PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix rtn bug (#966)
Wang, Mengnic4ef9a827bbb
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.pyindex db181c035..fe0939c1f 100644--- a/auto_round/compressors/base.py+++ b/auto_round/compressors/base.py@@ -1312,10 +1312,11 @@ class BaseCompressor(object):self.model.to("cpu")enable_imatrix = False- if has_gguf_k and not self.disable_opt_rtn:- enable_imatrix = True- if self.data_type == "int" and self.sym:- enable_imatrix = True+ if not self.disable_opt_rtn:+ if has_gguf_k:+ enable_imatrix = True+ elif self.data_type == "int" and self.sym:+ enable_imatrix = Trueif enable_imatrix:self._quant_rtn_with_imatrix(all_to_quantized_module_names)