Read from constants if it exists

test/stablehlo/test_export_llama.py

@@ -39,7 +39,8 @@ def test_llama_export(self):
     arg = (torch.randint(0, 1000, (8, 100)), torch.arange(0, 100), None)
     options = StableHLOExportOptions()
     options.override_tracing_arguments = arg
-    exported = torch.export.export(model, arg)
+    with torch.no_grad():
+      exported = torch.export.export(model, arg)
     with tempfile.TemporaryDirectory() as tempdir:
       save_as_stablehlo(exported, tempdir, options)
 

torch_xla/stablehlo.py

@@ -303,6 +303,12 @@ def _exported_program_to_stablehlo_bundle(exported_model,
   param_and_buffer_keys = exported_model.graph_signature.parameters + exported_model.graph_signature.buffers
   state_dict = pytree.tree_map_only(torch.Tensor, lambda x: x.to(device=device),
                                     exported_model.state_dict)
+
+  if (constants := getattr(exported_model, 'constants')) is not None:
+    state_dict.update(
+        pytree.tree_map_only(torch.Tensor, lambda x: x.to(device=device),
+                             constants))
+
   param_buffer_values = (state_dict[key] for key in param_and_buffer_keys)
 
   if hasattr(exported_model.graph_signature, "lifted_tensor_constants"):

torch_xla/tf_saved_model_integration.py

@@ -33,7 +33,7 @@ def inner(*args):
     call_args = stablehlo._extract_call_parameters(args, func.meta, bundle)
     return tfxla.call_module(
         tuple(call_args),
-        version=5,
+        version=6,
         Tout=Touts,  # dtype information
         Sout=Souts,  # Shape information
         function_list=[],