convert : minor fixes for numpy 2.x (ggml-org#23571)

Author: CISC

examples/convert_legacy_llama.py

@@ -1308,7 +1308,8 @@ def do_dump_model(model_plus: ModelPlus) -> None:
 
 def main(args_in: list[str] | None = None) -> None:
     output_choices = ["f32", "f16"]
-    if np.uint32(1) == np.uint32(1).newbyteorder("<"):
+    dummy_val = np.uint32(1)
+    if dummy_val == dummy_val.view(dummy_val.dtype.newbyteorder("<")):
         # We currently only support Q8_0 output on little endian systems.
         output_choices.append("q8_0")
     parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")

gguf-py/gguf/quants.py

@@ -28,6 +28,7 @@ def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizati
 # This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
 def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
     rows = arr.reshape((-1, arr.shape[-1]))
+    assert len(rows.shape)
     osize = 1
     for dim in oshape:
         osize *= dim