From: lixing-star <redacted>
Date: Wed, 23 Jul 2025 06:39:51 +0000 (+0800)
Subject: ggml: fix loongarch quantize_row_q8_1 error (#14827)
X-Git-Tag: upstream/0.0.6073~106
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=6c88b3bb2509d980e6a64c50fdf8dd304929f770;p=pkg%2Fggml%2Fsources%2Fllama.cpp

ggml: fix loongarch quantize_row_q8_1 error (#14827)
---

diff --git a/ggml/src/ggml-cpu/arch/loongarch/quants.c b/ggml/src/ggml-cpu/arch/loongarch/quants.c
index 9e33fb32..7908da4d 100644
--- a/ggml/src/ggml-cpu/arch/loongarch/quants.c
+++ b/ggml/src/ggml-cpu/arch/loongarch/quants.c
@@ -544,7 +544,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x1 ));
         const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats