1 files changed, 18 insertions, 7 deletions

diff --git a/tools/cnn_test.cc b/tools/cnn_test.cc
index 826d9ea..c504c3d 100644
--- a/tools/cnn_test.cc
+++ b/tools/cnn_test.cc
@@ -111,6 +111,12 @@ static bool parse_args(int argc, char** argv, Args* args) {
               args->cnn_version);
     }
     args->cnn_version = 2;
+
+    // Warn if --layers was specified (binary file config takes precedence)
+    if (args->num_layers != 3) {  // 3 is the default
+      fprintf(stderr, "WARNING: --layers %d ignored (--weights loads layer config from .bin)\n",
+              args->num_layers);
+    }
   }
 
   return true;
@@ -122,11 +128,11 @@ static void print_usage(const char* prog) {
   fprintf(stderr, "\nOPTIONS:\n");
   fprintf(stderr, "  --blend F                Final blend amount (0.0-1.0, default: 1.0)\n");
   fprintf(stderr, "  --format ppm|png         Output format (default: png)\n");
-  fprintf(stderr, "  --layers N               Number of CNN layers (1-10, default: 3)\n");
+  fprintf(stderr, "  --layers N               Number of CNN layers (1-10, default: 3, ignored with --weights)\n");
   fprintf(stderr, "  --save-intermediates DIR Save intermediate layers to directory\n");
   fprintf(stderr, "  --debug-hex              Print first 8 pixels as hex (debug)\n");
-  fprintf(stderr, "  --cnn-version N          CNN version: 1 (default) or 2\n");
-  fprintf(stderr, "  --weights PATH           Load weights from .bin file (forces CNN v2)\n");
+  fprintf(stderr, "  --cnn-version N          CNN version: 1 (default) or 2 (ignored with --weights)\n");
+  fprintf(stderr, "  --weights PATH           Load weights from .bin (forces CNN v2, overrides layer config)\n");
   fprintf(stderr, "  --help                   Show this help\n");
 }
 
@@ -678,15 +684,20 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
            info.out_channels, info.weight_count);
   }
 
-  // Create weights storage buffer
+  // Create weights storage buffer (skip header + layer info, upload only weights)
+  size_t header_size = 20;  // 5 u32
+  size_t layer_info_size = 20 * layer_info.size();  // 5 u32 per layer
+  size_t weights_offset = header_size + layer_info_size;
+  size_t weights_only_size = weights_size - weights_offset;
+
   WGPUBufferDescriptor weights_buffer_desc = {};
-  weights_buffer_desc.size = weights_size;
+  weights_buffer_desc.size = weights_only_size;
   weights_buffer_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
   weights_buffer_desc.mappedAtCreation = false;
 
   WGPUBuffer weights_buffer =
       wgpuDeviceCreateBuffer(device, &weights_buffer_desc);
-  wgpuQueueWriteBuffer(queue, weights_buffer, 0, weights_data, weights_size);
+  wgpuQueueWriteBuffer(queue, weights_buffer, 0, weights_data + weights_offset, weights_only_size);
 
   // Create input view
   const WGPUTextureViewDescriptor view_desc = {
@@ -1045,7 +1056,7 @@ static bool process_cnn_v2(WGPUDevice device, WGPUQueue queue,
 
     layer_bg_entries[3].binding = 3;
     layer_bg_entries[3].buffer = weights_buffer;
-    layer_bg_entries[3].size = weights_size;
+    layer_bg_entries[3].size = weights_only_size;
 
     layer_bg_entries[4].binding = 4;
     layer_bg_entries[4].buffer = layer_params_buffers[i];