gguf-py : fix passing non-native endian tensors (editor-gui and new-metadata) (#17553)

gguf_new_metadata.py reads data from reader.
Reader doesn't byteswap tensors to native endianness.
But writer does expect tensors in native endianness to convert them
into requested endianness.

There are two ways to fix this: update reader and do conversion to native endianness and back,
or skip converting endianness in writer in this particular USE-case.

gguf_editor_gui.py doesn't allow editing or viewing tensor data.
Let's go with skipping excessive byteswapping.

If eventually capability to view or edit tensor data is added,
tensor data should be instead byteswapped when reading it.

gguf-py/gguf/gguf_writer.py

@@ -371,10 +371,13 @@ class GGUFWriter:
 
     def add_tensor(
         self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
-        raw_dtype: GGMLQuantizationType | None = None,
+        raw_dtype: GGMLQuantizationType | None = None, tensor_endianess: GGUFEndian | None = None
     ) -> None:
-        if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \
-                (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'):
+        # if tensor endianness is not passed, assume it's native to system
+        if tensor_endianess is None:
+            tensor_endianess = GGUFEndian.BIG if sys.byteorder == 'big' else GGUFEndian.LITTLE
+
+        if tensor_endianess != self.endianess:
             # Don't byteswap inplace since lazy copies cannot handle it
             tensor = tensor.byteswap(inplace=False)
         if self.use_temp_file and self.temp_file is None:
@@ -397,13 +400,16 @@ class GGUFWriter:
         if pad != 0:
             fp.write(bytes([0] * pad))
 
-    def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any], tensor_endianess: GGUFEndian | None = None) -> None:
         if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS:
             raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}')
         assert self.fout is not None
 
-        if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \
-                (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'):
+        # if tensor endianness is not passed, assume it's native to system
+        if tensor_endianess is None:
+            tensor_endianess = GGUFEndian.BIG if sys.byteorder == 'big' else GGUFEndian.LITTLE
+
+        if tensor_endianess != self.endianess:
             # Don't byteswap inplace since lazy copies cannot handle it
             tensor = tensor.byteswap(inplace=False)
 

gguf-py/gguf/scripts/gguf_editor_gui.py

@@ -1552,7 +1552,7 @@ class GGUFEditorWindow(QMainWindow):
 
             # Add tensors (including data)
             for tensor in self.reader.tensors:
-                writer.add_tensor(tensor.name, tensor.data, raw_shape=tensor.data.shape, raw_dtype=tensor.tensor_type)
+                writer.add_tensor(tensor.name, tensor.data, raw_shape=tensor.data.shape, raw_dtype=tensor.tensor_type, tensor_endianess=self.reader.endianess)
 
             # Write header and metadata
             writer.open_output_file(Path(file_path))

gguf-py/gguf/scripts/gguf_new_metadata.py

@@ -94,7 +94,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
     writer.write_ti_data_to_file()
 
     for tensor in reader.tensors:
-        writer.write_tensor_data(tensor.data)
+        writer.write_tensor_data(tensor.data, tensor_endianess=reader.endianess)
         bar.update(tensor.n_bytes)
 
     writer.close()