[Transform] Implement multi-headed transforms

src/compressed_tensors/transform/factory/hadamard.py

@@ -18,9 +18,9 @@
 from compressed_tensors.transform import TransformArgs, TransformScheme
 from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
 from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
-from compressed_tensors.transform.utils.utils import (
+from compressed_tensors.transform.utils.matrix import (
     apply_transform_weight,
-    get_matrix_size,
+    get_transform_size,
 )
 from compressed_tensors.utils import get_execution_device, get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
@@ -52,15 +52,15 @@ def create_transform(self, module: Module, args: TransformArgs):
         :param args: defines how the transform will be applied to the module
         """
         assert isinstance(module, Linear)
-        size = get_matrix_size(module, args.location)
+        size = get_transform_size(module, args.location, self.scheme.head_dim)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
         exec_device = get_execution_device(module)
 
         factory_kwargs = {"construct_device": exec_device}
         weight = self.weights.get(size, dtype, device, factory_kwargs=factory_kwargs)
         perm = self.perms[weight] if self.scheme.randomize else None
-        return HadamardTransform(weight, perm, args)
+        return HadamardTransform(weight, perm, args, type(module))
 
     def _create_weight(
         self,
@@ -81,12 +81,17 @@ def _create_permutation(self, weight: Parameter) -> Parameter:
 
 class HadamardTransform(TransformBase):
     def __init__(
-        self, weight: Parameter, perm: Union[Parameter, None], args: TransformArgs
+        self,
+        weight: Parameter,
+        perm: Optional[Parameter],
+        args: TransformArgs,
+        module_type: type[torch.nn.Module],
     ):
         super().__init__()
         self.weight = weight
         self.perm = perm
         self.args = args
+        self.module_type = module_type
 
     def forward(self, value: Tensor) -> Tensor:
         weight = self.weight
@@ -97,4 +102,6 @@ def forward(self, value: Tensor) -> Tensor:
         if self.args.inverse:
             weight = weight.T
 
-        return apply_transform_weight(weight, value, self.args.location)
+        return apply_transform_weight(
+            weight, value, self.args.location, self.module_type
+        )

src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -17,9 +17,9 @@
 import torch
 from compressed_tensors.transform import TransformArgs, TransformScheme
 from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
-from compressed_tensors.transform.utils.utils import (
+from compressed_tensors.transform.utils.matrix import (
     apply_transform_weight,
-    get_matrix_size,
+    get_transform_size,
 )
 from compressed_tensors.utils import get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
@@ -51,15 +51,15 @@ def create_transform(self, module: Module, args: TransformArgs):
         :param args: defines how the transform will be applied to the module
         """
         assert isinstance(module, Linear)
-        size = get_matrix_size(module, args.location)
+        size = get_transform_size(module, args.location, self.scheme.head_dim)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
 
         weight = self.weights[size, dtype, device]
         if args.inverse:
             weight = self.inverses[weight]
 
-        return RandomMatrixTransform(weight, args)
+        return RandomMatrixTransform(weight, args, type(module))
 
     def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
         # TODO: verify that weight is invertible (has non-zero determinant)
@@ -74,17 +74,27 @@ def _create_inverse(self, weight: Parameter) -> Parameter:
 
 
 class RandomMatrixTransform(TransformBase):
-    def __init__(self, weight: Tensor, args: TransformArgs):
+    def __init__(
+        self,
+        weight: Tensor,
+        args: TransformArgs,
+        module_type: type[torch.nn.Module],
+    ):
         super().__init__()
         self.weight = weight  # is an inverse if args.inverse
         self.args = args
+        self.module_type = module_type
 
     def forward(self, value: Tensor) -> Parameter:
-        return apply_transform_weight(self.weight, value, self.args.location)
+        return apply_transform_weight(
+            self.weight, value, self.args.location, self.module_type
+        )
 
     def right_inverse(self, value: Tensor) -> Tensor:
         inverse = high_precision_invert(self.weight)
-        return apply_transform_weight(inverse, value, self.args.location)
+        return apply_transform_weight(
+            inverse, value, self.args.location, self.module_type
+        )
 
 
 def high_precision_invert(weight: Tensor) -> Tensor:

src/compressed_tensors/transform/transform_scheme.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import List
+from typing import List, Optional
 
