]

the output is very long.but finally seems output right

[iAdanos]

Very strange.

On my device (Samsung S23 Ultra) it still loops even after app update and model re-download:

JFYI, I stopped it looping manually.

[Juude]

Maybe switching to high inference precision could fix the problem. I’ll probably add that setting in the next release.

[iAdanos]

Changing Diffusion Memory Mode and updating an app again didn't change anything, BTW

[Juude]

Maybe switching to high inference precision could fix the problem. I’ll probably add that setting in the next release.

also try to set sampler to "penalty" and set the Penalty value to "1.2" for a try

[Juude]

Changing Diffusion Memory Mode and updating an app again didn't change anything, BTW

Diffusion Memory Mode will only affect stable diffusion models.

[iAdanos]

Maybe switching to high inference precision could fix the problem. I’ll probably add that setting in the next release.

How should I switch it?

[iAdanos]

Maybe switching to high inference precision could fix the problem. I’ll probably add that setting in the next release.

also try to set sampler to "penalty" and set the Penalty value to "1.2" for a try

Penalty sampler changed a picture a lot:

Without opencl an answer is much longer and it tried to get looped, but managed to get out:

With opencl it's much shorter (It would be great to have some guide for users on how such options change behavior), much slower, but it managed to not to loop on first shot:

[iAdanos]

For follow-ups, with opencl enabled it started thinking, but does not produce "out loud" (non-thinking) output.

[Juude]

@wangzhaode please explain dfferent behavior of opencl and cpu

[github-actions[bot]]

Marking as stale. No activity in 60 days.