diff --git a/llms/mlx_lm/examples/pipeline_generate.py b/llms/mlx_lm/examples/pipeline_generate.py
index 95f0fcb8..7f6f34db 100644
--- a/llms/mlx_lm/examples/pipeline_generate.py
+++ b/llms/mlx_lm/examples/pipeline_generate.py
@@ -51,7 +51,7 @@ mx.eval(model.parameters())
 
 # Synchronize processes before generation to avoid timeout if downloading
 # model for the first time.
-mx.eval(mx.distributed.all_sum(mx.array(1.0)))
+mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu))
 
 
 def rprint(*args, **kwargs):