Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

Fix some CI issues (#1842)

Liang Lvfd495491548f
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index 6f5dc2fda..f651df215 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -55,6 +55,7 @@ from auto_round.utils import (
is_hpex_available,
is_quantized_input_module,
memory_monitor,
+ preserve_original_visual_block_name,
revert_checkpoint_conversion_mapping,
)
from auto_round.utils.device import (
@@ -556,14 +557,25 @@ class BaseCompressor(object):
def configure_layer_config(self, enable_gguf_official_mixed: bool | None = True) -> None:
"""Build ``self.layer_config`` from the resolved scheme on the patched model."""
- is_gguf_format = (f := getattr(self.compress_context, "formats", None)) is not None and "gguf" in f
+ _formats = getattr(self.compress_context, "formats", None)
+ is_gguf_format = _formats is not None and any(
+ "gguf" in str(getattr(fmt, "output_format", "")) for fmt in _formats
+ )
predefined_ignore_layers = get_predefined_ignore_layers(self.model_context.model)
compressed_predefined_ignore_layers = compress_layer_names(predefined_ignore_layers)
if not is_gguf_format:
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]
+ predefined_ignore_layers = [
+ name
+ for name in predefined_ignore_layers
+ if any(name.startswith(prefix) for prefix in block_prefixes)
+ ]
+ predefined_ignore_layers = compress_layer_names(predefined_ignore_layers)
if predefined_ignore_layers:
logger.info(f"Using predefined ignore_layers: {compressed_predefined_ignore_layers}")
- tmp_str = ",".join(predefined_ignore_layers)
+ tmp_str = predefined_ignore_layers.replace(" ", "")
if self.ignore_layers == "":
self.ignore_layers = tmp_str
else:
@@ -1355,19 +1367,32 @@ class BaseCompressor(object):
if "scale_dtype" in serialization_dict.keys():
serialization_dict["scale_dtype"] = str(serialization_dict["scale_dtype"])
+ original_to_quant_block_names = serialization_dict.get("to_quant_block_names")
+ if isinstance(original_to_quant_block_names, list):
+ original_to_quant_block_names = original_to_quant_block_names[:]
+
# to match the original name
reverse_checkpoint_conversion_mapping = get_reverse_checkpoint_conversion_mapping(self.model)
if isinstance(serialization_dict["to_quant_block_names"], str):
- serialization_dict["to_quant_block_names"] = revert_checkpoint_conversion_mapping(
+ reverted_block_name = revert_checkpoint_conversion_mapping(
serialization_dict["to_quant_block_names"], reverse_checkpoint_conversion_mapping
)
+ serialization_dict["to_quant_block_names"] = preserve_original_visual_block_name(
+ original_to_quant_block_names, reverted_block_name
+ )
elif isinstance(serialization_dict["to_quant_block_names"], list):
for idx in range(len(serialization_dict["to_quant_block_names"])):
- serialization_dict["to_quant_block_names"][idx] = revert_checkpoint_conversion_mapping(
+ reverted_block_name = revert_checkpoint_conversion_mapping(
serialization_dict["to_quant_block_names"][idx], reverse_checkpoint_conversion_mapping
)
+ original_block_name = None
+ if isinstance(original_to_quant_block_names, list) and idx < len(original_to_quant_block_names):
+ original_block_name = original_to_quant_block_names[idx]
+ serialization_dict["to_quant_block_names"][idx] = preserve_original_visual_block_name(
+ original_block_name, reverted_block_name
+ )
compressed_model = format.save_quantized(
save_folder,
diff --git a/auto_round/export/export_to_autoround/export.py b/auto_round/export/export_to_autoround/export.py
index 5246c9503..a394a67d8 100644
--- a/auto_round/export/export_to_autoround/export.py
+++ b/auto_round/export/export_to_autoround/export.py
@@ -280,10 +280,12 @@ def save_quantized_as_autoround(
extra_config = {}
block_name_to_quantize = quantization_config["block_name_to_quantize"]
if isinstance(block_name_to_quantize, str):
- block_name_to_quantize = block_name_to_quantize.split(",")
+ block_name_to_quantize = [name.strip() for name in block_name_to_quantize.split(",")]
elif isinstance(block_name_to_quantize, list):
- for i in range(len(block_name_to_quantize)):
- block_name_to_quantize[i] = os.path.commonprefix(block_name_to_quantize[i]).rstrip(".")
