Add FP8 placeholder support to ExecuTorch serialization (pytorch#19043)

Change-Id: Ibb7ef4167ab96426133fce64e34366c365cd12ad
Signed-off-by: Yufeng Shi <yufeng.shi@arm.com>

Author: YufengShi-dudu

exir/print_program.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -60,6 +61,8 @@ def _scalar_type_str(scalar_type: ScalarType) -> str:
         ScalarType.QUINT8: "qui8",
         ScalarType.QINT32: "qi32",
         ScalarType.BFLOAT16: "bf16",
+        ScalarType.FLOAT8E5M2: "f8e5m2",
+        ScalarType.FLOAT8E4M3FN: "f8e4m3fn",
         ScalarType.QUINT4x2: "qui4x2",
         ScalarType.QUINT2x4: "qui2x4",
     }

exir/serde/export_serialize.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -149,6 +150,8 @@ def _reverse_map(d: Dict[Any, Enum]):
     torch.complex128: ScalarType.COMPLEXDOUBLE,
     torch.bool: ScalarType.BOOL,
     torch.bfloat16: ScalarType.BFLOAT16,
+    torch.float8_e5m2: ScalarType.FLOAT8E5M2,
+    torch.float8_e4m3fn: ScalarType.FLOAT8E4M3FN,
     torch.uint16: ScalarType.UINT16
 }
 

exir/serde/schema.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -16,7 +17,7 @@
 from executorch.exir.serde.union import _Union
 
 # NOTE: Please update this value if any modifications are made to the schema
-SCHEMA_VERSION = (5, 3)
+SCHEMA_VERSION = (5, 4)
 TREESPEC_VERSION = 1
 
 
@@ -36,6 +37,8 @@ class ScalarType(IntEnum):
     BOOL = 12
     BFLOAT16 = 13
     UINT16 = 14
+    FLOAT8E5M2 = 15
+    FLOAT8E4M3FN = 16
 
 class Layout(IntEnum):
     Unknown = 0

exir/tensor.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -292,6 +293,8 @@ def memory_format_enum(memory_format: torch.memory_format) -> int:
     torch.qint32: ScalarType.QINT32,
     torch.bfloat16: ScalarType.BFLOAT16,
     torch.quint4x2: ScalarType.QUINT4x2,
+    torch.float8_e5m2: ScalarType.FLOAT8E5M2,
+    torch.float8_e4m3fn: ScalarType.FLOAT8E4M3FN,
     torch.uint16: ScalarType.UINT16,
     torch.uint32: ScalarType.UINT32,
 }

exir/tests/test_tensor.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -90,6 +91,14 @@ def test_normal_tensor_conversion(self) -> None:
         # whereas strides for torch.memory_format = torch.channels_last is
         # (3*4*5, 1, 5*3, 3))
 
+    def test_fp8_tensor_conversion(self) -> None:
+        for dtype in (torch.float8_e5m2, torch.float8_e4m3fn):
+            normal_tensor = torch.randn(2, 2, 3, dtype=torch.float32).to(dtype)
+            flatbuffer_tensor = make_tensor_value(
+                1, 0, TensorSpec.from_tensor(normal_tensor)
+            )
+            self.compare_tensors(normal_tensor, flatbuffer_tensor)
+
     def test_allocation_info_succeeds(self) -> None:
         test_cases = (
             (