PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
remove result to cpu (#1115)
Yi Liu5346cb73b148
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.pyindex 0c5a7f3fb..fab7fd346 100644--- a/auto_round/compressors/base.py+++ b/auto_round/compressors/base.py@@ -1590,7 +1590,7 @@ class BaseCompressor(object):if hasattr(m, "imatrix"):m.imatrix /= m.imatrix_cntif hasattr(m, "tmp_name") and m.tmp_name in all_to_quantized_module_names:- self._quantize_layer_via_rtn(m.tmp_name, to_cpu=False)+ self._quantize_layer_via_rtn(m.tmp_name, to_cpu=self.low_gpu_mem_usage)all_to_quantized_module_names.remove(m.tmp_name)if not self.immediate_saving:mv_module_from_gpu(block)