Reddit reports GLM-5.2 crashes with all-NaN logits past 32k to 64k context

naunen · reddit · 2026-07-26

A Reddit user reports that GLM-5.2 runs normally at 8k context on a mixed CPU/GPU setup, but crashes as soon as context grows past roughly 32k–64k.

They describe an all-NaN failure in llama-sampling.cpp on the first generated token, after trying multiple fixes: toggling DSA and flash attention, shrinking batch sizes, moving KV cache to CPU, switching KV cache precision, and pulling recent GLM-DSA indexer fixes. The user suspects an FP16 overflow in the DSA/indexer path and asks whether there is any workaround short of CPU-only inference.

Original post →

More from Models

Models channel →