PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Fix special-model predefined ignore layer filtering (#1863)
Liang Lv930b4b905da0
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.pyindex 0f6eaadf9..a8a7bc5ab 100644--- a/auto_round/compressors/base.py+++ b/auto_round/compressors/base.py@@ -560,10 +560,12 @@ class BaseCompressor(object):predefined_ignore_layers = get_predefined_ignore_layers(self.model_context.model)if predefined_ignore_layers and self.quant_block_list:block_prefixes = [block for group in self.quant_block_list for block in group]+ # Only filter layers that are full paths clearly inside a block.predefined_ignore_layers = [namefor name in predefined_ignore_layersif any(name.startswith(prefix) for prefix in block_prefixes)+ or not any(prefix.startswith(name.split(".")[0]) for prefix in block_prefixes)]predefined_ignore_layers = compress_layer_names(predefined_ignore_layers)if predefined_ignore_layers:diff --git a/test/test_cpu/utils/test_utils.py b/test/test_cpu/utils/test_utils.pyindex 1f41074e1..8b9710f61 100644--- a/test/test_cpu/utils/test_utils.py+++ b/test/test_cpu/utils/test_utils.py@@ -1,7 +1,10 @@-from unittest.mock import patch+from unittest.mock import MagicMock, patch++import pytestimport auto_round.utils.device as auto_round_utilsfrom auto_round.utils.common import (+ compress_layer_names,preserve_original_visual_block_name,revert_checkpoint_conversion_mapping,)@@ -63,3 +66,123 @@ def test_preserve_original_mllm_language_block_name():)== "model.visual.blocks,model.language_model.layers,model.layers")+++class TestPredefinedIgnoreLayersBlockFilter:+ """Test the block-prefix filter in configure_layer_config."""++ @staticmethod+ def _make_compressor_stub(predefined_ignore_layers, quant_block_list):+ """Create a minimal stub that allows calling configure_layer_config."""+ from auto_round.compressors.base import BaseCompressor++ stub = object.__new__(BaseCompressor)+ # Minimal attributes required by configure_layer_config+ stub.ignore_layers = ""+ stub.quant_block_list = quant_block_list+ stub.is_auto_scheme = False+ stub.orig_scheme = "W4A16"+ stub.layer_config = None+ stub.quant_lm_head = False+ stub.scale_dtype = "fp16"++ # Mock compress_context (no gguf format)+ stub.compress_context = MagicMock()+ stub.compress_context.formats = None++ # Mock model_context with a fake model+ stub.model_context = MagicMock()+ stub.model_context.model = MagicMock()+ stub.model_context.is_mllm = False++ return stub++ @patch("auto_round.compressors.base.get_predefined_ignore_layers")+ @patch("auto_round.compressors.base.set_layer_config")+ @patch("auto_round.compressors.base._handle_special_schemes", return_value=None)+ def test_kimi_k25_ignore_layers_preserved(self, mock_handle, mock_set_lc, mock_get_predefined):+ """KIMI K2.5: vision_tower and mm_projector must end up in ignore_layers."""+ mock_get_predefined.return_value = ["vision_tower", "mm_projector"]+ mock_set_lc.return_value = ({}, False, None)++ stub = self._make_compressor_stub(+ predefined_ignore_layers=["vision_tower", "mm_projector"],+ quant_block_list=[["model.layers"]],+ )+ stub.configure_layer_config()++ assert (+ "vision_tower" in stub.ignore_layers+ ), f"vision_tower should be in ignore_layers, got: '{stub.ignore_layers}'"+ assert (+ "mm_projector" in stub.ignore_layers+ ), f"mm_projector should be in ignore_layers, got: '{stub.ignore_layers}'"++ @patch("auto_round.compressors.base.get_predefined_ignore_layers")+ @patch("auto_round.compressors.base.set_layer_config")+ @patch("auto_round.compressors.base._handle_special_schemes", return_value=None)+ def test_step3p5_ignore_layers_preserved(self, mock_handle, mock_set_lc, mock_get_predefined):+ """step3p5: short pattern names must not be dropped by block filter."""+ predefined = ["g_proj", "moe.gate", "eh_proj", "shared_head", "layers.45"]+ mock_get_predefined.return_value = predefined+ mock_set_lc.return_value = ({}, False, None)++ stub = self._make_compressor_stub(+ predefined_ignore_layers=predefined,+ quant_block_list=[["model.layers"]],+ )+ stub.configure_layer_config()++ for name in predefined:+ assert name in stub.ignore_layers, f"'{name}' should be in ignore_layers, got: '{stub.ignore_layers}'"++ @patch("auto_round.compressors.base.get_predefined_ignore_layers")+ @patch("auto_round.compressors.base.set_layer_config")+ @patch("auto_round.compressors.base._handle_special_schemes", return_value=None)+ def test_longcat_classifier_preserved(self, mock_handle, mock_set_lc, mock_get_predefined):+ """Longcat: 'classifier' must not be dropped."""+ mock_get_predefined.return_value = ["classifier"]+ mock_set_lc.return_value = ({}, False, None)++ stub = self._make_compressor_stub(+ predefined_ignore_layers=["classifier"],+ quant_block_list=[["model.layers"]],+ )+ stub.configure_layer_config()++ assert "classifier" in stub.ignore_layers, f"classifier should be in ignore_layers, got: '{stub.ignore_layers}'"++ @patch("auto_round.compressors.base.get_predefined_ignore_layers")+ @patch("auto_round.compressors.base.set_layer_config")+ @patch("auto_round.compressors.base._handle_special_schemes", return_value=None)+ def test_glm_flash_full_path_inside_block_preserved(self, mock_handle, mock_set_lc, mock_get_predefined):+ """GLM Flash: model.layers.0.mlp (full path inside block) must be kept."""+ mock_get_predefined.return_value = ["model.layers.0.mlp"]+ mock_set_lc.return_value = ({}, False, None)++ stub = self._make_compressor_stub(+ predefined_ignore_layers=["model.layers.0.mlp"],+ quant_block_list=[["model.layers"]],+ )+ stub.configure_layer_config()++ assert (+ "model.layers.0.mlp" in stub.ignore_layers+ ), f"model.layers.0.mlp should be in ignore_layers, got: '{stub.ignore_layers}'"++ @patch("auto_round.compressors.base.get_predefined_ignore_layers")+ @patch("auto_round.compressors.base.set_layer_config")+ @patch("auto_round.compressors.base._handle_special_schemes", return_value=None)+ def test_mllm_with_multiple_block_groups(self, mock_handle, mock_set_lc, mock_get_predefined):+ """MLLM with vision + language blocks: ignore layers outside both are preserved."""+ mock_get_predefined.return_value = ["vision_tower", "mm_projector"]+ mock_set_lc.return_value = ({}, False, None)++ stub = self._make_compressor_stub(+ predefined_ignore_layers=["vision_tower", "mm_projector"],+ quant_block_list=[["model.visual.blocks"], ["model.layers"]],+ )+ stub.configure_layer_config()++ assert "vision_tower" in stub.ignore_layers+ assert "mm_projector" in stub.ignore_layers