Fix delegate node metadata

Summary: The delegate node's metadata was set incorrectly, causing deserialization to fail

Reviewed By: mcr229

Differential Revision: D78350040

Author: angelayi

exir/backend/backend_api.py

@@ -235,7 +235,9 @@ def generate_debug_handle(ep: ExportedProgram) -> int:
             call_submodule_node.kwargs,
         )
         call_delegate_node.meta["debug_handle"] = generate_debug_handle(owning_program)
-        call_delegate_node.meta["val"] = submodule_output_node.meta["val"]
+        call_delegate_node.meta["val"] = [
+            out_arg.meta["val"] for out_arg in submodule_output_node.args[0]
+        ]
         call_submodule_node.replace_all_uses_with(call_delegate_node)
         owning_graph_module.graph.erase_node(call_submodule_node)
     if is_submodule:
@@ -476,7 +478,6 @@ def _create_partitions_in_graph_module(
         submodule_output_node = submodule.graph.output_node()
         # Copy the output node meta from the original output node, because
         # create_submodule_from_nodes doesn't cover the meta field
-        submodule_output_node.meta = tagged_graph_module_output_node.meta
         logging.debug(f"Partitioned graph module: {tagged_graph_module}")
         (
             submodule_program,

exir/backend/test/demos/test_xnnpack_qnnpack.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import tempfile
 import unittest
 
 import executorch.exir as exir
@@ -41,6 +42,18 @@
     prepare_fx,
 )
 
+from typing import Tuple
+
+import torch
+from executorch import exir
+from executorch.backends.xnnpack.partition.xnnpack_partitioner import (
+    XnnpackFloatingPointPartitioner,
+)
+from executorch.exir import (
+    EdgeCompileConfig,
+    EdgeProgramManager,
+    to_edge_transform_and_lower,
+)
 
 class TestXnnQnnBackends(unittest.TestCase):
     def test_add_xnnpack_and_dqlinear_qnn(self):
@@ -132,3 +145,42 @@ def forward(self, x, y):
         self.assertTrue(
             torch.allclose(model_output[0], ref_output, atol=1e-03, rtol=1e-03)
         )
+
+    def test_serde(self):
+        # The module with blank_logprobs() function
+        class BlankLogProbsModule(torch.nn.Module):
+            def __init__(self) -> None:
+                super().__init__()
+                self.linear = torch.nn.Linear(768, 1)
+                self.log_sigmoid = torch.nn.LogSigmoid()
+
+            def forward(self, joint_encodings: torch.Tensor) -> torch.Tensor:
+                tanh_out = torch.tanh(joint_encodings)
+                linear_out = self.linear(tanh_out)
+                blank_output = self.log_sigmoid(linear_out)
+                return blank_output
+
+        def get_blank_logprobs_inputs_fn() -> Tuple[torch.Tensor, ...]:
+            """
+            Get the input to the blank_logprobs() and nonblank_logprobs() functions.
+            """
+            return (torch.randn(1, 1, 1, 768),)
+
+        model = BlankLogProbsModule()
+        # Get the inputs for the logprobs function
+        logprobs_fake_inputs = get_blank_logprobs_inputs_fn()
+
+        # Export and partition
+        aten_prog = torch.export.export(model, logprobs_fake_inputs, strict=True)
+        partitioned_prog: EdgeProgramManager = to_edge_transform_and_lower(
+            aten_prog,
+            partitioner=[XnnpackFloatingPointPartitioner()],
+            compile_config=EdgeCompileConfig(_check_ir_validity=False, _use_edge_ops=True),
+        )
+
+        with tempfile.NamedTemporaryFile(suffix=".pt2") as f:
+            exir.save(partitioned_prog.exported_program(), f.name)
+            f.seek(0)
+            loaded_model = exir.load(f.name)
+        
+        self.assertTrue(torch.allclose(model(*logprobs_fake_inputs), loaded_model.module()(*logprobs_fake_inputs)))