PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Fix group_size issue (#1638)
Wang, Mengni7b0ff4159496
diff --git a/auto_round/__main__.py b/auto_round/__main__.pyindex 057641bdf..9da465810 100644--- a/auto_round/__main__.py+++ b/auto_round/__main__.py@@ -299,7 +299,7 @@ class BasicArgumentParser(argparse.ArgumentParser):scheme.add_argument("--group_size",default=None,- type=lambda s: int(s) if s.isdigit() else tuple([int(x.strip()) for x in s.split(",")]),+ type=lambda s: int(s) if s.lstrip("-").isdigit() else tuple([int(x.strip()) for x in s.split(",")]),help="Group size for weight quantization.",)scheme.add_argument("--asym", action="store_true", help="Use asymmetric quantization instead of symmetric.")diff --git a/test/test_cpu/quantization/test_block_fp.py b/test/test_cpu/quantization/test_block_fp.pyindex 19625911e..c703b6232 100644--- a/test/test_cpu/quantization/test_block_fp.py+++ b/test/test_cpu/quantization/test_block_fp.py@@ -1,4 +1,5 @@import shutil+import subprocessfrom math import ceilimport pytest@@ -91,3 +92,46 @@ class TestAutoRoundBlockFP:for j in range(N):scale_ref[i, j] = data[i * 128 : (i + 1) * 128, j * 128 : (j + 1) * 128].abs().max() / max_valassert (scale == scale_ref).all()++ def test_group_size_handler(self, tiny_qwen_model_path):+ scheme = {+ "data_type": "int",+ "bits": 4,+ "group_size": -1,+ "act_data_type": "int",+ "act_bits": 4,+ "act_group_size": -1,+ }+ autoround = AutoRound(+ tiny_qwen_model_path,+ scheme=scheme,+ iters=2,+ seqlen=2,+ )++ cmd = [+ "python3",+ "-m",+ "auto_round",+ "--model_name",+ tiny_qwen_model_path,+ "--bits",+ "4",+ "--group_size",+ "-1",+ "--act_bits",+ "4",+ "--act_group_size",+ "-1",+ "--format",+ "fake",+ "--iters",+ "0",+ ]+ subprocess.run(+ cmd,+ check=True,+ text=True,+ stdout=subprocess.PIPE,+ stderr=subprocess.PIPE,+ )