Implement aliasable mixin and alias activation ordering (python3.9 fi…

…x) (#218) * implement aliasable mixin and alias activation ordering Signed-off-by: Kyle Sayers <[email protected]> * update docstring Signed-off-by: Kyle Sayers <[email protected]> * fix docstring Signed-off-by: Kyle Sayers <[email protected]> * uncomment Signed-off-by: Kyle Sayers <[email protected]> * rename and make abstract Signed-off-by: Kyle Sayers <[email protected]> * remove property for clarity and to support python3.9 Signed-off-by: Kyle Sayers <[email protected]> --------- Signed-off-by: Kyle Sayers <[email protected]>
neuralmagic · Dec 2, 2024 · 2dcbc9d · 2dcbc9d
1 parent 8571339
commit 2dcbc9d
Show file tree

Hide file tree

Showing 4 changed files with 96 additions and 4 deletions.
diff --git a/src/compressed_tensors/quantization/quant_args.py b/src/compressed_tensors/quantization/quant_args.py
@@ -17,6 +17,7 @@
 from typing import Any, Dict, Optional, Union
 
 import torch
+from compressed_tensors.utils import Aliasable
 from pydantic import BaseModel, Field, field_validator, model_validator
 
 
@@ -53,17 +54,29 @@ class QuantizationStrategy(str, Enum):
     TOKEN = "token"
 
 
-class ActivationOrdering(str, Enum):
+class ActivationOrdering(Aliasable, str, Enum):
     """
     Enum storing strategies for activation ordering
 
     Group: reorder groups and weight\n
-    Weight: only reorder weight, not groups. Slightly lower latency and
-    accuracy compared to group actorder\n
+    Weight: only reorder weight, not groups. Slightly lower accuracy but also lower
+    latency when compared to group actorder\n
+    Dynamic: alias for Group\n
+    Static: alias for Weight\n
     """
 
     GROUP = "group"
     WEIGHT = "weight"
+    # aliases
+    DYNAMIC = "dynamic"
+    STATIC = "static"
+
+    @staticmethod
+    def get_aliases() -> Dict[str, str]:
+        return {
+            "dynamic": "group",
+            "static": "weight",
+        }
 
 
 class QuantizationArgs(BaseModel, use_enum_values=True):

diff --git a/src/compressed_tensors/quantization/quant_scheme.py b/src/compressed_tensors/quantization/quant_scheme.py
@@ -62,6 +62,7 @@ def validate_model_after(model: "QuantizationArgs") -> Dict[str, Any]:
 
         return model
 
+
 """
 Pre-Set Quantization Scheme Args
 """

diff --git a/src/compressed_tensors/utils/helpers.py b/src/compressed_tensors/utils/helpers.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Optional
+from typing import Any, Dict, Optional
 
 import torch
 from transformers import AutoConfig
@@ -24,6 +24,7 @@
     "tensor_follows_mask_structure",
     "replace_module",
     "is_compressed_tensors_config",
+    "Aliasable",
 ]
 
 FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
@@ -119,3 +120,34 @@ def is_compressed_tensors_config(compression_config: Any) -> bool:
         return isinstance(compression_config, CompressedTensorsConfig)
     except ImportError:
         return False
+
+
+class Aliasable:
+    """
+    A mixin for enums to allow aliasing of enum members
+
+    Example:
+    >>> class MyClass(Aliasable, int, Enum):
+    >>>     ...
+    """
+
+    @staticmethod
+    def get_aliases() -> Dict[str, str]:
+        raise NotImplementedError()
+
+    def __eq__(self, other):
+        if isinstance(other, self.__class__):
+            aliases = self.get_aliases()
+            return self.value == other.value or (
+                aliases.get(self.value, self.value)
+                == aliases.get(other.value, other.value)
+            )
+        else:
+            aliases = self.get_aliases()
+            self_value = aliases.get(self.value, self.value)
+            other_value = aliases.get(other, other)
+            return self_value == other_value
+
+    def __hash__(self):
+        canonical_value = self.aliases.get(self.value, self.value)
+        return hash(canonical_value)
diff --git a/tests/test_quantization/test_quant_args.py b/tests/test_quantization/test_quant_args.py
@@ -83,14 +83,28 @@ def test_actorder():
     # test group inference with actorder
     args = QuantizationArgs(group_size=128, actorder=ActivationOrdering.GROUP)
     assert args.strategy == QuantizationStrategy.GROUP
+    args = QuantizationArgs(group_size=128, actorder=ActivationOrdering.DYNAMIC)
+    assert args.strategy == QuantizationStrategy.GROUP
 
