I often play with samplers and text template with llama.cpp, but recently I found that newer models are very repetitive in their output, I chucked it to a stricter training and moved on.
Now I decided to give gemma 4 a go, and the 26B A4B was looping so I started by checking smaplers since I often run with weirder settings but not matter what I changed, the output did not change.
Even setting it to the extreme values, like temp 1000 with no other samplers, the output is coherent, which no matter what, it should not be.
Is it me, or are samplers somewhat broken?
[link] [comments]