1 files changed, 0 insertions, 54 deletions
diff --git a/ggml-quants.c b/ggml-quants.c
index 5f2bcf69..31817b1c 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -3718,60 +3718,6 @@ void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
     quantize_row_q8_K_reference(x, y, k);
 }
 
-//===================================== Q8_K64 ==============================================
-
-void quantize_row_q8_K64_reference(const float * restrict x, block_q8_K64 * restrict y, int64_t k) {
-    assert(k % 64 == 0);
-    const int64_t nb = k / 64;
-
-    // Check if a row-wise scale works. It almost does, PPL is only ~0.02 higher
-    //float amax = 0;
-    //for (int j = 0; j < k; ++j) {
-    //    float ax = fabsf(x[j]);
-    //    amax = MAX(ax, amax);
-    //}
-
-    //float d = amax/127;
-    //float id = d ? 1/d : 0.f;
-
-    //for (int i = 0; i < nb; i++) {
-    //    for (int j = 0; j < 64; ++j) y[i].qs[j] = nearest_int(id*x[j]);
-    //    y[i].d = d;
-    //    x += 64;
-    //}
-
-    for (int i = 0; i < nb; i++) {
-
-        float max = 0;
-        float amax = 0;
-        for (int j = 0; j < 64; ++j) {
-            float ax = fabsf(x[j]);
-            if (ax > amax) {
-                amax = ax; max = x[j];
-            }
-        }
-        if (!amax) {
-            y[i].d = 0;
-            memset(y[i].qs, 0, 64);
-            x += 64;
-            continue;
-        }
-        //const float iscale = -128.f/max;
-        // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
-        const float iscale = -127.f/max;
-        for (int j = 0; j < 64; ++j) {
-            int v = nearest_int(iscale*x[j]);
-            y[i].qs[j] = MIN(127, v);
-        }
-        y[i].d = 1/iscale;
-        x += 64;
-    }
-}
-
-void quantize_row_q8_K64(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q8_K64_reference(x, y, k);
-}
-
 //===================================== Dot ptoducts =================================
 
 //