Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

Merge branch 'main' of https://github.com/intel/auto-round into main

Zhang, Weiwei141b328c21df9
diff --cc examples/language-modeling/main.py
index 7892922fc,141fe9315..b4b224b12
--- a/examples/language-modeling/main.py
+++ b/examples/language-modeling/main.py
@@@ -83,14 -84,14 +84,14 @@@ if __name__ == '__main__'
parser.add_argument("--enable_minmax_tuning", action='store_true',
help="whether enable weight minmax tuning")
-
+
parser.add_argument("--deployment_device", default='fake', type=str,
- help="targeted inference acceleration platform,The options are 'fake', 'cpu' and 'gpu',"
+ help="targeted inference acceleration platform,The options are 'fake', 'cpu' and 'gpu'."
"default to 'fake', indicating that it only performs fake quantization and won't be exported to any device.")
-
+
parser.add_argument("--scale_dtype", default='fp32',
help="which scale data type to use for quantization, 'fp16', 'fp32' or 'bf16'.")
-
+
parser.add_argument("--tasks",
default=['wikitext2', 'ptb-new', 'c4-new', 'lambada_openai', 'hellaswag', 'winogrande', 'piqa',
"mmlu", "wikitext", "truthfulqa_mc1", "truthfulqa_mc2", "openbookqa", "boolq", "rte",
@@@ -109,8 -104,9 +104,10 @@@
parser.add_argument("--eval_legacy", action='store_true',
help="Whether to evaluate with a old lm_eval version(e.g. 0.3.0).")
+ parser.add_argument("--disable_lmeval", action='store_true',
+ help="Whether to do lmeval evaluation.")
+
args = parser.parse_args()
set_seed(args.seed)