PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix device bug (#873)
Xin Heeb5cdcd8244e
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.pyindex 97f7e8682..e2aabc706 100644--- a/auto_round/compressors/base.py+++ b/auto_round/compressors/base.py@@ -1420,7 +1420,7 @@ class BaseCompressor(object):m.zp = Noneelse:try:- m.to(self.device)+ m = m.to(self.device)m = WrapperLinear(m,enable_minmax_tuning=False,@@ -1858,6 +1858,7 @@ class BaseCompressor(object):from auto_round.data_type import QUANT_FUNC_WITH_DTYPElayer = get_module(self.model, layer_name)+ layer = layer.to(self.device)if _is_fp8_model(self.model):new_layer = convert_fp8_layer_to_linear(layer, self.amp_dtype).to(self.device)set_module(self.model, layer_name, new_layer)