[ET-VK][ops] Move group norm reduction lanes to Z

backends/vulkan/runtime/graph/ops/glsl/group_norm_reduce_texture.glsl

@@ -65,17 +65,16 @@ shared float shared_sum_sq[LOCAL_WORK_GROUP_SIZE];
  * N is the number of elements in the tensor buffer; each thread computes one
  * output element.
  *
- * Local work group size:  {1, float, 1}
- * float should be a power of 2, recommended 64 or 128 threads. This allows
- * efficient tree-based reduction in shared memory. Each local group will
- * cooperate to compute the output element.
+ * Local work group size:  {1, 1, 64}
+ * This allows efficient tree-based reduction in shared memory. Each local
+ * group cooperates to compute one output element.
  *
  * Each shader invocation will compute the mean and standard deviation for one
  * channel group in the input, and write out the corresponding result.
  */
 void group_norm_reduce_C_packed() {
   const int global_idx = int(gl_GlobalInvocationID.x);
-  const int local_idx = int(gl_LocalInvocationID.y);
+  const int local_idx = int(gl_LocalInvocationID.z);
 
   // Calculate group dimensions
   const int D = in_sizes.z / group;  // channels per group

backends/vulkan/runtime/graph/ops/impl/GroupNorm.cpp

@@ -29,13 +29,15 @@ utils::uvec3 group_norm_local_wg_size(
   (void)args;
   (void)resize_args;
 
-  return {1, 64, 1};
+  return {1u, 1u, 64u};
 }
 
 void resize_group_norm_texture_node(
     ComputeGraph* graph,
     const std::vector<ArgGroup>& args,
     const std::vector<ValueRef>& resize_args) {
+  VK_CHECK_COND(graph != nullptr);
+
   // Extract tensor references from args
   const ValueRef out = args.at(0).refs.at(0);
   const ValueRef in = args.at(1).refs.at(0);