PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Skip zero-point round-trip with compressed-tensors 0.20.0 (#2544)
Wang, Chang76a114e5cf68
diff --git a/test/unit/test_cpu/export/test_llmc_format.py b/test/unit/test_cpu/export/test_llmc_format.pyindex 98d9dcc78..325c4c0d9 100644--- a/test/unit/test_cpu/export/test_llmc_format.py+++ b/test/unit/test_cpu/export/test_llmc_format.py@@ -1,7 +1,7 @@import jsonimport osimport shutil-from test.helpers import forbid_threaded_packing, get_model_path, opt_name_or_path+from test.helpers import check_version, forbid_threaded_packing, get_model_path, opt_name_or_pathimport pytestimport torch@@ -99,6 +99,10 @@ class TestLLMCZpInt8:else:raise AssertionError("out-of-range zp must raise")+ @pytest.mark.skipif(+ check_version("compressed-tensors==0.20.0"),+ reason="compressed-tensors 0.20.0 has a zero-point unpacking bug",+ )def test_asym_zp_signed_convention_round_trip(self):"""AutoRound zp is unsigned (q in [0, 2^b-1], W = (q - zp) * s);compressed-tensors re-quantizes with the SIGNED range