Refactor serialize.py

extension/flat_tensor/serialize/flat_tensor.fbs

@@ -35,8 +35,8 @@ table TensorMetadata {
   // To retrieve a given tensor:
   // 1. segment_base_offset: from the file header.
   // 2. segment_offset: segments[segment_index].offset
-  // 3. tensor_offset: segments[segment_offset].tensor_metadata[j].offset
-  //    Find the relevant index j by matching on tensor fqn.
+  // 3. tensor_offset: the offset within the segment. If there is only one item
+  // in the segment, offset=0.
   offset: uint64;
 }
 
@@ -55,6 +55,15 @@ table DataSegment {
   size: uint64;
 }
 
+// Attributes a name to data referenced by FlatTensor.segments.
+table NamedData {
+  // The unique id of the data blob.
+  key: string;
+
+  // Index of the segment in FlatTensor.segments.
+  segment_index: uint32;
+}
+
 // FlatTensor is a flatbuffer-based format for storing and loading tensors.
 table FlatTensor {
   // Schema version.
@@ -70,6 +79,10 @@ table FlatTensor {
   // List of data segments that follow the FlatTensor data in this file, sorted by
   // offset. Elements in this schema can refer to these segments by index.
   segments: [DataSegment];
+
+  // List of blobs keyed by a unique name. Note that multiple 'NamedData'
+  // entries could point to the same segment index.
+  named_data: [NamedData];
 }
 
 root_type FlatTensor;

extension/flat_tensor/serialize/flat_tensor_schema.py

@@ -31,9 +31,16 @@ class DataSegment:
     size: int
 
 
+@dataclass
+class NamedData:
+    key: str
+    segment_index: int
+
+
 @dataclass
 class FlatTensor:
     version: int
     tensor_alignment: int
     tensors: List[TensorMetadata]
     segments: List[DataSegment]
+    named_data: List[NamedData]

extension/flat_tensor/serialize/serialize.py

@@ -10,29 +10,33 @@
 import os
 import tempfile
 from dataclasses import dataclass
-from typing import ClassVar, Dict, List, Literal, Optional
+from typing import ClassVar, Dict, List, Literal, Optional, Sequence
 
 import pkg_resources
 from executorch.exir._serialize._cord import Cord
 from executorch.exir._serialize._dataclass import _DataclassEncoder, _json_to_dataclass
 
 from executorch.exir._serialize._flatbuffer import _flatc_compile, _flatc_decompile
 from executorch.exir._serialize._program import _insert_flatbuffer_header
-from executorch.exir._serialize.data_serializer import DataPayload, DataSerializer
+from executorch.exir._serialize.data_serializer import (
+    DataPayload,
+    DataSerializer,
+    TensorEntry,
+)
 
 from executorch.exir._serialize.padding import aligned_size, pad_to, padding_required
 
-# Byte order of numbers written to flat tensor headers. Always little-endian
-# regardless of the host system, since all commonly-used modern CPUs are little
-# endian.
-_HEADER_BYTEORDER: Literal["little"] = "little"
-
 from executorch.extension.flat_tensor.serialize.flat_tensor_schema import (
     DataSegment,
     FlatTensor,
     TensorMetadata,
 )
 
+# Byte order of numbers written to flat tensor headers. Always little-endian
+# regardless of the host system, since all commonly-used modern CPUs are little
+# endian.
+_HEADER_BYTEORDER: Literal["little"] = "little"
+
 
 def _serialize_to_flatbuffer(flat_tensor: FlatTensor) -> Cord:
     """Serializes a FlatTensor to a flatbuffer and returns the serialized data."""
@@ -209,6 +213,62 @@ def _get_extended_header(flat_tensor_data: bytes) -> Optional[FlatTensorHeader]:
     return None
 
