diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 57ca2035fe2..8ddd895cb7c 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -371,10 +371,13 @@ def add_tensor_info( 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 @@ def write_padding(self, fp: IO[bytes], n: int, align: int | None = None) -> None 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) diff --git a/gguf-py/gguf/scripts/gguf_editor_gui.py b/gguf-py/gguf/scripts/gguf_editor_gui.py index 05f4db0f8cd..293316afed4 100755 --- a/gguf-py/gguf/scripts/gguf_editor_gui.py +++ b/gguf-py/gguf/scripts/gguf_editor_gui.py @@ -1552,7 +1552,7 @@ def save_file(self): # 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)) diff --git a/gguf-py/gguf/scripts/gguf_new_metadata.py b/gguf-py/gguf/scripts/gguf_new_metadata.py index 2fa5800cf74..c67436bad4a 100755 --- a/gguf-py/gguf/scripts/gguf_new_metadata.py +++ b/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()