Hi! I am new to local AI. Got a HP SFF server and added a Low profile RTX 3050 6GB to play around with AI. My expectations are according to my hardware. My use case is actually for Paperless-AI and Inbox Zero, which should be reasonably achievable. Now I am also playing around with OpenWeb UI and Perplexica, both working as expected. Weird issue: I got Qwen3.5:4B looping, asking it how to plant peas :D It started questioning itself, if it got the right peas, questioning itself all the time.

Wait, wait... Actually,  "Zuckereerbsen" is often used as a synonym for  "Mintpea". 
Okay, proceed. 

Wait, hold on. I need to check if there is any specific variety called "Zuckereeherbe" which is actually a  Garden Pea. 
Yes, they are essentially the same family. 
Okay, proceed. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a specific product:  "Zuckereeherbe" (Sugar Bean) from some seed companies. They are indeed peas. 
Okay, proceed. 

Wait, final check: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed. 

Wait, one last thing: Is there any chance they meant  "Zuckererbsen" (which is actually  Zuckerrübe)? No. 
Okay, write it. 

Wait, wait... Wait, I found a reference:  "Zuckereerbsen" (Sugar Peas) are sometimes grown for their sweet flesh, similar to "Snow Peas". 
Okay, proceed.

Somebody of you got experience with that? Does this happen often? Whats the best way to prevent it? It basically leads to a complete block and high energy usage of my graphics card.

  • iceberg314@midwest.social
    1·
    58 minutes ago

    I thought there was a parameter too that penalizes repetition. It usually in the same area as where you can adjust the temperature

  • Sims@lemmy.ml
    1·
    14 hours ago

    I would try a less quantized version. Sometimes inaccurate weights compound and a loop can occur.

    Bigger cloud models can do it too when they get perplexed/overwhelmed, or have countering instructions etc. It is possible to detect repeated sentences and ‘snap it out of it’ by inserting a rough/LOUD! order, or even mindful words like ‘caaaalm down… close your eyes and breath slowly…’ into the context. …or both. Feels odd, but the llm just reacts to the human language/intend anyway…

  • phanto@lemmy.ca
    1·
    20 hours ago

    I had the exact same issue recently. I was asking it about CPU vs GPU utilization on an LLM, and it started looking until it crashed. Might be a limitation of the model? I was running it in ollama. Tried the same prompt on the 7b version on a different machine, and it has no issue.