PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix missing change for compress_layer_names and add UTs (#1556)
Xin He6d17bbfbacba
diff --git a/auto_round/utils/common.py b/auto_round/utils/common.pyindex 96b6066e5..457a0e6a4 100644--- a/auto_round/utils/common.py+++ b/auto_round/utils/common.py@@ -554,7 +554,7 @@ def compress_layer_names(names: list) -> str:groups: dict = defaultdict(list)singles: list = []for name in names:- m = _re.match(r"^(.*?\.)?(\d+)(\..+)?$", name)+ m = _re.match(r"^(.*\.)?(\d+)(\..+)?$", name)if m:prefix = m.group(1) or ""num = int(m.group(2))diff --git a/test/test_cpu/utils/test_compress_layer_names.py b/test/test_cpu/utils/test_compress_layer_names.pynew file mode 100644index 000000000..bf1888f84--- /dev/null+++ b/test/test_cpu/utils/test_compress_layer_names.py@@ -0,0 +1,81 @@+import pytest++from auto_round.utils.common import compress_layer_names+++class TestCompressLayerNames:++ def test_simple_consecutive_layers(self):+ names = [f"model.layers.{i}.mlp.gate_proj" for i in range(4)]+ result = compress_layer_names(names)+ assert "model.layers.[0-3].mlp.gate_proj" in result++ def test_experts_same_suffix(self):+ names = [+ "mtp.layers.0.mlp.experts.248.up_proj",+ "mtp.layers.0.mlp.experts.87.up_proj",+ "mtp.layers.0.mlp.experts.23.up_proj",+ ]+ result = compress_layer_names(names)+ # Should group by expert number, not by layers index+ assert "mtp.layers.0.mlp.experts." in result+ assert ".up_proj" in result+ # Expert numbers should appear, not "layers.0"+ assert "23" in result and "87" in result and "248" in result++ def test_experts_multiple_suffixes(self):+ names = [+ "mtp.layers.0.mlp.experts.58.down_proj",+ "mtp.layers.0.mlp.experts.58.gate_proj",+ "mtp.layers.0.mlp.experts.37.up_proj",+ "mtp.layers.0.mlp.experts.37.down_proj",+ ]+ result = compress_layer_names(names)+ # down_proj and gate_proj should be separate groups+ assert "down_proj" in result+ assert "gate_proj" in result+ assert "up_proj" in result++ def test_experts_not_merged_across_layers(self):+ names = [+ "mtp.layers.0.mlp.experts.5.up_proj",+ "mtp.layers.1.mlp.experts.5.up_proj",+ ]+ result = compress_layer_names(names)+ # Different layer indices → different prefix → not merged into one group+ parts = [p.strip() for p in result.split(",")]+ assert len(parts) == 2++ def test_single_name(self):+ names = ["model.layers.3.mlp.gate_proj"]+ result = compress_layer_names(names)+ assert result == "model.layers.3.mlp.gate_proj"++ def test_non_consecutive_ranges(self):+ names = [f"model.layers.{i}.self_attn.q_proj" for i in [0, 1, 3, 4]]+ result = compress_layer_names(names)+ assert "[0-1,3-4]" in result++ def test_no_number_in_name(self):+ names = ["embed_tokens", "lm_head"]+ result = compress_layer_names(names)+ for name in names:+ assert name in result++ def test_empty_input(self):+ assert compress_layer_names([]) == ""++ def test_full_experts_example(self):+ """Reproduces the original bug report: expert index must be the grouped number."""+ names = [+ "mtp.layers.0.mlp.experts.248.up_proj",+ "mtp.layers.0.mlp.experts.87.up_proj",+ "mtp.layers.0.mlp.experts.58.down_proj",+ "mtp.layers.0.mlp.experts.37.up_proj",+ "mtp.layers.0.mlp.experts.37.down_proj",+ ]+ result = compress_layer_names(names)+ # All up_proj experts should be in one group+ assert "mtp.layers.0.mlp.experts.[" in result+ # The layer 0 index must NOT be the grouped number (i.e. no "[0]" grouping)+ assert "layers.[0]" not in result