diff --git a/tinygrad/device.py b/tinygrad/device.py
index 0c6c7f84f5..2b9babb560 100644
--- a/tinygrad/device.py
+++ b/tinygrad/device.py
@@ -99,9 +99,13 @@ class Buffer:
 def _internal_buffer_copy(dest:Buffer, src:Buffer):
   if hasattr(src.allocator, 'transfer') and type(dest.allocator) is type(src.allocator):  # noqa: E721
     # fast path, used on HIP between GPUs
-    # NOTE: it's important we use the dest device here to ensure the transfer is ready
+    # NOTE: we have to block here so the data isn't copied too early. this is probably due to buffer reuse
+    if hasattr(src.d, "block") and hasattr(dest.d, "event"): src.d.block(dest.d.event())
+    else: dest.d.synchronize()
     src.allocator.transfer(dest._buf, src._buf, dest.size*dest.dtype.itemsize)
+    # NOTE: we have to block here so the data is ready on dest when dest needs it
     if hasattr(dest.d, "block") and hasattr(src.d, "event"): dest.d.block(src.d.event())
+    else: src.d.synchronize()
     return
   if getenv("FROM_BUFFER") and hasattr(dest.allocator, 'from_buffer') and hasattr(dest.allocator, 'transfer') and hasattr(src.allocator, 'as_buffer'):
     # fast path, used on Metal in OS X Sonoma