diff --git a/test/test_search.py b/test/test_search.py
index 48d6849b22fc..9f99d232e90f 100644
--- a/test/test_search.py
+++ b/test/test_search.py
@@ -69,6 +69,7 @@ def test_get_linearizer_actions(self):
     if Opt(OptOps.GROUPTOP, 0, 0) in actions:
       assert len([x for x in lins if x.applied_opts[0] == Opt(OptOps.GROUPTOP, axis=0, amt=3)]) == 0, "did not de-dup GROUPTOP"
 
+  @unittest.skipIf(Device.DEFAULT in {"NV"}, "Tries to open CUDA. #4607")
   def test_filter_global_buffer(self):
     # taken from https://github.com/tinygrad/tinygrad/issues/4612
     ast = LazyOp(op=BufferOps.STORE, src=(LazyOp(op=ReduceOps.MAX, src=(LazyOp(op=BinaryOps.MUL, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BinaryOps.ADD, src=(LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=1, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(384768,), strides=(1,), offset=0, mask=((0, 64128),), contiguous=False), View(shape=(1, 501, 256), strides=(0, 1, 501), offset=256512, mask=None, contiguous=False))))), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=2, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(384768,), strides=(1,), offset=-64128, mask=((64128, 128256),), contiguous=False), View(shape=(1, 501, 256), strides=(0, 1, 501), offset=256512, mask=None, contiguous=False)))))), arg=None), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=3, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(384768,), strides=(1,), offset=-128256, mask=((128256, 192384),), contiguous=False), View(shape=(1, 501, 256), strides=(0, 1, 501), offset=256512, mask=None, contiguous=False)))))), arg=None), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=4, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(384768,), strides=(1,), offset=-192384, mask=((192384, 256512),), contiguous=False), View(shape=(1, 501, 256), strides=(0, 1, 501), offset=256512, mask=None, contiguous=False)))))), arg=None), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=5, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(384768,), strides=(1,), offset=-256512, mask=((256512, 320640),), contiguous=False), View(shape=(1, 501, 256), strides=(0, 1, 501), offset=256512, mask=None, contiguous=False)))))), arg=None), LazyOp(op=BufferOps.LOAD, src=(), arg=MemBuffer(idx=6, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(384768,), strides=(1,), offset=-320640, mask=((320640, 384768),), contiguous=False), View(shape=(1, 501, 256), strides=(0, 1, 501), offset=256512, mask=None, contiguous=False)))))), arg=None), LazyOp(op=BufferOps.CONST, src=(), arg=ConstBuffer(val=1.4285714285714286, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 501, 256), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),))))), arg=None),), arg=(1,)),), arg=MemBuffer(idx=0, dtype=dtypes.float, st=ShapeTracker(views=(View(shape=(1, 1, 256), strides=(0, 0, 1), offset=0, mask=None, contiguous=True),))))  # noqa: E501
@@ -77,7 +78,8 @@ def test_filter_global_buffer(self):
     bufs = bufs_from_lin(lin)
     best_lin = beam_search(lin, bufs, 3)
     assert best_lin
-    tm = time_linearizer(best_lin, bufs, allow_test_size=False, cnt=2, disable_cache=False)
+    # need disable_cache to trigger.
+    tm = time_linearizer(best_lin, bufs, allow_test_size=False, cnt=2, disable_cache=True)
     assert tm
 
 if __name__ == '__main__':