ggml : sync (unary ops refactor, static-correctness) (#2370)

* ggml : sync (unary ops, tests)

ggml-ci

* tests : remove unnecessary funcs

ggml-cuda.cu

@@ -3962,18 +3962,23 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             }
             func = ggml_cuda_mul;
             break;
-        case GGML_OP_GELU:
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(tensor)) {
+                case GGML_UNARY_OP_GELU:
                     if (!any_on_device) {
                         return false;
                     }
                     func = ggml_cuda_gelu;
                     break;
-        case GGML_OP_SILU:
+                case GGML_UNARY_OP_SILU:
                     if (!any_on_device) {
                         return false;
                     }
                     func = ggml_cuda_silu;
                     break;
+                default:
+                    return false;
+            } break;
         case GGML_OP_NORM:
             if (!any_on_device) {
                 return false;

ggml-metal.m

@@ -519,7 +519,9 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
-                    case GGML_OP_SILU:
+                    case GGML_OP_UNARY:
+                        switch (ggml_get_unary_op(gf->nodes[i])) {
+                            case GGML_UNARY_OP_SILU:
                                 {
                                     if (encoder == nil) {
                                         encoder = [command_buffer computeCommandEncoder];
@@ -533,7 +535,7 @@ void ggml_metal_graph_compute(
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
-                    case GGML_OP_RELU:
+                            case GGML_UNARY_OP_RELU:
                                 {
                                     if (encoder == nil) {
                                         encoder = [command_buffer computeCommandEncoder];
@@ -547,7 +549,7 @@ void ggml_metal_graph_compute(
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
-                    case GGML_OP_GELU:
+                            case GGML_UNARY_OP_GELU:
                                 {
                                     if (encoder == nil) {
                                         encoder = [command_buffer computeCommandEncoder];
@@ -561,6 +563,12 @@ void ggml_metal_graph_compute(
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
+                            default:
+                                {
+                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_ASSERT(false);
+                                }
+                        } break;
                     case GGML_OP_SOFT_MAX:
                         {
                             if (encoder == nil) {
@@ -979,10 +987,12 @@ void ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     default:
+                        {
                             fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                             GGML_ASSERT(false);
                         }
                 }
+            }
 
             if (encoder != nil) {
                 [encoder endEncoding];

ggml.h

@@ -330,16 +330,6 @@ extern "C" {
         GGML_OP_ARGMAX,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
-        GGML_OP_ABS,
-        GGML_OP_SGN,
-        GGML_OP_NEG,
-        GGML_OP_STEP,
-        GGML_OP_TANH,
-        GGML_OP_ELU,
-        GGML_OP_RELU,
-        GGML_OP_GELU,
-        GGML_OP_GELU_QUICK,
-        GGML_OP_SILU,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
@@ -378,6 +368,8 @@ extern "C" {
         GGML_OP_WIN_PART,
         GGML_OP_WIN_UNPART,
 
+        GGML_OP_UNARY,
+
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
@@ -391,6 +383,18 @@ extern "C" {
         GGML_OP_COUNT,
     };
 
+    enum ggml_unary_op {
+        GGML_UNARY_OP_ABS,
+        GGML_UNARY_OP_SGN,
+        GGML_UNARY_OP_NEG,
+        GGML_UNARY_OP_STEP,
+        GGML_UNARY_OP_TANH,
+        GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_QUICK,
+        GGML_UNARY_OP_SILU,
+    };
 
     // ggml object
     struct ggml_object {
@@ -535,6 +539,7 @@ extern "C" {
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
+    GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -558,6 +563,7 @@ extern "C" {
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
     GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
     GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
@@ -617,9 +623,11 @@ extern "C" {
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API const char *         ggml_get_name(const struct ggml_tensor * tensor);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
-    GGML_API struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name);
+
-    GGML_API struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * fmt, ...);
+    GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
     // operations on tensors with backpropagation
@@ -629,6 +637,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_dup_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_add(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -952,11 +965,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // a -> b, in-place, return view(b)
+    GGML_API struct ggml_tensor * ggml_cpy_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, in-place
+    GGML_API struct ggml_tensor * ggml_cont_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1268,6 +1292,16 @@ extern "C" {
     typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
     typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
 
