ggml : fix CPU implementation

Author: ggerganov

ggml.c

@@ -10337,19 +10337,17 @@ static void ggml_compute_forward_out_prod(
 static void ggml_compute_forward_scale_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
-    GGML_ASSERT(ggml_is_scalar(src1));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
     // scale factor
-    const float v = *(float *) src1->data;
+    const float v = *(float *) dst->op_params;
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -10380,12 +10378,11 @@ static void ggml_compute_forward_scale_f32(
 static void ggml_compute_forward_scale(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_scale_f32(params, src0, src1, dst);
+                ggml_compute_forward_scale_f32(params, src0, dst);
             } break;
         default:
             {
@@ -14395,7 +14392,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_SCALE:
             {
-                ggml_compute_forward_scale(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_scale(params, tensor->src[0], tensor);
             } break;
         case GGML_OP_SET:
             {

tests/test-backend-ops.cpp

@@ -766,18 +766,19 @@ struct test_bin_bcast : public test_case {
 struct test_scale : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
+    float scale;
 
     std::string vars() override {
-        return VARS_TO_STR2(type, ne);
+        return VARS_TO_STR3(type, ne, scale);
     }
 
     test_scale(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 10, 10, 10})
-        : type(type), ne(ne) {}
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            float scale = 2.0f)
+        : type(type), ne(ne), scale(scale) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * scale = ggml_new_tensor_1d(ctx, type, 1);
         ggml_tensor * out = ggml_scale(ctx, a, scale);
         return out;
     }

tests/test-grad0.cpp

@@ -887,13 +887,14 @@ int main(int argc, const char ** argv) {
             ne2[0] = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
+                const float s = -1.0f + 2.0f*frand();
+
                 ggml_set_param(ctx0, x[0]);
                 ggml_set_param(ctx0, x[1]);
 
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], x[1]));
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], s));
 
                 check_gradient("scale", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
             }
@@ -1395,7 +1396,7 @@ int main(int argc, const char ** argv) {
                                                 ggml_add1(ctx0,
                                                     ggml_scale(ctx0,
                                                         ggml_soft_max(ctx0, x[0]),
-                                                        ggml_new_f32(ctx0, 1.0f - eps)),
+                                                        1.0f - eps),
                                                     ggml_new_f32(ctx0, eps))));
 
                 check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);