vulkan: support SET_ROWS (#14587)

* vulkan: support SET_ROWS

Add variants of the copy_to_quant shader that do the SET_ROWS operation.
Change these shaders to spread the work across the workgroup.
The memory access pattern is probably not great (one thread per quant block),
but should be fine for now.

* vulkan: optimize set_rows

Larger workgroups for non-quant types.
Set "norepeat" (there is manual repeat logic).
Use fastmod.

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -437,6 +437,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_set_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
@@ -2749,19 +2750,41 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+    }
+
+    if (device->float_controls_rte_fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F32],  "set_rows_f32",  set_rows_f32_rte_len,  set_rows_f32_rte_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F16],  "set_rows_f16",  set_rows_f16_rte_len,  set_rows_f16_rte_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_BF16], "set_rows_bf16", set_rows_bf16_rte_len, set_rows_bf16_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_0], "set_rows_q4_0", set_rows_q4_0_rte_len, set_rows_q4_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_1], "set_rows_q4_1", set_rows_q4_1_rte_len, set_rows_q4_1_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_0], "set_rows_q5_0", set_rows_q5_0_rte_len, set_rows_q5_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_1], "set_rows_q5_1", set_rows_q5_1_rte_len, set_rows_q5_1_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q8_0], "set_rows_q8_0", set_rows_q8_0_rte_len, set_rows_q8_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_IQ4_NL], "set_rows_iq4_nl", set_rows_iq4_nl_rte_len, set_rows_iq4_nl_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F32],  "set_rows_f32",  set_rows_f32_len,  set_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F16],  "set_rows_f16",  set_rows_f16_len,  set_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_BF16], "set_rows_bf16", set_rows_bf16_len, set_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_0], "set_rows_q4_0", set_rows_q4_0_len, set_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_1], "set_rows_q4_1", set_rows_q4_1_len, set_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_0], "set_rows_q5_0", set_rows_q5_0_len, set_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_1], "set_rows_q5_1", set_rows_q5_1_len, set_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q8_0], "set_rows_q8_0", set_rows_q8_0_len, set_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_IQ4_NL], "set_rows_iq4_nl", set_rows_iq4_nl_len, set_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
@@ -6527,6 +6550,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CONT:
     case GGML_OP_DUP:
         return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
+    case GGML_OP_SET_ROWS:
+        return ctx->device->pipeline_set_rows[dst->type];
     case GGML_OP_SILU_BACK:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_silu_back_f32;
@@ -6765,6 +6790,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_RMS_NORM:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_IM2COL:
+    case GGML_OP_SET_ROWS:
         return true;
     default:
         return false;
@@ -7078,6 +7104,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                     ne *= ggml_type_size(src0->type) / 2;
                 }
             }
+            // copy_to_quant has block size of 32, and each thread does QUANT_K elements.
+            // Splitting into 512x512xZ wouldn't work well since each workgroup does 1024 elements.
+            // So divide by block size here before splitting into 512x512 groups.
+            if (op == GGML_OP_CPY && !ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
+                ne = CEIL_DIV(ne, ggml_blck_size(dst->type));
+            }
             if (ne > 262144) {
                 elements = { 512, 512, CEIL_DIV(ne, 262144) };
             } else if (ne > 512) {
@@ -7086,6 +7118,25 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                 elements = { ne, 1, 1 };
             }
         } break;
+    case GGML_OP_SET_ROWS:
+        {
+            uint32_t ne = ggml_nelements(src0);
+            if (ggml_is_quantized(dst->type)) {
+                // quants run 32 threads each doing QUANT_K elements
+                ne = CEIL_DIV(ne, 32 * ggml_blck_size(dst->type));
+            } else {
+                // scalar types do one element per thread, running 512 threads
+                ne = CEIL_DIV(ne, 512);
+            }
+            if (ne > 262144) {
+                elements = { 512, 512, CEIL_DIV(ne, 262144) };
+            } else if (ne > 512) {
+                elements = { 512, CEIL_DIV(ne, 512), 1 };
+            } else {
+                elements = { ne, 1, 1 };
+            }
+        }
+        break;
     default:
         elements = { (uint32_t)ggml_nelements(src0), 1, 1 };
         break;
@@ -7648,6 +7699,21 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SET_ROWS, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    }, dryrun);
+}
+
 static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
@@ -8968,6 +9034,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_CPY:
+    case GGML_OP_SET_ROWS:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_SILU_BACK:
@@ -9034,6 +9101,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_CLAMP:
         case GGML_OP_PAD:
         case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
         case GGML_OP_CONT:
         case GGML_OP_DUP:
         case GGML_OP_SILU_BACK:
@@ -9142,6 +9210,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_DUP:
         ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_SET_ROWS:
+        ggml_vk_set_rows(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_SILU_BACK:
         ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9357,6 +9429,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_CPY:
+    case GGML_OP_SET_ROWS:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_SILU_BACK:
@@ -10422,9 +10495,20 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             } break;
         case GGML_OP_SET_ROWS:
             {
-                // TODO: add support
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
+                switch (op->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        return false;
+                }
             } break;
         case GGML_OP_CONT:
         case GGML_OP_CPY:
@@ -11039,6 +11123,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
         }
+    } else if (tensor->op == GGML_OP_SET_ROWS) {
+        tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONT) {
         tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
     } else if (tensor->op == GGML_OP_RESHAPE) {