+    GGML_API struct ggml_tensor * ggml_unary(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_unary_op op);
+
+    GGML_API struct ggml_tensor * ggml_unary_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_unary_op op);
+
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,

tests/test-grad0.c

@@ -64,7 +64,7 @@ void get_random_dims(int64_t * dims, int ndims) {
     }
 }
 
-struct ggml_tensor * get_random_tensor(
+struct ggml_tensor * get_random_tensor_f32(
         struct ggml_context * ctx0,
         int ndims,
         int64_t ne[],
@@ -112,7 +112,55 @@ struct ggml_tensor * get_random_tensor(
     return result;
 }
 
-struct ggml_tensor * get_random_tensor_int(
+struct ggml_tensor * get_random_tensor_f16(
+        struct ggml_context * ctx0,
+        int ndims,
+        int64_t ne[],
+        float fmin,
+        float fmax) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F16, ndims, ne);
+
+    switch (ndims) {
+        case 1:
+            for (int i0 = 0; i0 < ne[0]; i0++) {
+                ((ggml_fp16_t *)result->data)[i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < ne[1]; i1++) {
+                for (int i0 = 0; i0 < ne[0]; i0++) {
+                    ((ggml_fp16_t *)result->data)[i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+                }
+            }
+            break;
+        case 3:
+            for (int i2 = 0; i2 < ne[2]; i2++) {
+                for (int i1 = 0; i1 < ne[1]; i1++) {
+                    for (int i0 = 0; i0 < ne[0]; i0++) {
+                        ((ggml_fp16_t *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+                    }
+                }
+            }
+            break;
+        case 4:
+            for (int i3 = 0; i3 < ne[3]; i3++) {
+                for (int i2 = 0; i2 < ne[2]; i2++) {
+                    for (int i1 = 0; i1 < ne[1]; i1++) {
+                        for (int i0 = 0; i0 < ne[0]; i0++) {
+                            ((ggml_fp16_t *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            assert(false);
+    };
+
+    return result;
+}
+
+struct ggml_tensor * get_random_tensor_i32(
         struct ggml_context * ctx0,
         int ndims,
         int64_t ne[],
@@ -160,23 +208,6 @@ struct ggml_tensor * get_random_tensor_int(
     return result;
 }
 
-float get_element(const struct ggml_tensor * t, int idx) {
-    if (t->type == GGML_TYPE_F32) {
-        return ((float *)t->data)[idx];
-    }
-
-    if (t->type == GGML_TYPE_I32) {
-        return ((int32_t *)t->data)[idx];
-    }
-
-    assert(false);
-    return INFINITY;
-}
-
-void set_element(struct ggml_tensor * t, int idx, float value) {
-    ((float *)t->data)[idx] = value;
-}
-
 void print_elements(const char* label, const struct ggml_tensor * t) {
     if (!t) {
         printf("%s: %s = null\n", __func__, label);
@@ -186,7 +217,7 @@ void print_elements(const char* label, const struct ggml_tensor * t) {
     printf("%s: %s = [", __func__, label);
     for (int k = 0; k < nelements; ++k) {
         if (k > 0) { printf(", "); }
-        printf("%.5f", get_element(t, k));
+        printf("%.5f", ggml_get_f32_1d(t, k));
     }
     printf("] shape: [");
     for (int k = 0; k < t->n_dims; ++k) {
@@ -237,23 +268,23 @@ bool check_gradient(
         const int nelements = ggml_nelements(x[i]);
         for (int k = 0; k < nelements; ++k) {
             // compute gradient using finite differences
-            const float x0 = get_element(x[i], k);
+            const float x0 = ggml_get_f32_1d(x[i], k);
             const float xm = x0 - eps;
             const float xp = x0 + eps;
-            set_element(x[i], k, xp);
+            ggml_set_f32_1d(x[i], k, xp);
 
             ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
             const float f0 = ggml_get_f32_1d(f, 0);
 
-            set_element(x[i], k, xm);
+            ggml_set_f32_1d(x[i], k, xm);
 
             ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
             const float f1 = ggml_get_f32_1d(f, 0);
             const float g0 = (f0 - f1)/(2.0f*eps);
 
-            set_element(x[i], k, x0);
+            ggml_set_f32_1d(x[i], k, x0);
 
             // compute gradient using backward graph
             ggml_graph_reset  (&gf);
@@ -261,7 +292,7 @@ bool check_gradient(
 
             ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
 
-            const float g1 = get_element(x[i]->grad, k);
+            const float g1 = ggml_get_f32_1d(x[i]->grad, k);
 
             const float error_abs = fabsf(g0 - g1);
             const float error_rel = g0 != 0 ? fabsf(g0 - g1)/fabsf(g0) : 0;
@@ -392,19 +423,35 @@ int main(int argc, const char ** argv) {
 
         struct ggml_tensor * x[MAX_NARGS];
 
-        // add
+        // add f32
         {
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
 
-                check_gradient("add", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f);
+                check_gradient("add f32", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f);
+            }
+        }
+
+        // add f16
+        {
+            const int nargs = 2;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f16(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
+
+                check_gradient("add f16", ctx0, x, f, ndims, nargs, 1e-1f, 2e-1f, 2e-1f);
             }
         }
 
@@ -414,7 +461,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -430,7 +477,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -446,7 +493,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, 0.5f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 0.5f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -462,7 +509,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -478,7 +525,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -494,7 +541,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -510,7 +557,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -527,7 +574,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -537,6 +584,40 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // mean, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_mean(ctx0, x[0]));
+
+                check_gradient("mean", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // argmax
+        if (0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_argmax(ctx0, x[0]));
+
+                check_gradient("argmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
         // repeat
         {
             int64_t ne2[4];
@@ -549,15 +630,36 @@ int main(int argc, const char ** argv) {
 
             const int nargs = 1;
             for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[1], ggml_repeat(ctx0, x[0], x[1]))));
 
                 check_gradient("repeat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
             }
+        }
 
+        // repeat back
+        {
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            ne2[0] = ne[0] * ne2[0];
+            ne2[1] = ne[1] * ne2[1];
+            ne2[2] = 1;
+            ne2[3] = 1;
+
+            const int nargs = 1;
+            for (int ndims = 1; ndims <= 2; ++ndims) {
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[0], ggml_repeat_back(ctx0, x[1], x[0]))));
+
+                check_gradient("repeat back", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
+            }
         }
 
         // abs (finite differences do not work)
@@ -566,7 +668,7 @@ int main(int argc, const char ** argv) {
 
         //    for (int ndims = 1; ndims <= 2; ++ndims) {
         //        for (int i = 0; i < nargs; ++i) {
-        //            x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+        //            x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
         //            ggml_set_param(ctx0, x[i]);
         //        }
 
@@ -576,17 +678,82 @@ int main(int argc, const char ** argv) {
         //    }
         //}
 
+        // sgn
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_sgn(ctx0, x[0]));
+
+                check_gradient("sgn", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // neg
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_neg(ctx0, x[0]));
+
+                check_gradient("neg", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // step
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_step(ctx0, x[0]));
+
+                check_gradient("step", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // tanh, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_tanh(ctx0, x[0]));
+
+                check_gradient("tanh", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
         // mul_mat
         {
             const int nargs = 2;
 
             for (int ndims = 2; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 {
                     int64_t ne2[4];
                     get_random_dims(ne2, 4);
                     ne2[0] = ne[0];
-                    x[1] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                    x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 }
 
                 ggml_set_param(ctx0, x[0]);
@@ -602,13 +769,63 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // elu, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_elu(ctx0, x[0]));
+
+                check_gradient("elu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
+        // relu
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_relu(ctx0, x[0]));
+
+                check_gradient("relu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            }
+        }
+
+        // gelu, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 1;
+
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+
+                struct ggml_tensor* f = ggml_sum(ctx0, ggml_gelu(ctx0, x[0]));
+
+                check_gradient("gelu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+            }
+        }
+
         // silu
         {
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -629,7 +846,7 @@ int main(int argc, const char ** argv) {
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
 
