resolve against #15489, sync after clearing partial sums

Author: jeffbolznv

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

@@ -7953,7 +7953,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     if (op == GGML_OP_ADD || op == GGML_OP_RMS_NORM) {
         vk_buffer d_A = ctx->do_add_rms_partials ? ctx->prealloc_add_rms_partials : d_X;
         size_t a_buf_offset = ctx->do_add_rms_partials ? ctx->prealloc_size_add_rms_partials_offset : 0;
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
             { vk_subbuffer{ d_X, x_buf_offset, x_sz },
               vk_subbuffer{ d_Y, y_buf_offset, y_sz },
@@ -11297,6 +11296,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         }
         // initialize partial sums to zero.
         ggml_vk_buffer_memset_async(compute_ctx, ctx->prealloc_add_rms_partials, 0, 0, ctx->prealloc_size_add_rms_partials);
+        ggml_vk_sync_buffers(ctx, compute_ctx);
     }
 
     // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.