PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
rm fp16 export in autoround format (#525)
Wenhua Cheng5d8226661386
diff --git a/auto_round/export/export_to_autoround/export.py b/auto_round/export/export_to_autoround/export.pyindex 6f8401a7d..beffe62a4 100644--- a/auto_round/export/export_to_autoround/export.py+++ b/auto_round/export/export_to_autoround/export.py@@ -324,7 +324,7 @@ def save_quantized_as_autoround(output_dir, inplace=True, backend="auto_round:exif quantization_config.get("act_bits", 16) <= 8:dtype = torch.bfloat16else:- dtype = torch.float16 ##force dtype to fp16+ dtype = Nonesave(model, output_dir, safe_serialization=safe_serialization, dtype=dtype)return modeldiff --git a/auto_round/inference/backend.py b/auto_round/inference/backend.pyindex fe26dd3f9..ed2d3fac1 100644--- a/auto_round/inference/backend.py+++ b/auto_round/inference/backend.py@@ -180,7 +180,7 @@ BackendInfos['auto_awq:gemm'] = BackendInfo(device=["cuda"], sym=[True, False],bits=[4], group_size=None,priority=4,dtype=["float16"],- alias=["auto_awq:gemm", "awq",+ alias=["auto_awq:gemm", "awq","awq:gemm","auto_awq"],requirements=["autoawq"])