Skip TRTInterpreter.validate_conversion when partition already proved full support.

py/torch_tensorrt/dynamo/_compiler.py

@@ -1351,6 +1351,9 @@ def compile_module(
     num_supported_ops, total_ops = partitioning.get_graph_converter_support(
         gm, settings.torch_executed_ops
     )
+    skip_conversion_validation = (
+        settings.require_full_compilation and num_supported_ops == total_ops
+    )
 
     dryrun_tracker.total_ops_in_graph = total_ops
     dryrun_tracker.supported_ops_in_graph = num_supported_ops
@@ -1422,6 +1425,7 @@ def contains_metadata(gm: torch.fx.GraphModule) -> bool:
             settings=settings,
             name="_run_on_acc_0",
             engine_cache=engine_cache,
+            skip_conversion_validation=skip_conversion_validation,
         )
 
     # Store the original input spec for later use
@@ -1622,6 +1626,7 @@ def preserve_module_specs(
                 settings=settings,
                 name=name,
                 engine_cache=engine_cache,
+                skip_conversion_validation=skip_conversion_validation,
             )
 
             trt_modules[name] = trt_module

py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py

@@ -168,6 +168,7 @@ def __init__(
         input_binding_names: Optional[Sequence[str]] = None,
         output_binding_names: Optional[Sequence[str]] = None,
         _debugger_config: Optional[DebuggerConfig] = None,
+        skip_conversion_validation: bool = False,
     ):
         super().__init__(module)
 
@@ -194,12 +195,16 @@ def __init__(
         # an xdist worker), making those ops incorrectly appear as disallowed.
         CONVERTERS.set_compilation_settings(compilation_settings)
         self.validate_compile_settings()
-        missing_ops = self.validate_conversion()
-        if missing_ops:
-            warnings.warn(
-                "Interpretation will fail due to missing operations \n"
-                + "\n".join(f"{i}" for i in missing_ops)
-            )
+        # compile_module already walked CONVERTERS via get_graph_converter_support.
+        # When that count was complete under require_full_compilation, this extra
+        # get() pass cannot find missing ops; skip it.
+        if not skip_conversion_validation:
+            missing_ops = self.validate_conversion()
+            if missing_ops:
+                warnings.warn(
+                    "Interpretation will fail due to missing operations \n"
+                    + "\n".join(f"{i}" for i in missing_ops)
+                )
 
         # Optimization profiles. Profiles are an ordered list on
         # ``Input.profiles``; profile index i is built from each input's

py/torch_tensorrt/dynamo/conversion/_conversion.py

@@ -212,13 +212,16 @@ def interpret_module_to_result(
     *,
     input_binding_names: Optional[Sequence[str]] = None,
     output_binding_names: Optional[Sequence[str]] = None,
+    skip_conversion_validation: bool = False,
 ) -> SerializedInterpreterResult:
     """Interpret an FX module to a TRTInterpreterResult
     Args:
         module: FX GraphModule to interpret
         inputs: It requires a sequence of FLATTENED Inputs representing inputs to the module. It should include both arg_inputs and kwarg_inputs, if applicable.
         settings: Compilation settings
         engine_cache: Engine cache instance
+        skip_conversion_validation: If True, skip TRTInterpreter.validate_conversion.
+            Set by compile_module when require_full_compilation already proved full converter coverage.
     Returns:
         SerializedInterpreterResult
     """
@@ -291,6 +294,7 @@ def interpret_module_to_result(
         engine_cache=engine_cache,
         input_binding_names=input_binding_names,
         output_binding_names=output_binding_names,
+        skip_conversion_validation=skip_conversion_validation,
     )
 
     interpreter_result = interpreter.run()
@@ -349,6 +353,7 @@ def convert_module(
     settings: CompilationSettings = CompilationSettings(),
     name: str = "",
     engine_cache: Optional[BaseEngineCache] = None,
+    skip_conversion_validation: bool = False,
 ) -> TorchTensorRTModule:
     """Convert an FX module to a TRT module
     Args:
@@ -357,11 +362,16 @@ def convert_module(
         settings: Compilation settings
         name: TRT engine name
         engine_cache: Engine cache instance
+        skip_conversion_validation: If True, skip TRTInterpreter.validate_conversion.
     Returns:
         TorchTensorRTModule
     """
     serialized_interpreter_result = interpret_module_to_result(
-        module, inputs, settings, engine_cache=engine_cache
+        module,
+        inputs,
+        settings,
+        engine_cache=engine_cache,
+        skip_conversion_validation=skip_conversion_validation,
     )
 
     if not ENABLED_FEATURES.torch_tensorrt_runtime: