Make GGUFWriter accept tensors in native endianness instead of little-endian

With this change if no byteswapping is actually needed, 2 excessive byteswaps can be omitted on s390x

Author: AlekseiNikiforovIBM

convert_hf_to_gguf.py

@@ -615,12 +615,6 @@ def prepare_tensors(self):
                 # reverse shape to make it similar to the internal ggml dimension order
                 shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}"
 
-                if sys.byteorder == 'big':
-                    # Switch data back to little-endian.
-                    # gguf_writer.add_tensor later switches it back to big endian if needed.
-                    # Don't byteswap inplace since it cannot handle lazy copies
-                    data = data.byteswap(inplace=False)
-
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
 

gguf-py/gguf/gguf_writer.py

@@ -4,6 +4,7 @@
 import os
 import shutil
 import struct
+import sys
 import tempfile
 from dataclasses import dataclass
 from enum import Enum, auto
@@ -372,7 +373,8 @@ def add_tensor(
         self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
         raw_dtype: GGMLQuantizationType | None = None,
     ) -> None:
-        if self.endianess == GGUFEndian.BIG:
+        if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \
+                (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'):
             # 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:
@@ -400,7 +402,8 @@ def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
             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:
+        if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \
+                (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'):
             # Don't byteswap inplace since lazy copies cannot handle it
             tensor = tensor.byteswap(inplace=False)