@@ -647,8 +864,8 @@ int main(int argc, const char ** argv) {
             ne2[0] = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
                 ggml_set_param(ctx0, x[1]);
@@ -659,20 +876,37 @@ int main(int argc, const char ** argv) {
             }
         }
 
-        // cpy
+        // cpy f32
         {
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
                 for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                     ggml_set_param(ctx0, x[i]);
                 }
                 // x[1] is overwritten by x[0], so the gradients don't propagate to x[1]
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
 
-                check_gradient("cpy", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("cpy f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            }
+        }
+
+        // cpy f16
+        {
+            const int nargs = 2;
+
+            for (int ndims = 1; ndims <= 2; ++ndims) {
+                for (int i = 0; i < nargs; ++i) {
+                    x[i] = get_random_tensor_f16(ctx0, ndims, ne, -1.0f, 1.0f);
+                    ggml_set_param(ctx0, x[i]);
+                }
+                // x[1] is overwritten by x[0], so the gradients don't propagate to x[1]
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
+
+                check_gradient("cpy f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
             }
         }
 
@@ -689,8 +923,8 @@ int main(int argc, const char ** argv) {
                 for (int i = 0; i < ndims; ++i) {
                     ne2[0] *= ne[i];
                 }
-                x[0] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
 
@@ -712,8 +946,8 @@ int main(int argc, const char ** argv) {
                 for (int i = 0; i < ndims; ++i) {
                     ne2[0] *= ne[i];
                 }
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
 
@@ -729,7 +963,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 1);
@@ -737,7 +971,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 1);
                 }
 
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 const int max_offset = MAX(0, ggml_nelements(x[0]) - ggml_nelements(x[1]));
@@ -758,7 +992,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 2; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 2);
@@ -766,7 +1000,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 2);
                 }
 
-                x[1] = get_random_tensor(ctx0, 2, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 2, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -790,7 +1024,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 3; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 3);
@@ -798,7 +1032,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 3);
                 }
 
-                x[1] = get_random_tensor(ctx0, 3, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 3, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -824,7 +1058,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 4; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 4);
@@ -832,7 +1066,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 4);
                 }
 
-                x[1] = get_random_tensor(ctx0, 4, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -858,7 +1092,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 2;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 1);
@@ -866,7 +1100,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 1);
                 }
 
-                x[1] = get_random_tensor(ctx0, 1, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 const int max_offset = MAX(0, ggml_nelements(x[0]) - ggml_nelements(x[1]));
@@ -887,7 +1121,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 2; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 get_random_dims(ne2, 2);
@@ -895,7 +1129,7 @@ int main(int argc, const char ** argv) {
                     get_random_dims(ne2, 2);
                 }
 
-                x[1] = get_random_tensor(ctx0, 2, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, 2, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[1]);
 
                 max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
@@ -915,7 +1149,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
 
@@ -941,7 +1175,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 get_random_dims(ne2, 2);
                 while (ne2[0]*ne2[1] > ggml_nelements(x[0])) {
@@ -971,7 +1205,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
-                x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
                 get_random_dims(ne2, 3);
                 while (ne2[0]*ne2[1]*ne2[2] > ggml_nelements(x[0])) {
@@ -1010,7 +1244,7 @@ int main(int argc, const char ** argv) {
                 for (int i=ndims; i<4; ++i) {
                     ne2[i] = 1;
                 }
-                x[0] = get_random_tensor(ctx0, 4, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
 
@@ -1043,7 +1277,7 @@ int main(int argc, const char ** argv) {
                 for (int i=ndims; i<4; ++i) {
                     ne2[i] = 1;
                 }
-                x[0] = get_random_tensor(ctx0, 4, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
 
                 ggml_set_param(ctx0, x[0]);
 
@@ -1060,8 +1294,8 @@ int main(int argc, const char ** argv) {
             int64_t ne3[4] = {1+irand(ne[1]), 1, 1, 1};
             const int nargs = 1;
             const int ndims = 2;
-            x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+            x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-            x[1] = get_random_tensor_int(ctx0, 1, ne3, 0, ne2[1]);
+            x[1] = get_random_tensor_i32(ctx0, 1, ne3, 0, ne2[1]);
 
             ggml_set_param(ctx0, x[0]);
 
