fix main test failure bug (#3590)

Author: lanluo-nvidia

.github/workflows/linux-test.yml

@@ -155,7 +155,7 @@ jobs:
           path: ${{ env.RUNNER_TEST_RESULTS_DIR }}/**/*.xml
           summary: true
           display-options: fEs
-          fail-on-empty: false
+          fail-on-empty: true
 
       - name: Prepare artifacts for upload
         working-directory: ${{ inputs.repository }}

py/torch_tensorrt/dynamo/_defaults.py

@@ -1,5 +1,5 @@
 import os
-import pwd
+import platform
 import tempfile
 
 import torch
@@ -57,8 +57,18 @@
 L2_LIMIT_FOR_TILING = -1
 USE_DISTRIBUTED_MODE_TRACE = False
 OFFLOAD_MODULE_TO_CPU = False
+
+if platform.system() == "Linux":
+    import pwd
+
+    current_user = pwd.getpwuid(os.getuid())[0]
+else:
+    import getpass
+
+    current_user = getpass.getuser()
+
 DEBUG_LOGGING_DIR = os.path.join(
-    tempfile.gettempdir(), pwd.getpwuid(os.getuid())[0], "torch_tensorrt/debug_logs"
+    tempfile.gettempdir(), f"torch_tensorrt_{current_user}/debug_logs"
 )
 
 

py/torch_tensorrt/dynamo/conversion/impl/unsqueeze.py

@@ -22,11 +22,13 @@ def unsqueeze(
     input: TRTTensor,
     dim: int,
 ) -> TRTTensor:
-    from importlib.metadata import version
+    from importlib import metadata
 
-    if version("tensorrt") < "10.7.0":
+    from packaging.version import Version
+
+    if Version(metadata.version("tensorrt")) < Version("10.7.0"):
         logger.warning(
-            f"IUnsqueezeLayer is supported starting from TensorRT 10.7.0, using the old unsqueeze implementation in the current TensorRT version: {version('tensorrt')}"
+            f"IUnsqueezeLayer is supported starting from TensorRT 10.7.0, using the old unsqueeze implementation in the current TensorRT version: {metadata.version('tensorrt')}"
         )
         return unsqueeze_old(ctx, target, source_ir, name, input, dim)
     axes = get_trt_tensor(ctx, dim, f"{name}_axes")

py/torch_tensorrt/runtime/_cudagraphs.py

@@ -71,7 +71,7 @@ def __init__(self, compiled_module: torch.nn.Module) -> None:
         self.cudagraphs_module: Optional[CudaGraphsTorchTensorRTModule] = None
         self.old_module = None
 
-    def __enter__(self) -> torch.nn.Module | torch.fx.GraphModule:
+    def __enter__(self) -> Union[torch.nn.Module, torch.fx.GraphModule]:
 
         if isinstance(self.compiled_module, torch_tensorrt.MutableTorchTensorRTModule):
             self.old_module = self.compiled_module.gm
@@ -92,7 +92,7 @@ def __exit__(self, *args: Any) -> None:
 
 def get_cuda_graph_module(
     compiled_module: torch.fx.GraphModule,
-) -> torch.nn.Module | torch.fx.GraphModule:
+) -> Union[torch.nn.Module, torch.fx.GraphModule]:
     global _PY_RT_CUDAGRAPHS
 
     num_torch_module = 0

setup.py

@@ -429,6 +429,7 @@ def run(self):
     "torch_tensorrt.dynamo.conversion.impl.slice",
     "torch_tensorrt.dynamo.conversion.impl.unary",
     "torch_tensorrt.dynamo.conversion.plugins",
+    "torch_tensorrt.dynamo.debug",
     "torch_tensorrt.dynamo.lowering",
     "torch_tensorrt.dynamo.lowering.passes",
     "torch_tensorrt.dynamo.partitioning",
@@ -458,6 +459,7 @@ def run(self):
     "torch_tensorrt.dynamo.conversion.impl.slice": "py/torch_tensorrt/dynamo/conversion/impl/slice",
     "torch_tensorrt.dynamo.conversion.impl.unary": "py/torch_tensorrt/dynamo/conversion/impl/unary",
     "torch_tensorrt.dynamo.conversion.plugins": "py/torch_tensorrt/dynamo/conversion/plugins",
+    "torch_tensorrt.dynamo.debug": "py/torch_tensorrt/dynamo/debug",
     "torch_tensorrt.dynamo.lowering": "py/torch_tensorrt/dynamo/lowering",
     "torch_tensorrt.dynamo.lowering.passes": "py/torch_tensorrt/dynamo/lowering/passes",
     "torch_tensorrt.dynamo.partitioning": "py/torch_tensorrt/dynamo/partitioning",