Update QuantizationScheme defaults (#157)

* remove weight and input details

* add default test

* PR comments

Author: dsikka

src/compressed_tensors/quantization/quant_scheme.py

@@ -57,15 +57,9 @@ def default_scheme(
             # default to quantizing all Linear layers
             targets = ["Linear"]
 
-        # default to 8 bit integer symmetric quantization
-        # for weights
-        weights = QuantizationArgs(num_bits=8, symmetric=True)
-
-        # default to 8 bit integer asymmetric quantization
-        input_activations = QuantizationArgs(num_bits=8, symmetric=True)
-
-        # Do not quantize the output activations
-        # by default
+        # by default, activations and weights are left unquantized
+        weights = None
+        input_activations = None
         output_activations = None
 
         return cls(
@@ -110,6 +104,7 @@ def is_preset_scheme(name: str) -> bool:
     """
     return name.upper() in PRESET_SCHEMES
 
+
 UNQUANTIZED = dict()
 
 # 8 bit integer weights and 8 bit activations quantization

tests/test_quantization/test_quant_scheme.py

@@ -13,7 +13,11 @@
 # limitations under the License.
 
 import pytest
-from compressed_tensors.quantization import QuantizationArgs, QuantizationScheme
+from compressed_tensors.quantization import (
+    QuantizationArgs,
+    QuantizationConfig,
+    QuantizationScheme,
+)
 from pydantic import ValidationError
 
 
@@ -49,3 +53,11 @@ def test_full_scheme():
 def test_needs_targets():
     with pytest.raises(ValidationError):
         _ = QuantizationScheme()
+
+
+def test_defaults():
+    targets = ["Linear"]
+    output = QuantizationScheme.default_scheme(targets=targets)
+    assert output.weights is None
+    assert output.input_activations is None
+    assert output.output_activations is None