c

Author: AlexanderDokuchaev

src/nncf/torch/quantization/quantize_functions.py

@@ -435,8 +435,11 @@ def pack_uint4(tensor: torch.Tensor) -> torch.Tensor:
     :raises nncf.errors.ValidationError: If the input tensor is not of type `torch.uint8`.
     """
     if tensor.dtype != torch.uint8:
-        msg = f"Invalid tensor dtype {tensor.type}. torch.uint8 type is supported."
+        msg = f"Invalid tensor dtype {tensor.dtype}. torch.uint8 type is supported."
         raise ValidationError(msg)
+    if torch.any((tensor < 0) | (tensor > 15)):
+        msg = "Tensor values are not in [0, 15]."
+        raise ValueError(msg)
     packed_tensor = tensor.contiguous()
     packed_tensor = packed_tensor.reshape(-1, 2)
     packed_tensor = torch.bitwise_and(packed_tensor[..., ::2], 15) | packed_tensor[..., 1::2] << 4
@@ -476,8 +479,11 @@ def pack_int4(tensor: torch.Tensor) -> torch.Tensor:
     :raises nncf.errors.ValidationError: If the input tensor is not of type `torch.int8`.
     """
     if tensor.dtype != torch.int8:
-        msg = f"Invalid tensor dtype {tensor.type}. torch.int8 type is supported."
+        msg = f"Invalid tensor dtype {tensor.dtype}. torch.int8 type is supported."
         raise ValidationError(msg)
+    if torch.any((tensor < -8) | (tensor > 7)):
+        msg = "Tensor values are not in [-8, 7]."
+        raise ValueError(msg)
     tensor = tensor + 8
     return pack_uint4(tensor.type(torch.uint8))
 

tests/torch/quantization/test_functions.py

@@ -880,6 +880,18 @@ def test_pack_invalid_dtype_raises(desc: PackUnpackDesc):
         desc.pack_fn(tensor)
 
 
+@pytest.mark.parametrize("desc", PACK_UNPACK_DESCS, ids=str)
+def test_pack_invalid_value(desc: PackUnpackDesc):
+    tensor = torch.tensor([desc.max_value + 1, 0, 0, 0], dtype=desc.input_dtype)
+    with pytest.raises(ValueError, match=r"Tensor values are not in"):
+        desc.pack_fn(tensor)
+
+    if desc.input_dtype.is_signed:
+        tensor = torch.tensor([desc.min_value - 1, 0, 0, 0], dtype=desc.input_dtype)
+        with pytest.raises(ValueError, match=r"Tensor values are not in"):
+            desc.pack_fn(tensor)
+
+
 def test_pack_uint4_layout():
     tensor = torch.tensor([1, 2, 3, 4], dtype=torch.uint8)
     # Two consecutive values [low, high] are packed into one byte as: low | (high << 4).