ggml : sync with ggml repo (warning fixes + asserts)

Author: ggerganov

ggml.c

@@ -8245,8 +8245,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
         ggml_fp16_t * d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
         ggml_fp16_t * d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
         float       * d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
-#else
-        float * const wdata = params->wdata;
 #endif
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -8263,15 +8261,20 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                             wdata[id++] = GGML_FP32_TO_FP16(*(float *) ((char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11 + i00*nb10));
                         }
                     }
+
+                    assert(id*sizeof(ggml_fp16_t) <= params->wsize);
                 }
 #else
+                float * const wdata = params->wdata;
                 {
                     size_t id = 0;
                     for (int64_t i01 = 0; i01 < ne01; ++i01) {
                         for (int64_t i00 = 0; i00 < ne00; ++i00) {
                             wdata[id++] = GGML_FP16_TO_FP32(*(ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00));
                         }
                     }
+
+                    assert(id*sizeof(float) <= params->wsize);
                 }
 #endif
 
@@ -8537,7 +8540,10 @@ static void ggml_compute_forward_mul_mat_q_f32(
                         dequantize_row_q((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
                         id += ne00;
                     }
+
+                    assert(id*sizeof(float) <= params->wsize);
                 }
+
                 const float * x = wdata;
 #endif
 
@@ -9118,7 +9124,7 @@ static void ggml_compute_forward_alibi_f32(
     //const int nb3 = src0->nb[3];
 
     assert(nb0 == sizeof(float));
-    assert(ne1+n_past == ne0);
+    assert(ne1 + n_past == ne0); (void) n_past;
 
     // add alibi to src0 (KQ_scaled)
     const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@@ -9179,7 +9185,7 @@ static void ggml_compute_forward_alibi_f16(
     //const int nb3 = src0->nb[3];
 
     assert(nb0 == sizeof(ggml_fp16_t));
-    assert(ne1+n_past == ne0);
+    assert(ne1 + n_past == ne0); (void) n_past;
 
     // add alibi to src0 (KQ_scaled)
     const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@@ -11571,12 +11577,12 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1; // TODO: this actually is doing nothing
                                                    //       the threads are still spinning
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
+#if defined(GGML_USE_CUBLAS)
+                                // with cuBLAS, we need memory for the full 3D / 4D data of src1
+                                cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
+#else
                                 // here we need memory just for single 2D matrix from src0
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
-#else
-                                // with GPU, we need memory for the full 3D / 4D data
-                                cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*MAX(ggml_nelements(node->src1), ggml_nelements(node->src0));
 #endif
                             } else {
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
@@ -11586,7 +11592,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 #endif
                         } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
                             cur = 0;
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
                             }

ggml.h

@@ -701,8 +701,8 @@ extern "C" {
             struct ggml_tensor  * c1);
 
     // Mapping operations
-    GGML_API typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
-    GGML_API typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+    typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
+    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
 
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,