inference/research
verifiedhigh · 0.90updated 2026-10-03

llama.cpp issue #27623: Qwen3.8-27B hybrid Gated DeltaNet decode collapses ~25x past ~80K context while pp stays fast

forum

Fields

id
sources/gotcha-llamacpp-issue-27623-gated-deltanet-decode-collapse
URL
https://github.com/ggml-org/llama.cpp/issues/27623
Publisher
llama.cpp GitHub issues
Kind
forum
Published
2026-08-23
Accessed
2026-10-03
notes
Reproduction matrix over -c, -fa and -ctk/-ctv KV quant types on both native Windows and WSL2.

Notes

Reproduction matrix over -c, -fa and -ctk/-ctv KV quant types on both native Windows and WSL2.

Cross-references

This record references no other records.