Files
ik_llama.cpp/github-data/pull_requests/137 - Fix AVX2 implementation of iq4_nl_r4.md
2025-07-23 13:31:53 +02:00

486 B

🐛 #137 - Fix AVX2 implementation of iq4_nl_r4

Author ikawrakow
State Closed
Created 2024-12-11
Updated 2024-12-11

Description

The implementation was using _mm256_maddubs_epi16, which overflows (and gets saturated) with the unsigned version of the non-linear quants IQ4_NL lookup table. This PR fixes it without a noticeable performance loss.