fix: GdnCuTileEngine import path in cutile conftest and _pygraph error msg

python/cudnn/_pygraph.py

@@ -2039,7 +2039,7 @@ def push_output_attrs(out_t: Tensor, cpp_t: Any) -> None:
                     raise cudnn.cudnnGraphNotSupportedError(
                         f"[cudnn_frontend] Error: No valid engine configs for {method.upper()}: "
                         f"{method} has no cuDNN backend lowering; register a python engine "
-                        f"(e.g. cudnn.engines.GdnCuTileEngine)"
+                        f"(e.g. cudnn.linear_attention.cutile.GdnCuTileEngine)"
                     )
                 kw = {"name": node.name}
                 if not spec.get("no_cdt") and node.compute_data_type is not None:

test/python/linear_attention/cutile/conftest.py

@@ -12,7 +12,7 @@
 
 @pytest.fixture(autouse=True, scope="package")
 def _pin_cutile_engines():
-    from cudnn.engines import GdnCuTileEngine, KdaCuTileEngine
+    from cudnn.linear_attention.cutile import GdnCuTileEngine, KdaCuTileEngine
     from cudnn.linear_attention.ops import gdn, kda
 
     saved = gdn._engines, kda._engines

test/python/linear_attention/frost/test_gdn2_prefill_kernel.py

@@ -15,6 +15,7 @@
 import torch.nn.functional as F
 
 import cudnn  # noqa: F401  (conftest extends cudnn.__path__ with the source tree)
+from cudnn.linear_attention.frost import Gdn2FrostEngine
 
 from linear_attention.common import assert_bitwise_runs, assert_concurrent_stream_runs, assert_engine_declines
 from linear_attention.conftest import multidist_randu

test/python/linear_attention/frost/test_gdn_bprop_kernel.py

@@ -15,6 +15,7 @@
 import torch.nn.functional as F
 
 import cudnn  # noqa: F401  (conftest extends cudnn.__path__ with the source tree)
+from cudnn.linear_attention.frost import GdnFrostEngine
 
 from linear_attention.common import assert_bitwise_runs
 from linear_attention.conftest import multidist_randu

test/python/linear_attention/frost/test_kda_prefill_kernel.py

@@ -15,6 +15,7 @@
 import torch.nn.functional as F
 
 import cudnn  # noqa: F401  (conftest extends cudnn.__path__ with the source tree)
+from cudnn.linear_attention.frost import KdaFrostEngine
 
 from linear_attention.common import assert_bitwise_runs, assert_concurrent_stream_runs, assert_engine_declines
 from linear_attention.conftest import multidist_randu