Merge pull request openvinotoolkit#102 from eaidova/ea/add_all_layers…

…_option

add --all_layers option

llm_bench/python/convert.py

@@ -1151,8 +1151,6 @@ def main():
     parser.add_argument('-o', '--output_dir', required=True, help='output directory for saving model')
     parser.add_argument('--save_orig', action='store_true', help='save pytorch model on disk')
     parser.add_argument('-p', '--precision', choices=['FP32', 'FP16'], default='FP32', help='base conversion precision')
-    parser.add_argument('--bettertransformer', action='store_true',
-                        help='Apply bettertransformer to enable ScaledDotProductAttention operation for a part of the models')
 
     compression_group = parser.add_argument_group('Weights compression parameters')
     compression_group.add_argument(
@@ -1187,6 +1185,11 @@ def main():
         default=None,
         type=int,
     )
+    compression_group.add_argument(
+        "--all_layers",
+        action="store_true",
+        help="Compress all layers including embeddings and prediction head"
+    )
     add_stateful_model_arguments(parser)
 
     args = parser.parse_args()

llm_bench/python/utils/conversion_utils/helpers.py

@@ -97,6 +97,8 @@ def compress_ov_model_weights_helper(ov_model, tok, config, out_path, compress_w
             compression_args["ratio"] = args.ratio
         if args.group_size is not None:
             compression_args["group_size"] = args.group_size
+    if args.all_layers:
+        compression_args["all_layers"] = True
     log.info("Compression options:")
     log.info(compression_args)
     compressed_ov_model = compress_weights(ov_model, **compression_args)