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: update tensor summary creation to directly use tensors_to_restor… (#2401)

Xin Hebff3deab0705
diff --git a/auto_round/utils/missing_tensors.py b/auto_round/utils/missing_tensors.py
index 63d06bdc7..832aae485 100644
--- a/auto_round/utils/missing_tensors.py
+++ b/auto_round/utils/missing_tensors.py
@@ -121,7 +121,7 @@ def _restore_special_fp32_tensors(
os.remove(temporary_path)
if tensors_to_restore:
- tensor_summary = compress_layer_names([name.rsplit(".", 1)[0] for name in tensors_to_restore])
+ tensor_summary = compress_layer_names(list(tensors_to_restore))
logger.info(
f"Restored {len(tensors_to_restore)} tensor(s) from FP16/BF16 to their original FP32 values: "
f"{tensor_summary}."
diff --git a/test/unit/common/utils/test_missing_tensors.py b/test/unit/common/utils/test_missing_tensors.py
index e651f4f00..2867450a4 100644
--- a/test/unit/common/utils/test_missing_tensors.py
+++ b/test/unit/common/utils/test_missing_tensors.py
@@ -479,7 +479,7 @@ class TestCopyMissingTensorsFromSource:
copy_missing_tensors_from_source(src, tgt)
assert not os.path.exists(os.path.join(tgt, "model_extra_tensors.safetensors"))
- def test_restores_fp32_tensor_saved_as_fp16(self, tmp_path):
+ def test_restores_fp32_tensor_saved_as_fp16(self, tmp_path, caplog, _autoround_log_propagate):
src, tgt = str(tmp_path / "src"), str(tmp_path / "tgt")
os.makedirs(src)
os.makedirs(tgt)
@@ -501,6 +501,7 @@ class TestCopyMissingTensorsFromSource:
copy_missing_tensors_from_source(src, tgt)
result = _load_safetensors(os.path.join(tgt, "model.safetensors"))
+ assert "model.special.weight" in caplog.text
assert result["model.special.weight"].dtype == torch.float32
assert torch.equal(result["model.special.weight"], original)
assert torch.equal(result["model.unchanged.weight"], torch.tensor([9.0], dtype=torch.float32))