Added: tri, cumsum. Still a mess.

Author: pwilkin

ggml/include/ggml.h

@@ -243,6 +243,8 @@
 
 #define GGML_MROPE_SECTIONS   4
 
+#define GGML_DELTA_NET_CHUNK    64
+
 #define GGML_UNUSED(x) (void)(x)
 #ifdef __CUDACC__
 template<typename... Args>
@@ -472,6 +474,7 @@ extern "C" {
         GGML_OP_COS,
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
+        GGML_OP_CUMSUM,
         GGML_OP_MEAN,
         GGML_OP_ARGMAX,
         GGML_OP_COUNT_EQUAL,
@@ -527,6 +530,7 @@ extern "C" {
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
         GGML_OP_LEAKY_RELU,
+        GGML_OP_TRI,
 
         GGML_OP_FLASH_ATTN_EXT,
         GGML_OP_FLASH_ATTN_BACK,
@@ -539,6 +543,7 @@ extern "C" {
         GGML_OP_RWKV_WKV6,
         GGML_OP_GATED_LINEAR_ATTN,
         GGML_OP_RWKV_WKV7,
+        GGML_OP_DELTA_NET,
 
         GGML_OP_UNARY,
 
@@ -612,6 +617,13 @@ extern "C" {
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
+    enum ggml_tri_type {
+        GGML_TRI_TYPE_UPPER_DIAG        = 0,
+        GGML_TRI_TYPE_UPPER             = 1,
+        GGML_TRI_TYPE_LOWER_DIAG        = 2,
+        GGML_TRI_TYPE_LOWER             = 3
+    };
+
     struct ggml_init_params {
         // memory pool
         size_t mem_size;   // bytes
@@ -975,6 +987,10 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_cumsum(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
     // mean along rows
     GGML_API struct ggml_tensor * ggml_mean(
             struct ggml_context * ctx,
@@ -2119,6 +2135,17 @@ extern "C" {
             int                   shift2,
             int                   shift3);
 
+    // Make matrix into a triangular one (upper, upper + diagonal, lower or lower + diagonal) with constant value
+    GGML_API struct ggml_tensor * ggml_tri(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 constant,
+            enum ggml_tri_type    tritype);
+
+    GGML_API struct ggml_tensor * ggml_tri_keep(
+            struct ggml_context * ctx,
+            struct ggml_tensor * a,
+            enum ggml_tri_type tritype);
 
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
@@ -2289,6 +2316,18 @@ extern "C" {
             struct ggml_tensor  * b,
             struct ggml_tensor  * state);
 
+    GGML_API struct ggml_tensor * ggml_delta_net(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * g,
+            struct ggml_tensor  * beta,
+            struct ggml_tensor  * state,
+            bool                  use_qk_l2norm,
+            float                 scale,
+            float                 eps_norm);
+
     // custom operators
 
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1731,6 +1731,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_sum_rows(params, tensor);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                ggml_compute_forward_cumsum(params, tensor);
+            } break;
         case GGML_OP_MEAN:
             {
                 ggml_compute_forward_mean(params, tensor);
@@ -1943,6 +1947,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
             } break;
+        case GGML_OP_TRI:
+            {
+                ggml_compute_forward_tri(params, tensor);
+            } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_compute_forward_flash_attn_ext(params, tensor);
@@ -1998,6 +2006,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv7(params, tensor);
             } break;
+        case GGML_OP_DELTA_NET:
+            {
+                ggml_compute_forward_delta_net_f32(params, tensor);
+            } break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2153,6 +2165,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_ARGMAX:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
             {
                 n_tasks = 1;
             } break;
@@ -2297,6 +2311,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_GET_REL_POS:
+        case GGML_OP_DELTA_NET:
             {
                 n_tasks = 1;
             } break;