 from compressed_tensors.transform import TransformArgs
 from pydantic import BaseModel, Field
@@ -40,3 +40,4 @@ class TransformScheme(BaseModel):
     apply: List[TransformArgs] = Field(default_factory=list)
     randomize: bool = Field(default=False)
     requires_grad: bool = Field(default=False)
+    head_dim: Optional[int] = Field(default=None)

src/compressed_tensors/transform/utils/matrix.py

@@ -0,0 +1,147 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Tuple
+
+import torch
+from compressed_tensors.transform import TransformLocation
+
+
+__all__ = ["get_transform_size", "apply_transform_weight"]
+
+
+def get_transform_size(
+    module: torch.nn.Module,
+    location: TransformLocation,
+    head_dim: Optional[int] = None,
+) -> int:
+    """
+    Determine the size of a transform matrix given its location on the module
+
+    :param module: module that matrix will be applied to
+    :param location: location on module
+    :param head_dim: size of head when transform is applied to mha
+    :return: size of matrix
+    """
+    if isinstance(module, torch.nn.Linear):
+        if location in (TransformLocation.INPUT, TransformLocation.WEIGHT_INPUT):
+            size = module.in_features
+        else:
+            size = module.out_features
+    else:
+        raise NotImplementedError(f"Transforms on {type(module)} are not supported")
+
+    if head_dim is not None:
+        if size % head_dim != 0:
+            raise ValueError(
+                f"{head_dim} must divide {size} for {type(module)} at {location}"
+            )
+
+        size = head_dim
+
+    return size
+
+
+def apply_transform_weight(
+    weight: torch.Tensor,
+    value: torch.Tensor,
+    location: TransformLocation,
+    module_type: type[torch.nn.Module],
+) -> torch.Tensor:
+    """
+    :param weight: transform weight to apply
+    :param value: value to apply weight to
+    :param location: determines how weight should be applied
+    :param model_type: result of type(module), passed in to determine application of
+        weight transform
+    :return: value after weight has been applied
+    """
+    fn, axis = _get_transform_method(module_type, location)
+
+    assert weight.shape[0] == weight.shape[1]
+    head_dim = weight.shape[0]
+    num_heads = value.shape[axis] // head_dim
+
+    value = value.unflatten(axis, (num_heads, head_dim))
+    value = fn(weight, value)
+    value = value.flatten(axis - 1, axis)
+
+    return value
+
+
+def _get_transform_method(
+    module_type: type[torch.nn.Module],
+    location: TransformLocation,
+) -> Tuple[Callable[[torch.Tensor, torch.Tensor], torch.Tensor], int]:
+    """
+    Using the transform location, determine how to apply the transform weight to the
+    given value wrt linear weights. For more info on input and output transforms,
+    see `TransformLocation`
+
+    The following explains how weights should be applied to values according to location
+
+    let  x          be input activation
+         W          be weight,
+         yh, xh, Wh be transformed output, input, weight
+
+    note that
+         y  = (x W.T)        // torch.nn.Linear
+
+    Choose values for yh, xh, and Wh which incorporate matrix transforms
+
+    let  V, Vi      be transform matrices on input side
+         U, Ui      be transform matrices on output side
+
+    pick xh = (x V)
+         Wh = (U.T W Vi.T)
+         yh = (y U)
+
+    The following shows that `yh = (xh) (Wh).T` for the chosen values of yh, xh, and Wh
+
+    (xh) (Wh).T = (x V) (U.T W Vi.T).T
+                = (x V) (Vi W.T U)        // transpose matrix product identity
+                = (x W.T) U
+                = y U
+                = yh
+
+    :param weight: transform weight to apply
+    :param value: value to apply weight to
+    :param location: determines how weight should be applied
+    :return: value after transform weight has been applied
+    """
+    fn = axis = None
+
+    if module_type == torch.nn.Linear:
+        if location == TransformLocation.INPUT:
+            fn = lambda weight, value: value @ weight
+            axis = -1
+
+        elif location == TransformLocation.WEIGHT_INPUT:
+            fn = lambda weight, value: value @ weight.T
+            axis = -1
+
+        elif location == TransformLocation.WEIGHT_OUTPUT:
+            fn = lambda weight, value: weight.T @ value
+            axis = -2
+
+        elif location == TransformLocation.OUTPUT:
+            fn = lambda weight, value: value @ weight
+            axis = -1
+
+    if fn is None:
+        raise NotImplementedError(
+            f"Applying transforms to {module_type} {location} is not supported"
+        )
+
+    return fn, axis