PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Fix checkpoint remap (#2001)
Yi Liuec73a8ddae8d
diff --git a/auto_round/utils/common.py b/auto_round/utils/common.pyindex 5e7a50e6a..236d56c05 100644--- a/auto_round/utils/common.py+++ b/auto_round/utils/common.py@@ -1192,7 +1192,18 @@ def revert_checkpoint_conversion_mapping(name: str, key_mapping: dict[str, str])for target_pattern in target_patterns:source_pattern = source_pattern.lstrip("^") # strip off un-needed chars and patternssource_pattern = re.sub(r"\(.*\)", "", source_pattern)- name, n_replace = re.subn(source_pattern, target_pattern, name)++ # Weight-conversion reverse mappings may expose bare tensor names+ # such as "weight" -> ".weight_packed". Treat those as terminal+ # tensor suffixes so they do not rewrite already-quantized names+ # like ".weight_packed" into malformed keys such as+ # "..weight_packed_packed".+ if re.fullmatch(r"[A-Za-z_][A-Za-z0-9_]*", source_pattern) and target_pattern.startswith("."):+ match_pattern = rf"\.{re.escape(source_pattern)}$"+ else:+ match_pattern = source_pattern++ name, n_replace = re.subn(match_pattern, target_pattern, name)# Early exit of the loopif n_replace > 0:return namediff --git a/test/test_cpu/utils/test_utils.py b/test/test_cpu/utils/test_utils.pyindex 8b9710f61..587fe57d7 100644--- a/test/test_cpu/utils/test_utils.py+++ b/test/test_cpu/utils/test_utils.py@@ -39,6 +39,23 @@ def test_revert_checkpoint_conversion_mapping_handles_comma_separated_block_nameassert converted == "model.visual.blocks,model.layers"+def test_revert_checkpoint_conversion_mapping_does_not_rewrite_quantized_tensor_suffixes():+ mapping = {"weight": [".weight_packed", ".weight_scale", ".weight_shape"]}++ assert (+ revert_checkpoint_conversion_mapping("model.layers.0.mlp.down_proj.weight", mapping)+ == "model.layers.0.mlp.down_proj.weight_packed"+ )+ assert (+ revert_checkpoint_conversion_mapping("model.layers.0.mlp.down_proj.weight_packed", mapping)+ == "model.layers.0.mlp.down_proj.weight_packed"+ )+ assert (+ revert_checkpoint_conversion_mapping("model.layers.0.mlp.down_proj.weight_scale", mapping)+ == "model.layers.0.mlp.down_proj.weight_scale"+ )++def test_preserve_original_visual_block_name():# Single visual block nameassert preserve_original_visual_block_name("model.visual.blocks", "visual.blocks") == "model.visual.blocks"