diff --git a/tools/Polygraphy/CHANGELOG.md b/tools/Polygraphy/CHANGELOG.md
index d6da60ea3..957750031 100644
--- a/tools/Polygraphy/CHANGELOG.md
+++ b/tools/Polygraphy/CHANGELOG.md
@@ -13,6 +13,8 @@ Dates are in YYYY-MM-DD format.
 
 ### Fixed
 - Fixed issue when `polygraphy multi-device shard` would exceed python recursive depth limit on large models.
+- Fixed a bug where `polygraphy convert` would fail when the input model was already a
+  TensorRT engine, e.g. `polygraphy convert model.engine --convert-to trt -o out.plan`.
 - Fixed a bug where setting `POLYGRAPHY_ASK_BEFORE_INSTALL=0` would enable asking before installing dependencies instead of disabling it.
 
 
diff --git a/tools/Polygraphy/polygraphy/tools/args/backend/trt/loader.py b/tools/Polygraphy/polygraphy/tools/args/backend/trt/loader.py
index e78b8fa38..87cf7db9a 100644
--- a/tools/Polygraphy/polygraphy/tools/args/backend/trt/loader.py
+++ b/tools/Polygraphy/polygraphy/tools/args/backend/trt/loader.py
@@ -712,14 +712,14 @@ def add_to_script_impl(self, script, network_name=None):
 
     def load_engine_bytes(self, network=None):
         """
-        Loads a TensorRT engine according to arguments provided on the command-line.
+        Loads serialized TensorRT engine bytes according to arguments provided on the command-line.
 
         Args:
             network (Tuple[trt.Builder, trt.INetworkDefinition, Optional[parser]]):
                     A tuple containing a TensorRT builder, network and optionally parser.
 
         Returns:
-            tensorrt.ICudaEngine: The engine.
+            bytes: The serialized engine.
         """
         loader = args_util.run_script(self.add_to_script, network)
         return loader()
diff --git a/tools/Polygraphy/polygraphy/tools/convert/convert.py b/tools/Polygraphy/polygraphy/tools/convert/convert.py
index 50ba59bb5..6fb7e3bbf 100644
--- a/tools/Polygraphy/polygraphy/tools/convert/convert.py
+++ b/tools/Polygraphy/polygraphy/tools/convert/convert.py
@@ -101,11 +101,11 @@ def run_impl(self, args):
             model = self.arg_groups[OnnxLoadArgs].load_onnx()
             self.arg_groups[OnnxSaveArgs].save_onnx(model, args.output)
         elif convert_type.is_trt():
-            with self.arg_groups[
+            serialized_engine = self.arg_groups[
                 TrtLoadEngineBytesArgs
-            ].load_engine_bytes() as serialized_engine:
-                self.arg_groups[TrtSaveEngineBytesArgs].save_engine_bytes(
-                    serialized_engine, args.output
-                )
+            ].load_engine_bytes()
+            self.arg_groups[TrtSaveEngineBytesArgs].save_engine_bytes(
+                serialized_engine, args.output
+            )
         else:
             G_LOGGER.critical(f"Cannot convert to model type: {convert_type}")
diff --git a/tools/Polygraphy/tests/tools/test_convert.py b/tools/Polygraphy/tests/tools/test_convert.py
index 19f313deb..5739bf8a8 100644
--- a/tools/Polygraphy/tests/tools/test_convert.py
+++ b/tools/Polygraphy/tests/tools/test_convert.py
@@ -145,6 +145,19 @@ def test_modify_onnx_outputs(self, poly_convert):
             model = onnx.load(outmodel.name)
             assert len(model.graph.output) == 2
 
+    def test_engine_to_engine(self, poly_convert):
+        # Engine inputs yield plain bytes, not an object that supports the
+        # context manager protocol.
+        engine_bytes = b"serialized-engine"
+        with util.NamedTemporaryFile(
+            "w+b", suffix=".engine"
+        ) as inmodel, util.NamedTemporaryFile(suffix=".plan") as outmodel:
+            inmodel.write(engine_bytes)
+            inmodel.flush()
+
+            poly_convert([inmodel.name, "--model-type=engine", "-o", outmodel.name])
+            assert BytesFromPath(outmodel.name)() == engine_bytes
+
 
 class TestConvertToOnnxLikeTrt:
     @pytest.mark.parametrize(