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

llama.cpp issue #25618: speculative decoding (draft-mtp / draft-dspark) greedy output diverges from vanilla on quantized targets

forum

Fields

id
sources/qgotcha-llamacpp-25618-spec-quant-greedy-divergence
URL
https://github.com/ggml-org/llama.cpp/issues/25618
Publisher
llama.cpp GitHub issues
Kind
forum
Accessed
2026-10-03
notes
Reproducer matrix rules out ngram speculation, permanent weight corruption, FA, and vanilla Q4 nondeterminism. bf16 target matches 4/4 while Q4_K_M and Q4 diverge with MTP and DSpark.

Notes

Reproducer matrix rules out ngram speculation, permanent weight corruption, FA, and vanilla Q4 nondeterminism. bf16 target matches 4/4 while Q4_K_M and Q4 diverge with MTP and DSpark.

Cross-references

This record references no other records.