minimal diff

Signed-off-by: shanjiaz <[email protected]>

Author: shanjiaz

tests/test_compressors/model_compressors/test_model_compressor.py

@@ -572,12 +572,9 @@ def test_decompress_model(model_stub, comp_stub):
     # equivalent to decompressing from disk
     assert decompressed.keys() == true_decompressed.keys()
     for key in decompressed.keys():
-        # Skip dtype check for weight_shape - int32/int64 are functionally equivalent
-        # torch.Size() works identically with both, old checkpoints use int64, new use int32
-        if not key.endswith("weight_shape"):
-            assert (
-                decompressed[key].dtype == true_decompressed[key].dtype
-            ), f"{key} dtypes not equal"
+        assert (
+            decompressed[key].dtype == true_decompressed[key].dtype
+        ), f"{key} dtypes not equal"
         assert torch.all(
             decompressed[key] == true_decompressed[key]
         ), f"{key} values not equal"

tests/test_compressors/quantized_compressors/test_packed_asym_decompression.py

@@ -114,7 +114,6 @@ def test_end_to_end_asymmetric_quantization(
     # Verify compression created zero-point parameters
     assert hasattr(model.layer1, "weight_zero_point")
     assert hasattr(model.layer2, "weight_zero_point")
-    # For asymmetric GROUP/CHANNEL quantization, zero_point should be packed to int32
     assert model.layer1.weight_zero_point.dtype == torch.int32
     assert model.layer2.weight_zero_point.dtype == torch.int32