ggml : fix quants nans when all the group weights are very close to zero (#7313)

1 files changed, 18 insertions, 12 deletions

diff --git a/ggml-quants.c b/ggml-quants.c
index 9b291d52..7008e5d8 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -14,6 +14,12 @@
 #include <stdlib.h> // for qsort
 #include <stdio.h>  // for GGML_ASSERT
 
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid warnings for hundreds of casts
 // we should just be careful :)
@@ -1109,7 +1115,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (amax < 1e-30f) { // all zero
+    if (amax < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) {
             L[i] = 0;
         }
@@ -1177,7 +1183,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (!amax) { // all zero
+    if (amax < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) { L[i] = 0; }
         return 0.f;
     }
@@ -1646,7 +1652,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t *
             break;
         }
     }
-    return sumlx / suml2;
+    return sumlx/suml2;
 }
 
 static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
@@ -2653,7 +2659,7 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         }
 
-        if (!max_abs_scale) {
+        if (max_abs_scale < GROUP_MAX_EPS) {
             memset(&y[i], 0, sizeof(block_q6_K));
             y[i].d = GGML_FP32_TO_FP16(0.f);
             x += QK_K;
@@ -2805,7 +2811,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
 
         }
 
-        if (!max_abs_scale) {
+        if (max_abs_scale < GROUP_MAX_EPS) {
             memset(&y[i], 0, sizeof(block_q6_K));
             y[i].d = GGML_FP32_TO_FP16(0.f);
             x += QK_K;
@@ -12599,7 +12605,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
                 memset(L, 0, 32);
                 continue;
@@ -12775,7 +12781,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
                 memset(L, 0, 16);
                 continue;
@@ -13216,7 +13222,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ3_XXS) {
                 scales[ib] = 0;
                 memset(L, 0, 32);
                 continue;
@@ -13756,7 +13762,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
             float max = fabsf(xb[0]);
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ1_S) {
                 scales[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
@@ -13944,7 +13950,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
             }
             float max = fabsf(xb[0]);
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ1_M) {
                 scales[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
@@ -14208,7 +14214,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
                 amax = ax; max = xb[j];
             }
         }
-        if (!amax) {
+        if (amax < GROUP_MAX_EPS) {
             scales[ib] = 0;
             continue;
         }
@@ -14429,7 +14435,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ2_S) {
                 scales[ib] = 0;
                 continue;
             }