Add logs for helping debug address space overflow issue

Author: jackzhxng

exir/emit/_emitter.py

@@ -79,6 +79,7 @@
     TensorShapeDynamism,
 )
 from executorch.exir.tensor import (
+    AddressSpaceOverflowException,
     layout_enum,
     make_allocation_info,
     make_tensor_value,
@@ -349,7 +350,20 @@ def _tensor_spec_to_evalue(self, spec: TensorSpec) -> EValue:
                 self.node,
                 f"Non-const tensor should be an activation tensor: mem_offset {spec.mem_offset}",
             )
-            allocation_info = make_allocation_info(spec.mem_id, spec.mem_offset)
+            try:
+                allocation_info = make_allocation_info(spec.mem_id, spec.mem_offset)
+            except AddressSpaceOverflowException as e:
+                raise InternalError(
+                    self._emit_node_specific_error(
+                        self.node,
+                        (
+                            f"{e}\nHint: If you are using a memory pass based on dynamic shape bounds, "
+                            f"such as ConstraintBasedSymShapeEvalPass, this may be the cause of an "
+                            f"unbacked SymInt with its upper bound lazily set to 2^64-1 (uint64 max) "
+                            "during torch.export()."
+                        )
+                    )
+                )
 
         if spec.const:
             # Tensor with a blob we need to serialize. May not actually be constant at runtime

exir/passes/sym_shape_eval_pass.py

@@ -181,7 +181,7 @@ class HintBasedSymShapeEvalPass(PassBase):
 
     Warning: if you're using torch.export with constrain API, this method doesn't respect the input constraints.
 
-    Not inherit from ExportPass since we simply need a way to iterate thru
+    Not inherited from ExportPass since we simply need a way to iterate thru
     every node's output. PassBase is easier for that purpose.
     """
 
@@ -245,7 +245,7 @@ class ConstraintBasedSymShapeEvalPass(PassBase):
     formula. We should convert those symbolic formula to concrete value for
     static/upperbound tensors so we can properly do memory planning for them.
 
-    Not inherit from ExportPass since we simply need a way to iterate thru
+    Not inherited from ExportPass since we simply need a way to iterate through
     every node's output. PassBase is easier for that purpose.
     """
 

exir/tensor.py

@@ -21,6 +21,9 @@
 from executorch.exir.schema import ScalarType, TensorShapeDynamism
 from executorch.exir.sym_util import eval_shape
 
+class AddressSpaceOverflowException(Exception):
+    pass
+
 
 def num_bytes_from_shape_and_dtype(shape: torch.Size, dtype: torch.dtype) -> int:
     """
@@ -297,7 +300,7 @@ def make_allocation_info(mem_id: int, mem_offset: int) -> schema.AllocationDetai
     memory_offset_low = mem_offset & ((1 << 32) - 1)
     memory_offset_high = mem_offset >> 32
     if memory_offset_high >= 1 << 32:
-        raise ValueError(f"mem_offset {mem_offset} does not fit in 64 bits")
+        raise AddressSpaceOverflowException(f"mem_offset {mem_offset} does not fit in 64 bits")
 
     allocation_info = schema.AllocationDetails(
         memory_id=mem_id,