@@ -1075,7 +1309,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             const int ndims = 2;
 
-            x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+            x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
             ggml_set_param(ctx0, x[0]);
 
             int n_past = irand(ne[0]);
@@ -1090,7 +1324,7 @@ int main(int argc, const char ** argv) {
             const int nargs = 1;
             const int ndims = 2;
 
-            x[0] = get_random_tensor(ctx0, ndims, ne, -1.0f, 1.0f);
+            x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
             ggml_set_param(ctx0, x[0]);
 
             int n_past = irand(ne[0]);
@@ -1108,7 +1342,7 @@ int main(int argc, const char ** argv) {
             get_random_dims(ne2, 4);
 
             for (int ndims = 1; ndims <= 3; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_soft_max(ctx0, x[0]));
@@ -1125,8 +1359,8 @@ int main(int argc, const char ** argv) {
             get_random_dims(ne2, 4);
 
             for (int ndims = 1; ndims <= 3; ++ndims) {
-                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
@@ -1136,7 +1370,7 @@ int main(int argc, const char ** argv) {
             }
         }
 
-        // rope
+        // rope f32
         {
             const int nargs = 1;
 
@@ -1148,7 +1382,7 @@ int main(int argc, const char ** argv) {
             for (int ndims = 3; ndims <= 4; ++ndims) {
                 for (int mode = 0; mode < 4; ++mode) {
                     for (int n_past = 1; n_past < ne2[2]; ++n_past) {
-                        x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                        x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
 
                         ggml_set_param(ctx0, x[0]);
 
@@ -1163,14 +1397,48 @@ int main(int argc, const char ** argv) {
 
                         struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
 
-                        GGML_PRINT_DEBUG("rope: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
+                        GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
-                        check_gradient("rope", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
+                        check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
                     }
                 }
             }
         }
 
-        // flash_attn
+        // rope f16
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+            ne2[0] += ne2[0] % 2;
+            int n_rot = ne2[0];
+
+            for (int ndims = 3; ndims <= 4; ++ndims) {
+                for (int mode = 0; mode < 4; ++mode) {
+                    for (int n_past = 1; n_past < ne2[2]; ++n_past) {
+                        x[0] = get_random_tensor_f16(ctx0, ndims, ne2, -1.0f, 1.0f);
+
+                        ggml_set_param(ctx0, x[0]);
+
+                        const bool skip_past = (mode & 1);
+                        if (skip_past) {
+                            // we have no past, so this would have to work on uninitialized memory.
+                            // we only test the gradients here;
+                            // skip_past should have no influence on gradient computation.
+                            // so when other modes work, we assume that this does as well.
+                            continue;
+                        }
+
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+
+                        GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
+                        check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
+                    }
+                }
+            }
+        }
+
+        // flash_attn f32
         {
             const int nargs = 3;
 
@@ -1196,16 +1464,57 @@ int main(int argc, const char ** argv) {
                         nek[3] = 1;
                         nev[3] = 1;
                     }
-                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
                     ggml_set_param(ctx0, x[0]);
                     ggml_set_param(ctx0, x[1]);
                     ggml_set_param(ctx0, x[2]);
 
                     struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
 
-                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                    check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
+
+        // flash_attn f16, not yet fully implemented
+        if(0)
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor_f16(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor_f16(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor_f16(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
                 }
             }
         }

tests/test-opt.c

@@ -125,9 +125,9 @@ int main(void) {
     };
     struct ggml_context * ctx = ggml_init(params);
 
-    int64_t ne1[4] = {4, 1024, 1, 1};
+    int64_t ne1[4] = {4, 128, 1, 1};
-    int64_t ne2[4] = {4, 2048, 1, 1};;
+    int64_t ne2[4] = {4, 256, 1, 1};;
-    int64_t ne3[4] = {1024, 2048, 1, 1};
+    int64_t ne3[4] = {128, 256, 1, 1};
 
     struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
     struct ggml_tensor * b = get_random_tensor(ctx, 2, ne2, -1, +1);