     # test invalid pairings
+    with pytest.raises(ValueError):
+        QuantizationArgs(group_size=None, actorder="group")
     with pytest.raises(ValueError):
         QuantizationArgs(group_size=None, actorder="weight")
+    with pytest.raises(ValueError):
+        QuantizationArgs(group_size=None, actorder="static")
+    with pytest.raises(ValueError):
+        QuantizationArgs(group_size=-1, actorder="group")
     with pytest.raises(ValueError):
         QuantizationArgs(group_size=-1, actorder="weight")
+    with pytest.raises(ValueError):
+        QuantizationArgs(group_size=-1, actorder="static")
+    with pytest.raises(ValueError):
+        QuantizationArgs(strategy="tensor", actorder="group")
     with pytest.raises(ValueError):
         QuantizationArgs(strategy="tensor", actorder="weight")
+    with pytest.raises(ValueError):
+        QuantizationArgs(strategy="tensor", actorder="static")
 
     # test boolean and none defaulting
     assert (
@@ -101,6 +115,38 @@ def test_actorder():
     assert QuantizationArgs(group_size=1, actorder=None).actorder is None
 
 
+def test_actorder_aliases():
+    assert (
+        ActivationOrdering.GROUP
+        == ActivationOrdering.DYNAMIC
+        == ActivationOrdering.GROUP
+    )
+    assert (
+        ActivationOrdering.WEIGHT
+        == ActivationOrdering.STATIC
+        == ActivationOrdering.WEIGHT
+    )
+
+    assert ActivationOrdering.GROUP == "dynamic" == ActivationOrdering.GROUP
+    assert ActivationOrdering.DYNAMIC == "dynamic" == ActivationOrdering.DYNAMIC
+    assert ActivationOrdering.GROUP == "group" == ActivationOrdering.GROUP
+    assert ActivationOrdering.DYNAMIC == "group" == ActivationOrdering.DYNAMIC
+
+    assert ActivationOrdering.WEIGHT == "static" == ActivationOrdering.WEIGHT
+    assert ActivationOrdering.STATIC == "static" == ActivationOrdering.STATIC
+    assert ActivationOrdering.WEIGHT == "weight" == ActivationOrdering.WEIGHT
+    assert ActivationOrdering.STATIC == "weight" == ActivationOrdering.STATIC
+
+    assert ActivationOrdering.WEIGHT != "dynamic" != ActivationOrdering.WEIGHT
+    assert ActivationOrdering.STATIC != "dynamic" != ActivationOrdering.STATIC
+    assert ActivationOrdering.WEIGHT != "group" != ActivationOrdering.WEIGHT
+    assert ActivationOrdering.STATIC != "group" != ActivationOrdering.STATIC
+    assert ActivationOrdering.GROUP != "static" != ActivationOrdering.GROUP
+    assert ActivationOrdering.DYNAMIC != "static" != ActivationOrdering.DYNAMIC
+    assert ActivationOrdering.GROUP != "weight" != ActivationOrdering.GROUP
+    assert ActivationOrdering.DYNAMIC != "weight" != ActivationOrdering.DYNAMIC
+
+
 def test_invalid():
     with pytest.raises(ValidationError):
         QuantizationArgs(type="invalid")