 
+def _extract_tensors(
+    fqn_to_tensor: Dict[str, TensorEntry],
+    buffers: Sequence[bytes],
+    segments: List[Cord],
+    tensor_alignment: int,
+) -> List[TensorMetadata]:
+    """Places tensors into a single segment, aligned to tensor_alignment within
+        the segment.
+
+    Args:
+        fqn_to_tensor: A map from fully qualified names to tensor entries.
+        buffers: A sequence of tensor buffers.
+        segments: A list of segments to append the tensor data to. Modified in-place.
+        tensor_alignment: The alignment of the tensor data.
+
+    Returns:
+        A list of TensorMetadata, which describes the tensors in the segment.
+    """
+    tensor_data: Cord = Cord()
+    tensors: List[TensorMetadata] = []
+    # {idx, offset}
+    saved_offsets: Dict[int, int] = {}
+    for fqn, tensor_entry in fqn_to_tensor.items():
+        assert tensor_entry.layout is not None
+        # Check index into the tensor buffers is valid.
+        assert tensor_entry.buffer_index < len(
+            buffers
+        ), f"Invalid index {tensor_entry.buffer_index} is greater than tensor buffer size {len(buffers)}."
+
+        # Check if the tensor has already been appended to the flat_tensor_data.
+        offset = saved_offsets.get(tensor_entry.buffer_index, -1)
+        if offset == -1:
+            if len(tensor_data) > 0:
+                # Add padding to round off the previous tensor offset.
+                pad_length = padding_required(len(tensor_data), tensor_alignment)
+                tensor_data.append(b"\x00" * pad_length)
+            # Add to saved offsets.
+            offset = len(tensor_data)
+            saved_offsets[tensor_entry.buffer_index] = offset
+            # Append to flat_tensor_data at the offset.
+            tensor_data.append(buffers[tensor_entry.buffer_index])
+
+        tensors.append(
+            TensorMetadata(
+                fully_qualified_name=fqn,
+                scalar_type=tensor_entry.layout.scalar_type,
+                sizes=tensor_entry.layout.sizes,
+                dim_order=tensor_entry.layout.dim_order,
+                segment_index=len(segments),
+                offset=offset,
+            )
+        )
+    segments.append(tensor_data)
+    return tensors
+
+
 class FlatTensorSerializer(DataSerializer):
     """A concrete implementation of the DataSerializer interface that
     serializes and deserializes data to/from the FlatTensor format.
@@ -227,61 +287,46 @@ def serialize(
         self,
         data: DataPayload,
     ) -> Cord:
-        """Serializes a list of tensor metadata and tensors into a blob."""
-
-        flat_tensor_metadata: List[TensorMetadata] = []
-        flat_tensor_data: Cord = Cord()
-
-        # {idx, offset}
-        saved_offsets: Dict[int, int] = {}
-
-        for fqn, tensor_entry in data.fqn_to_tensor.items():
-            assert tensor_entry.layout is not None
-            # Check index into the tensor buffers is valid.
-            assert tensor_entry.buffer_index < len(
-                data.buffers
-            ), f"Invalid index {tensor_entry.buffer_index} is greater than tensor buffer size {len(data.buffers)}."
-
-            # Check if the tensor has already been appended to the flat_tensor_data.
-            offset = saved_offsets.get(tensor_entry.buffer_index, -1)
-            if offset == -1:
-                if len(flat_tensor_data) > 0:
-                    # Add padding to round off the previous tensor offset.
-                    pad_length = padding_required(
-                        len(flat_tensor_data), self.config.tensor_alignment
-                    )
-                    flat_tensor_data.append(b"\x00" * pad_length)
-                # Add to saved offsets.
-                offset = len(flat_tensor_data)
-                saved_offsets[tensor_entry.buffer_index] = offset
-                # Append to flat_tensor_data at the offset.
-                flat_tensor_data.append(data.buffers[tensor_entry.buffer_index])
-
-            flat_tensor_metadata.append(
-                TensorMetadata(
-                    fully_qualified_name=fqn,
-                    scalar_type=tensor_entry.layout.scalar_type,
-                    sizes=tensor_entry.layout.sizes,
-                    dim_order=tensor_entry.layout.dim_order,
-                    segment_index=0,
-                    offset=offset,
+        """Serializes a list of tensors and named data into a blob."""
+
+        segments: List[Cord] = []
+        tensors = _extract_tensors(
+            data.fqn_to_tensor,
+            data.buffers,
+            segments,
+            self.config.tensor_alignment,
+        )
+
+        data_segments: List[DataSegment] = []
+        segment_data = Cord()
+        for segment in segments:
+            prev_end = (
+                (data_segments[-1].offset + data_segments[-1].size)
+                if data_segments
+                else 0
+            )
+            data_segments.append(
+                DataSegment(
+                    offset=aligned_size(prev_end, self.config.segment_alignment),
+                    size=len(segment),
                 )
             )
-
-        # Pad flat_tensor_data to segment alignment.
-        segment_pad_length = padding_required(
-            len(flat_tensor_data), self.config.segment_alignment
-        )
-        if segment_pad_length > 0:
-            flat_tensor_data.append(b"\x00" * segment_pad_length)
+            # Pad segment_data to segment alignment.
+            segment_pad_length = padding_required(
+                len(segment_data), self.config.segment_alignment
+            )
+            if segment_pad_length > 0:
+                segment_data.append(b"\x00" * segment_pad_length)
+            segment_data.append(segment)
 
         # Create FlatTensor, which describes of the contents of the file and
         # points to all the data segments. It will be serialized to flatbuffer.
         flat_tensor = FlatTensor(
             version=0,  # Keep in sync with c++ version number in serialize.h
             tensor_alignment=self.config.tensor_alignment,
-            tensors=flat_tensor_metadata,
-            segments=[DataSegment(offset=0, size=len(flat_tensor_data))],
+            tensors=tensors,
+            segments=data_segments,
+            named_data=[],
         )
 
         flatbuffer_payload = _serialize_to_flatbuffer(flat_tensor)
@@ -306,7 +351,7 @@ def serialize(
             flatbuffer_offset=padded_header_length,
             flatbuffer_size=len(flatbuffer_payload),
             segment_base_offset=segment_base_offset,
-            segment_data_size=len(flat_tensor_data),
+            segment_data_size=len(segment_data),
         ).to_bytes()
 
         # Pad header and payload to segment alignment.
@@ -326,15 +371,15 @@ def serialize(
         assert eh.flatbuffer_size == original_flatbuffer_payload_size
         assert eh.segment_base_offset == segment_base_offset
         assert eh.flatbuffer_offset == padded_header_length
-        assert eh.segment_data_size == len(flat_tensor_data)
+        assert eh.segment_data_size == len(segment_data)
 
         del header_data
         del flatbuffer_payload
 
         # Place everything into one segment.
         payload = Cord()
         payload.append(injected_flatbuffer_data)
-        payload.append(flat_tensor_data)
+        payload.append(segment_data)
 
         return payload