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

[regression] fix copy missing tensors for llama4 (#1579)

Xin He38c7035525be
diff --git a/auto_round/utils/common.py b/auto_round/utils/common.py
index 23685d2f2..0a699ef84 100644
--- a/auto_round/utils/common.py
+++ b/auto_round/utils/common.py
@@ -795,4 +795,5 @@ def compress_layer_names(names: list) -> str:
range_str = ",".join(ranges)
parts.append(f"{prefix}[{range_str}]{suffix}")
parts.extend(singles)
+ parts.sort()
return ", ".join(parts)
diff --git a/auto_round/utils/missing_tensors.py b/auto_round/utils/missing_tensors.py
index a7f758b3c..481bf7068 100644
--- a/auto_round/utils/missing_tensors.py
+++ b/auto_round/utils/missing_tensors.py
@@ -189,24 +189,34 @@ def copy_missing_tensors_from_source(
shortcut_block_prefix: set = {name.split(".", 1)[1] for name in saved_block_prefix if "." in name}
saved_block_prefix.update(shortcut_block_prefix)
- def _is_truly_missing(name: str) -> bool:
- # Special case: google/gemma-3-4b-it
- # saved tensors have "model.language_model" prefix
+ def _name_aliases(name: str) -> set:
+ """Return a set of equivalent tensor-name variants to match against saved tensors.
+
+ Handles models where the source and saved prefixes differ, e.g.
+ google/gemma-3-4b-it: ``language_model.model.*`` ↔ ``model.language_model.*``.
+ """
+ aliases = {name}
if name.startswith("language_model.model."):
- name = name.replace("language_model.model.", "model.language_model.")
+ aliases.add("model.language_model." + name[len("language_model.model.") :])
+ elif name.startswith("model.language_model."):
+ aliases.add("language_model.model." + name[len("model.language_model.") :])
+ return aliases
+
+ def _is_truly_missing(name: str) -> bool:
# Special case: Qwen/Qwen3-0.6B-FP8
# lm_head is tied but still in source_dir → not missing
if name == "lm_head.weight":
return False
- if name in saved_tensor_names:
+ aliases = _name_aliases(name)
+ if aliases & saved_tensor_names:
return False
- parent = name.rsplit(".", 1)[0]
- if parent in saved_parent_layers:
+ parents = {a.rsplit(".", 1)[0] for a in aliases}
+ if parents & saved_parent_layers:
return False
# For split experts, name is changed but block name is the same.
- src_block = _first_numeric_prefix(name)
- if src_block is not None:
- return src_block not in saved_block_prefix
+ blocks = {_first_numeric_prefix(a) for a in aliases} - {None}
+ if blocks:
+ return not (blocks & saved_block_prefix)
return True
missing_tensor_names: list = [name for name in source_tensor_to_file if _is_truly_missing(name)]
@@ -599,9 +609,12 @@ def _woq_quantize_missing_tensors(target_dir: str, missing_tensors_dict: dict) -
ec.setdefault(layer_name, {"bits": 16, "data_type": "fp"})
qcfg["extra_config"] = ec
if show_log:
+ # 7 is the len('.weight')
+ ignored_layers = [k[:-7] for k in ignored_weight_keys]
+ compressed_ignored_layers = compress_layer_names(ignored_layers)
logger.info(
f"Updated extra_config for {len(ignored_weight_keys)} ignored layer(s): "
- f"{[k[: -len('.weight')] for k in ignored_weight_keys]}"
+ f"{compressed_ignored_layers}"
)
if weight_keys:
existing = qcfg.get("block_name_to_quantize") or []
diff --git a/test/test_cpu/utils/test_missing_tensors.py b/test/test_cpu/utils/test_missing_tensors.py
index 83293af9a..190404f86 100644
--- a/test/test_cpu/utils/test_missing_tensors.py
+++ b/test/test_cpu/utils/test_missing_tensors.py
@@ -618,7 +618,8 @@ class TestCopyMissingTensorsFromSource(unittest.TestCase):
# Target has a different sub-layer inside the same block
_save_safetensors(
{
- "model.language_model.layers.0.mlp.gate_proj.weight": torch.randn(32, 64),
+ "language_model.model.layers.0.mlp.gate_proj.weight": torch.randn(32, 64), # llama4
+ "model.language_model.layers.0.mlp.gate_proj.weight": torch.randn(32, 64), # gemma
"model.language_model.norm.weight": torch.randn(64),
},
os.path.join(target_dir, "model.safetensors"),