+ block_name_to_quantize = [
+ os.path.commonprefix(item).rstrip(".") if isinstance(item, list) else item
+ for item in block_name_to_quantize
+ ]
scheme_keys = [f.name for f in fields(QuantizationScheme)]
for layer_name, cfg in layer_config.items():
diff --git a/auto_round/export/export_to_gguf/convert.py b/auto_round/export/export_to_gguf/convert.py
index 78e0089e0..5245f93c4 100644
--- a/auto_round/export/export_to_gguf/convert.py
+++ b/auto_round/export/export_to_gguf/convert.py
@@ -146,7 +146,7 @@ def get_tensors(cls) -> Iterator[tuple[str, Tensor]]:
if not hasattr(cls.model, "tensor_name_list"):
cls.model.tensor_name_list = []
- reverse_key_mapping = {v: k for k, v in cls.model._checkpoint_conversion_mapping.items()}
+ reverse_key_mapping = {v: k for k, v in getattr(cls.model, "_checkpoint_conversion_mapping", {}).items()}
for name, tensor in cls.model.named_parameters():
if name not in cls.model.tensor_name_list:
cls.model.tensor_name_list.append(name)
diff --git a/auto_round/utils/common.py b/auto_round/utils/common.py
index 14e9033ac..b80a5ba6f 100644
--- a/auto_round/utils/common.py
+++ b/auto_round/utils/common.py
@@ -1161,6 +1161,9 @@ def get_reverse_checkpoint_conversion_mapping(model):
def revert_checkpoint_conversion_mapping(name: str, key_mapping: dict[str, str]) -> str:
+ if "," in name:
+ return ",".join(revert_checkpoint_conversion_mapping(part, key_mapping) for part in name.split(","))
+
for source_pattern, target_patterns in key_mapping.items():
if isinstance(target_patterns, str):
target_patterns = [target_patterns]
@@ -1174,6 +1177,42 @@ def revert_checkpoint_conversion_mapping(name: str, key_mapping: dict[str, str])
return name
+def preserve_original_visual_block_name(original_name: str | None, reverted_name: str) -> str:
+ """Keep composite multimodal block prefixes stable in serialized quant configs.
+
+ Some multimodal models expose block names under the composite model path
+ (for example ``model.visual.*`` or ``model.language_model.*``) during
+ quantization, but checkpoint conversion rules can rewrite those config-only
+ block prefixes to text-submodel paths such as ``visual.*`` or
+ ``model.layers``. The direct multimodal loaders expect the composite path to
+ remain intact in ``block_name_to_quantize``.
+ """
+ if not (isinstance(original_name, str) and isinstance(reverted_name, str)):
+ return reverted_name
+
+ original_parts = [part.strip() for part in original_name.split(",")]
+ reverted_parts = [part.strip() for part in reverted_name.split(",")]
+ if len(original_parts) != len(reverted_parts):
+ return reverted_name
+
+ preserved_parts = []
+ for original_part, reverted_part in zip(original_parts, reverted_parts):
+ if original_part.startswith("model.visual.") and reverted_part == original_part[len("model.") :]:
+ preserved_parts.append(original_part)
+ elif original_part.startswith("model.language_model.") and reverted_part.startswith("model.layers"):
+ preserved_parts.append(original_part)
+ preserved_parts.append(reverted_part)
+ else:
+ preserved_parts.append(reverted_part)
+
+ deduped_parts = []
+ for preserved_part in preserved_parts:
+ if preserved_part not in deduped_parts:
+ deduped_parts.append(preserved_part)
+
+ return ",".join(deduped_parts)
+
+
def apply_checkpoint_conversion_mapping(name: str, key_mapping: dict[str, str]) -> str:
for source_pattern, target_patterns in key_mapping.items():
if isinstance(target_patterns, str):
diff --git a/test/test_cpu/export/test_gguf_format.py b/test/test_cpu/export/test_gguf_format.py
