From 1f14f50dfdd251a4d02a4fd241bce6af924a152d Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Mon, 27 Oct 2025 11:39:18 +0200
Subject: [PATCH] Try removing copy indirection

---
 ggml/src/ggml-cuda.cu      | 6 +++---
 ggml/src/ggml-cuda/cpy.cuh | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index b47b7dd5..b21d95d7 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -3558,7 +3558,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_CPY) {
+        if (false && node->op == GGML_OP_CPY) {
 
             // Store the pointers which are updated for each token, such that these can be sent
             // to the device and accessed using indirection from CUDA graph
@@ -3602,7 +3602,7 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
 
 static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
     if (node->data != graph_node_properties->node_address &&
-          node->op != GGML_OP_CPY &&
+          //node->op != GGML_OP_CPY &&
           node->op != GGML_OP_VIEW) {
         return false;
     }
@@ -3623,7 +3623,7 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (node->src[i] &&
             node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_CPY &&
+            //node->op != GGML_OP_CPY &&
             node->op != GGML_OP_VIEW
         ) {
             return false;
diff --git a/ggml/src/ggml-cuda/cpy.cuh b/ggml/src/ggml-cuda/cpy.cuh
index 0bd3c0c6..9bd34b7c 100644
--- a/ggml/src/ggml-cuda/cpy.cuh
+++ b/ggml/src/ggml-cuda/cpy.cuh
@@ -2,7 +2,7 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = true); //false);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);