Create convert_gguf_to_safetensor.py

tools/convert_gguf_to_safetensor.py

@@ -0,0 +1,98 @@
+import sys
+import gguf  # Ensure this is the gguf module from llama.cpp
+from gguf.constants import GGUFValueType
+from safetensors.numpy import save_file
+import numpy as np
+
+def reconstruct_field_value(field):
+    # Reconstruct the value from the field
+    types = field.types
+    if not types:
+        return None
+    gtype = types[0]
+    if gtype == GGUFValueType.ARRAY:
+        # Handle arrays
+        raw_itype = field.parts[0][0]  # np.uint32
+        array_length = field.parts[1][0]  # np.uint64
+        # Now read the elements
+        array_values = []
+        idx = 2
+        for _ in range(array_length):
+            element = field.parts[idx][0]
+            array_values.append(int(element))
+            idx += 1
+        return array_values
+    elif gtype in [
+        GGUFValueType.UINT8, GGUFValueType.INT8,
+        GGUFValueType.UINT16, GGUFValueType.INT16,
+        GGUFValueType.UINT32, GGUFValueType.INT32,
+        GGUFValueType.UINT64, GGUFValueType.INT64,
+        GGUFValueType.FLOAT32, GGUFValueType.FLOAT64,
+        GGUFValueType.BOOL,
+    ]:
+        # Scalar types
+        value = field.parts[0][0]
+        return value
+    elif gtype == GGUFValueType.STRING:
+        # String type
+        value = field.parts[1].tobytes().decode('utf-8')
+        return value
+    else:
+        # Other types can be added as needed
+        return None
+
+def bfloat16_to_float32(bf16_array):
+    """Converts an array of bfloat16 stored as uint16 to float32."""
+    # Create an empty float32 array
+    float32_array = np.empty(bf16_array.shape, dtype=np.float32)
+    # View the float32 array as uint32
+    uint32_view = float32_array.view(np.uint32)
+    # Assign the high 16 bits from the bf16 data and zero out the low 16 bits
+    uint32_view[:] = bf16_array.astype(np.uint32) << 16
+    return float32_array
+
+def main():
+    import numpy as np
+    print(f"NumPy version: {np.__version__}")
+
+    if len(sys.argv) != 3:
+        print("Usage: python convert_gguf_to_safetensor.py <input.gguf> <output.safetensors>")
+        sys.exit(1)
+
+    input_file = sys.argv[1]
+    output_file = sys.argv[2]
+
+    print(f"Reading GGUF file: {input_file}")
+    reader = gguf.GGUFReader(input_file)
+
+    tensors = {}
+    print(f"Number of tensors: {len(reader.tensors)}")
+
+    for tensor in reader.tensors:
+        name = tensor.name.decode('utf-8') if isinstance(tensor.name, bytes) else tensor.name
+        print(f"Reading tensor: {name}")
+        data = tensor.data
+
+        # Check for original shape metadata
+        orig_shape_key = f"comfy.gguf.orig_shape.{name}"
+        if orig_shape_key in reader.fields:
+            field = reader.get_field(orig_shape_key)
+            orig_shape = reconstruct_field_value(field)
+            print(f"Reshaping tensor '{name}' to original shape: {orig_shape}")
+            data = data.reshape(orig_shape)
+        else:
+            print(f"No original shape found for tensor '{name}', using current shape.")
+
+        # Convert data types if necessary
+        if data.dtype.name == 'bfloat16':
+            print(f"Converting tensor '{name}' from bfloat16 to float32")
+            data = bfloat16_to_float32(data)
+
+        tensors[name] = data
+
+    print(f"Saving to safetensors file: {output_file}")
+    save_file(tensors, output_file)
+    print("Conversion completed successfully.")
+
+if __name__ == "__main__":
+    main()