fix: remove legacy conv converter

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -2454,6 +2454,8 @@ def aten_ops_le(
 def conv_param_validator(
     conv_node: Node, settings: Optional[CompilationSettings] = None
 ) -> bool:
+
+    # return True
     return conv_node.args[7] in ([0], [0, 0], [0, 0, 0])
 
 
@@ -2505,6 +2507,7 @@ def aten_ops_convolution(
             stride=args[3],
             padding=args[4],
             dilation=args[5],
+            # output_padding=args[7],
             groups=args[8],
         )
 

py/torch_tensorrt/fx/converters/aten_ops_converters.py

@@ -103,60 +103,60 @@ def aten_ops_batch_norm(
     )
 
 
-@tensorrt_converter(torch.ops.aten.convolution.default)
-def aten_ops_convolution(
-    network: TRTNetwork,
-    target: Target,
-    args: Tuple[Argument, ...],
-    kwargs: Dict[str, Argument],
-    name: str,
-) -> Union[TRTTensor, Sequence[TRTTensor]]:
-    kwargs_new = {
-        "input": args[0],
-        "weight": args[1],
-        "bias": args[2],
-        "stride": args[3],
-        "padding": args[4],
-        "dilation": args[5],
-        "groups": args[8],
-    }
-    # we do not handle transposed.
-    if args[6] is True:
-        raise RuntimeError(f"Target {target} does not support `transposed=True` ")
-    # we do not handle output_padding.
-    if args[7] not in ([0], [0, 0], [0, 0, 0]):
-        raise RuntimeError(f"Target {target} has non-0 output_padding")
-
-    if len(kwargs_new["stride"]) == 1:
-        return convolution.convNd(
-            network,
-            target,
-            source_ir=SourceIR.ATEN,
-            name=name,
-            is_conv1d=True,
-            input_val=kwargs_new["input"],
-            weight=kwargs_new["weight"],
-            bias=kwargs_new["bias"],
-            stride=kwargs_new["stride"],
-            padding=kwargs_new["padding"],
-            dilation=kwargs_new["dilation"],
-            groups=kwargs_new["groups"],
-        )
-    else:
-        return convolution.convNd(
-            network,
-            target,
-            source_ir=SourceIR.ATEN,
-            name=name,
-            is_conv1d=False,
-            input_val=kwargs_new["input"],
-            weight=kwargs_new["weight"],
-            bias=kwargs_new["bias"],
-            stride=kwargs_new["stride"],
-            padding=kwargs_new["padding"],
-            dilation=kwargs_new["dilation"],
-            groups=kwargs_new["groups"],
-        )
+# @tensorrt_converter(torch.ops.aten.convolution.default)
+# def aten_ops_convolution(
+#     network: TRTNetwork,
+#     target: Target,
+#     args: Tuple[Argument, ...],
+#     kwargs: Dict[str, Argument],
+#     name: str,
+# ) -> Union[TRTTensor, Sequence[TRTTensor]]:
+#     kwargs_new = {
+#         "input": args[0],
+#         "weight": args[1],
+#         "bias": args[2],
+#         "stride": args[3],
+#         "padding": args[4],
+#         "dilation": args[5],
+#         "groups": args[8],
+#     }
+#     # we do not handle transposed.
+#     if args[6] is True:
+#         raise RuntimeError(f"Target {target} does not support `transposed=True` ")
+#     # we do not handle output_padding.
+#     if args[7] not in ([0], [0, 0], [0, 0, 0]):
+#         raise RuntimeError(f"Target {target} has non-0 output_padding")
+
+#     if len(kwargs_new["stride"]) == 1:
+#         return convolution.convNd(
+#             network,
+#             target,
+#             source_ir=SourceIR.ATEN,
+#             name=name,
+#             is_conv1d=True,
+#             input_val=kwargs_new["input"],
+#             weight=kwargs_new["weight"],
+#             bias=kwargs_new["bias"],
+#             stride=kwargs_new["stride"],
+#             padding=kwargs_new["padding"],
+#             dilation=kwargs_new["dilation"],
+#             groups=kwargs_new["groups"],
+#         )
+#     else:
+#         return convolution.convNd(
+#             network,
+#             target,
+#             source_ir=SourceIR.ATEN,
+#             name=name,
+#             is_conv1d=False,
+#             input_val=kwargs_new["input"],
+#             weight=kwargs_new["weight"],
+#             bias=kwargs_new["bias"],
+#             stride=kwargs_new["stride"],
+#             padding=kwargs_new["padding"],
+#             dilation=kwargs_new["dilation"],
+#             groups=kwargs_new["groups"],
+#         )
 
 
 @tensorrt_converter(torch.ops.aten.div.default)

run_aot_moria_2dunet.py

@@ -0,0 +1,43 @@
+from monai.networks.nets import UNet
+import torch
+import torch_tensorrt
+
+device = "cuda:0"
+
+# Define the 2D U-Net model
+model = UNet(
+    spatial_dims=2,
+    in_channels=3,
+    out_channels=2,
+    channels=(16, 32, 64, 128),
+    strides=(2, 2, 2),
+    num_res_units=2,
+    act="relu",
+    norm="batch",
+    dropout=0.1,
+).to(device).half().eval()
+
+# (batch size, channels, height, width)
+input_tensor = torch.randn(1, 3, 256, 256, device=device).half()
+
+backend = "torch_tensorrt"
+
+# Compile the model with Torch-TensorRT backend
+model = torch.compile(
+    model,
+    backend=backend,
+    options={
+        "use_python_runtime": False,
+        "enabled_precisions": {torch.float16},
+        "truncate_double": True,
+        "debug": True,
+        "min_block_size": 1,
+    },
+    dynamic=False,
+)
+
+# Perform inference with the compiled model
+with torch.no_grad():
+    output = model(input_tensor)
+
+print(output)