index e7c9a0584..6eee4cdd1 100644
--- a/test/test_cpu/export/test_gguf_format.py
+++ b/test/test_cpu/export/test_gguf_format.py
@@ -218,7 +218,7 @@ class TestGGUF:
assert ar.layer_config["model.embed_tokens"]["bits"] == 8
assert "lm_head" not in ar.layer_config
- model_name = "Qwen/Qwen3-0.6B"
+ model_name = get_model_path("Qwen/Qwen3-0.6B")
ar = AutoRound(model=model_name, scheme="gguf:q4_0", iters=0)
ar.formats = ["gguf:q4_0"]
ar.layer_config, _, _ = set_layer_config(
diff --git a/test/test_cpu/models/test_mllm.py b/test/test_cpu/models/test_mllm.py
index 49b33332a..674b1d718 100644
--- a/test/test_cpu/models/test_mllm.py
+++ b/test/test_cpu/models/test_mllm.py
@@ -236,7 +236,14 @@ class TestAutoRound:
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
quantized_model_path, torch_dtype="auto", device_map="auto"
)
- assert model.config.quantization_config.block_name_to_quantize == "model.visual.blocks,model.layers"
+ block_name_to_quantize = model.config.quantization_config.block_name_to_quantize
+ if isinstance(block_name_to_quantize, str):
+ block_name_to_quantize = block_name_to_quantize.split(",")
+ assert block_name_to_quantize == [
+ "model.visual.blocks",
+ "model.language_model.layers",
+ "model.layers",
+ ]
image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
processor = AutoProcessor.from_pretrained(quantized_model_path)
messages = [
diff --git a/test/test_cpu/utils/test_utils.py b/test/test_cpu/utils/test_utils.py
index 3dec97010..1f41074e1 100644
--- a/test/test_cpu/utils/test_utils.py
+++ b/test/test_cpu/utils/test_utils.py
@@ -1,6 +1,10 @@
from unittest.mock import patch
import auto_round.utils.device as auto_round_utils
+from auto_round.utils.common import (
+ preserve_original_visual_block_name,
+ revert_checkpoint_conversion_mapping,
+)
class TestPackingWithNumba:
@@ -19,3 +23,43 @@ class TestPackingWithNumba:
@patch.object(auto_round_utils, "is_numba_available", lambda: False)
def test_numba_not_installed(self):
assert auto_round_utils.can_pack_with_numba() is False, "`can_pack_with_numba` should return False."
+
+
+def test_revert_checkpoint_conversion_mapping_handles_comma_separated_block_names():
+ mapping = {
+ r"^visual\.": "model.visual.",
+ r"^model\.language_model\.layers": "model.layers",
+ }
+
+ converted = revert_checkpoint_conversion_mapping("visual.blocks,model.language_model.layers", mapping)
+
+ assert converted == "model.visual.blocks,model.layers"
+
+
+def test_preserve_original_visual_block_name():
+ # Single visual block name
+ assert preserve_original_visual_block_name("model.visual.blocks", "visual.blocks") == "model.visual.blocks"
+ # Comma-separated: visual prefix is preserved and language-model keeps an
+ # extra model.layers alias for runtimes that do not use the composite path.
+ assert (
+ preserve_original_visual_block_name(
+ "model.visual.blocks,model.language_model.layers", "visual.blocks,model.layers"
+ )
+ == "model.visual.blocks,model.language_model.layers,model.layers"
+ )
+ # Direct language-model composite paths are preserved and also expose the
+ # runtime alias used by sglang's text submodel.
+ assert (
+ preserve_original_visual_block_name("model.language_model.layers", "model.layers")
+ == "model.language_model.layers,model.layers"
+ )
+
+
+def test_preserve_original_mllm_language_block_name():
+ assert (
+ preserve_original_visual_block_name(
+ "model.visual.blocks,model.language_model.layers",
+ "model.visual.blocks,model.layers",
+ )
+ == "model.visual.blocks,model.language_model.layers,model.layers"
+ )