ggml: new optimization interface (ggml/988)

tests/test-backend-ops.cpp

@@ -811,11 +811,11 @@ struct test_case {
 
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx, gf, gb, false);
+        ggml_build_backward_expand(ctx, ctx, gb, false);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
             for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-                GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || t->grad->op != GGML_OP_NONE);
+                GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || ggml_graph_get_grad(gb, t)->op != GGML_OP_NONE);
             }
         }
 
@@ -862,7 +862,13 @@ struct test_case {
             const char * bn = ggml_backend_name(backend);
             const int64_t ne = ggml_nelements(t);
 
-            std::vector<float> ga = tensor_to_float(t->grad);
+            std::vector<float> ga;
+            struct ggml_tensor * grad = ggml_graph_get_grad(gb, t);
+            if (grad) {
+                ga = tensor_to_float(grad);
+            } else {
+                ga.resize(ne); // default value is 0.0f
+            }
 
             for (int64_t i = 0; i < ne; ++i) { // gradient algebraic
                 // check for nans
@@ -2500,6 +2506,35 @@ struct test_sum_rows : public test_case {
     }
 };
 
+// GGML_OP_MEAN
+struct test_mean : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_mean(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_param(ctx, a);
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_mean(ctx, a);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    float grad_eps() override {
+        return 0.1f * ne[0]*ne[1]*ne[2]*ne[3];
+    }
+};
+
 // GGML_OP_UPSCALE
 struct test_upscale : public test_case {
     const ggml_type type;
@@ -2834,24 +2869,14 @@ struct test_cross_entropy_loss : public test_case {
 struct test_opt_step_adamw : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
-    const float alpha;
-    const float beta1;
-    const float beta2;
-    const float eps;
-    const float wd;
 
     std::string vars() override {
-        return VARS_TO_STR7(type, ne, alpha, beta1, beta2, eps, wd);
+        return VARS_TO_STR2(type, ne);
     }
 
     test_opt_step_adamw(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 5, 4, 3},
-            float alpha = 1e-3f,
-            float beta1 = 0.9f,
-            float beta2 = 0.999f,
-            float eps = 1e-8f,
-            float wd = 0.0f)
-        : type(type), ne(ne), alpha(alpha), beta1(beta1), beta2(beta2), eps(eps), wd(wd) {}
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
+        : type(type), ne(ne) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
@@ -2861,7 +2886,16 @@ struct test_opt_step_adamw : public test_case {
         ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
         ggml_set_name(grad, "grad");
 
-        ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, alpha, beta1, beta2, eps, wd);
+        ggml_tensor * grad_m = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad_m, "grad_m");
+
+        ggml_tensor * grad_v = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad_v, "grad_v");
+
+        ggml_tensor * adamw_params = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 7);
+        ggml_set_name(adamw_params, "adamw_params");
+
+        ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, grad_m, grad_v, adamw_params);
         ggml_set_name(out, "out");
 
         return out;
@@ -2869,7 +2903,7 @@ struct test_opt_step_adamw : public test_case {
 
     void initialize_tensors(ggml_context * ctx) override {
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-            init_tensor_uniform(t, 0.0f, 1.0f); // grad_v needs non-negative values.
+            init_tensor_uniform(t, 0.0f, 1.0f); // grad_v and adamw_params need non-negative values.
         }
     }
 
@@ -3735,6 +3769,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
+    test_cases.emplace_back(new test_mean());
     test_cases.emplace_back(new test_upscale());
     test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
     test_cases.emplace_back(new test_upscale_ext());
@@ -3766,9 +3801,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     test_cases.emplace_back(new test_cross_entropy_loss());
-    for (float wd : {0.0f, 1e-2f}) {
-        test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}, 1.0f, 1e-3f, 0.9f, 0.999f, wd));
-    }
+    test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}));
 
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
@@ -3938,6 +3971,8 @@ int main(int argc, char ** argv) {
         ggml_backend_free(backend);
     }
 
+    ggml_quantize_free();
+
     printf("%zu/%zu backends passed\n", n_ok, ggml_backend_dev_count());
 
     if (n_ok != ggml_backend_dev_count()) {
@@ -3945,8 +3980,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    ggml_quantize_free();
-
     printf("\033[1;32mOK\033[0m\n");
     return 0;
 }