diff --git a/xpra/codecs/cuda_common/cuda_context.py b/xpra/codecs/cuda_common/cuda_context.py
index 67f9ad8ff1..9c253ec368 100755
--- a/xpra/codecs/cuda_common/cuda_context.py
+++ b/xpra/codecs/cuda_common/cuda_context.py
@@ -10,13 +10,16 @@
 import os
 from time import monotonic
 from threading import RLock
-import pycuda               #@UnresolvedImport
-from pycuda import driver   #@UnresolvedImport
 
+from xpra.codecs.nv_util import numpy_import_lock
 from xpra.util import engs, print_nested_dict, envint, csv, first_time
 from xpra.os_util import load_binary_file
 from xpra.log import Logger
 
+with numpy_import_lock:
+    import pycuda               #@UnresolvedImport
+    from pycuda import driver   #@UnresolvedImport
+
 log = Logger("cuda")
 
 MIN_FREE_MEMORY = envint("XPRA_CUDA_MIN_FREE_MEMORY", 10)
diff --git a/xpra/codecs/nvfbc/cuda_image_wrapper.py b/xpra/codecs/nvfbc/cuda_image_wrapper.py
index b754ba72e1..ba63e4a4f6 100644
--- a/xpra/codecs/nvfbc/cuda_image_wrapper.py
+++ b/xpra/codecs/nvfbc/cuda_image_wrapper.py
@@ -5,7 +5,6 @@
 # later version. See the file COPYING for details.
 
 from time import monotonic
-from pycuda import driver       #@UnresolvedImport
 
 from xpra.codecs.nv_util import numpy_import_lock
 from xpra.codecs.image_wrapper import ImageWrapper
@@ -15,6 +14,7 @@
 
 with numpy_import_lock:
     import numpy
+    from pycuda import driver       #@UnresolvedImport
 
 
 class CUDAImageWrapper(ImageWrapper):