bitnet: scale is per row, not per tensor

2 files changed, 3 insertions, 3 deletions

diff --git a/ggml-quants.c b/ggml-quants.c
index 552d6198..fc3155ef 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -3731,7 +3731,7 @@ void quantize_row_q8_K64_reference(const float * restrict x, block_q8_K64 * rest
         for (int j = 0; j < 64; ++j) {
             float ax = fabsf(x[j]);
             if (ax > amax) {
-                amax = ax; max = x[j]; 
+                amax = ax; max = x[j];
             }
         }
         if (!amax) {
diff --git a/iqk_mul_mat.cpp b/iqk_mul_mat.cpp
index fa1a0c83..ed8ed2e8 100644
--- a/iqk_mul_mat.cpp
+++ b/iqk_mul_mat.cpp
@@ -1347,12 +1347,12 @@ static void mul_mat_iq1bn_q8_K64(int n, const void * vx, size_t bx, const DataIn
     typedef union { float f; uint32_t i; } scale_t;
 
     scale_t scale;
-    uint16_t u = x[0].extra & 0xff;
-    scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
 
     for (int ix = 0; ix < nrc_x; ++ix) {
 
         x = (const block_iq1_bn *)((const char *)vx + ix*bx);
+        uint16_t u = x[0].extra & 0xff;
+        scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
 
         for (int iy = 0; iy < nrc_y; ++iy) accd[iy] = _mm256_setzero_ps();