unsloth

/

GLM-4.7-Flash-GGUF

like 607

Follow

Unsloth AI 19.2k

I'm noticing significantly worse quality at ~70k tokens and more frequent looping and full de-railing at ~95k tokens, when using MXFP4 version with llama.cpp. Can someone help me with the correct scaling parameters / ROPE or similar that needs to be set in llama.cpp?