Address comment

Author: Danqing Wang (MPK)

data/bin/__init__.py

@@ -0,0 +1,44 @@
+# This file should be written to the wheel package as
+# `executorch/data/bin/__init__.py`.
+#
+# Setuptools will expect to be able to say something like `from
+# executorch.data.bin import mybin; mybin()` for each entry listed in the
+# [project.scripts] section of pyproject.toml. This file makes the `mybin()`
+# function execute the binary at `executorch/data/bin/mybin` and exit with that
+# binary's exit status.
+
+import subprocess
+import os
+import sys
+import types
+
+# This file should live in the target `bin` directory.
+_bin_dir = os.path.join(os.path.dirname(__file__))
+
+def _find_executable_files_under(dir):
+    """Lists all executable files in the given directory."""
+    bin_names = []
+    for filename in os.listdir(dir):
+        filepath = os.path.join(dir, filename)
+        if os.path.isfile(filepath) and os.access(filepath, os.X_OK):
+            # Remove .exe suffix on windows.
+            filename_without_ext = os.path.splitext(filename)[0]
+            bin_names.append(filename_without_ext)
+    return bin_names
+
+# The list of binaries to create wrapper functions for.
+_bin_names = _find_executable_files_under(_bin_dir)
+
+# We'll define functions named after each binary. Make them importable.
+__all__ = _bin_names
+
+def _run(name):
+    """Runs the named binary, which should live under _bin_dir.
+
+    Exits the current process with the return code of the subprocess.
+    """
+    raise SystemExit(subprocess.call([os.path.join(_bin_dir, name)] + sys.argv[1:], close_fds=False))
+
+# Define a function named after each of the binaries.
+for bin_name in _bin_names:
+    exec(f"def {bin_name}(): _run('{bin_name}')")

data/bin/flatc

@@ -0,0 +1,97 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+
+//
+// See README.md before modifying this file.
+//
+
+include "scalar_type.fbs";
+
+namespace bundled_program_flatbuffer;
+
+// Identifier of a valid bundled program schema.
+file_identifier "BP08";
+// Extension of written files.
+file_extension "bpte";
+
+// Reason for basic struct: union value type can only be table/struct/string
+table Int {
+  int_val:long;
+}
+
+table Bool {
+  bool_val:bool;
+}
+
+table Double {
+  double_val:double;
+}
+
+// All information we need to bundle for a tensor EValue input.
+table Tensor {
+  // The scalar type of Tensor
+  scalar_type: executorch_flatbuffer.ScalarType;
+  // The target sizes of the tensor.
+  sizes: [int];
+  // The contents of the corresponding input tensor.
+  data: [ubyte] (force_align: 16);
+  dim_order:[ubyte];
+}
+
+union ValueUnion {
+  Tensor,
+  Int,
+  Bool,
+  Double,
+}
+
+// Abstraction for BundledMethodTestCase values
+table Value {
+  val: ValueUnion;
+}
+
+// A single test for a method. The provided inputs should produce the
+// expected outputs.
+table BundledMethodTestCase {
+  // The inputs to provide to the method. The number and types of inputs must
+  // match the schema of the method under test.
+  inputs: [Value];
+
+  // The expected outputs generated while running the model in eager mode using
+  // the inputs provided. Its length should be equal to the length of program
+  // outputs.
+  expected_outputs: [Value];
+}
+
+// Collection of test cases for a program method.
+table BundledMethodTestSuite {
+  // The name of the method to test; e.g., "forward" for the forward() method
+  // of an nn.Module. This name match a method defined by the ExecuTorch
+  // program.
+  method_name: string;
+
+  // Individual test cases for the method.
+  test_cases: [BundledMethodTestCase];
+}
+
+
+// Executorch program bunlded with data for verification.
+table BundledProgram {
+  // Schema version.
+  version:uint;
+
+  // Test sets to run against the program.
+  // Each BundledMethodTestSuite should be used for the method of program sharing same name.
+  method_test_suites: [BundledMethodTestSuite];
+
+  // The binary data of a serialized Executorch program.
+  // The following `force_align` may sliently override any larger force_align
+  // used in the program. Therefore, to keep the data (including constant
+  // tensor, delegate data, etc, see schema.fbs for more info) in the
+  // executorch program keeps the same alignment as original no matter how
+  // the program schema changes, we need to make the force_align here the max
+  // one around all kinds of force_align in the current and future program
+  // schema, so we use the 32 as the force_align here.
+  program: [ubyte] (force_align: 32);
+}
+
+root_type BundledProgram;

devtools/bundled_program/serialize/bundled_program_schema.fbs

@@ -0,0 +1,43 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+
+//
+// See README.md before modifying this file.
+//
+
+namespace executorch_flatbuffer;
+
+// The scalar data type.
+// Must match executorch/runtime/core/portable_type/tensor_impl.h
+enum ScalarType : byte {
+  BYTE = 0,
+  CHAR = 1,
+  SHORT = 2,
+  INT = 3,
+  LONG = 4,
+  HALF = 5,
+  FLOAT = 6,
+  DOUBLE = 7,
+  BOOL = 11,
+  QINT8 = 12,
+  QUINT8 = 13,
+  QINT32 = 14,
+  QUINT4X2 = 16,
+  QUINT2X4 = 17,
+  BITS16 = 22,
+  FLOAT8E5M2 = 23,
+  FLOAT8E4M3FN = 24,
+  FLOAT8E5M2FNUZ = 25,
+  FLOAT8E4M3FNUZ = 26,
+  UINT16 = 27,
+  UINT32 = 28,
+  UINT64 = 29,
+  // Types currently not implemented.
+  // COMPLEXHALF = 8,
+  // COMPLEXFLOAT = 9,
+  // COMPLEXDOUBLE = 10,
+  // BFLOAT16 = 15,
+  // BITS1x8 = 18,
+  // BITS2x4 = 19,
+  // BITS4x2 = 20,
+  // BITS8 = 21,
+}

devtools/bundled_program/serialize/scalar_type.fbs

@@ -94,7 +94,7 @@
     "static_llama",
     "qwen2_5",
     "phi-4-mini",
-    "smollm",
+    "smolllm2",
 ]
 TORCHTUNE_DEFINED_MODELS = ["llama3_2_vision"]
 

examples/models/llama/export_llama_lib.py

@@ -42,11 +42,6 @@ def smollm_tune_to_meta(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.
         new_key = get_mapped_key(key, inverted_mapping_dict)
         converted_state_dict[new_key] = value
 
-    # Input and output embeddings are tied.
-    converted_state_dict["output.weight"] = converted_state_dict[
-        "tok_embeddings.weight"
-    ]
-
     return converted_state_dict
 
 
@@ -68,7 +63,7 @@ def main():
         checkpoint_dir=args.input_dir,
         checkpoint_files=["model.safetensors"],
         output_dir=".",
-        model_type="MISTRAL",
+        model_type="LLAMA",
     )
 
     print("Loading checkpoint...")