compressed-tensors 0.10.3a20250715__tar.gz → 0.10.3a20250716__tar.gz

{compressed_tensors-0.10.3a20250715/src/compressed_tensors.egg-info → compressed_tensors-0.10.3a20250716}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250715
+Version: 0.10.3a20250716
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/src/compressed_tensors/transform/factory/base.py

@@ -117,10 +117,8 @@ class TransformFactory(RegistryMixin, ABC):
             TransformLocation.WEIGHT_INPUT,
             TransformLocation.WEIGHT_OUTPUT,
         ):
-            assert isinstance(module, torch.nn.Linear)
-            assert module.bias is None
-
             # fuse transform into weight
+            assert hasattr(module, "weight")
             with torch.no_grad(), align_module_device(module):
                 update_offload_parameter(module, "weight", transform(module.weight))
 

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/src/compressed_tensors/transform/factory/hadamard.py

@@ -14,13 +14,14 @@
 
 from typing import Optional, Union
 
+import math
 import torch
 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
@@ -51,8 +52,8 @@ class HadamardFactory(TransformFactory):
         :param module: parent module that transform will be applied to
         :param args: defines how the transform will be applied to the module
         """
-        assert isinstance(module, Linear)
-        size = get_matrix_size(module, args.location)
+        assert hasattr(module, "weight")
+        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)
@@ -60,7 +61,7 @@ class HadamardFactory(TransformFactory):
         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 +82,18 @@ class HadamardFactory(TransformFactory):
 
 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
+        self._scale = math.sqrt(weight.size(0))
 
     def forward(self, value: Tensor) -> Tensor:
         weight = self.weight
@@ -96,5 +103,7 @@ class HadamardTransform(TransformBase):
 
         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
+        ) / self._scale

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -17,9 +17,9 @@ from typing import Optional
 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
@@ -50,8 +50,8 @@ class RandomMatrixFactory(TransformFactory):
         :param module: parent module that transform will be applied to
         :param args: defines how the transform will be applied to the module
         """
-        assert isinstance(module, Linear)
-        size = get_matrix_size(module, args.location)
+        assert hasattr(module, "weight")
+        size = get_transform_size(module, args.location, self.scheme.head_dim)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
 
@@ -59,7 +59,7 @@ class RandomMatrixFactory(TransformFactory):
         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 @@ class RandomMatrixFactory(TransformFactory):
 
 
 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:

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/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)

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/src/compressed_tensors/transform/utils/hadamard.py

@@ -59,7 +59,7 @@ def deterministic_hadamard_matrix(
     for _ in range(log2):
         H = torch.vstack((torch.hstack((H, H)), torch.hstack((H, -H))))
 
-    return H / math.sqrt(size)
+    return H
 
 
 def random_hadamard_matrix(
@@ -86,7 +86,7 @@ def random_hadamard_matrix(
     Q = Q.to(device=device)
     Q = Q * 2 - 1
     Q = torch.diag(Q)
-    return _matmul_hadU(Q) / math.sqrt(size)
+    return _matmul_hadU(Q)
 
 
 def is_pow2(n: int) -> bool:

compressed_tensors-0.10.3a20250716/src/compressed_tensors/transform/utils/matrix.py

@@ -0,0 +1,179 @@
+# 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
+    elif isinstance(module, torch.nn.Embedding):
+        if location in (TransformLocation.INPUT, TransformLocation.WEIGHT_INPUT):
+            size = module.num_embeddings
+        else:
+            size = module.embedding_dim
+    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(
+    transform_weight: torch.Tensor,
+    value: torch.Tensor,
+    location: TransformLocation,
+    module_type: type[torch.nn.Module],
+) -> torch.Tensor:
+    """
+    Using the transform location, 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 transform_weight: transform weight to apply
+    :param value: value to apply transform_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 transform_weight has been applied
+    """
+
+    assert transform_weight.shape[0] == transform_weight.shape[1]
+
+    if module_type == torch.nn.Linear:
+        if location == TransformLocation.INPUT:
+            return _multihead_matmul(value, transform_weight)
+
+        elif location == TransformLocation.WEIGHT_INPUT:
+            # equivalent to (transform_weight @ value.T).T
+            return _multihead_matmul(value, transform_weight.T)
+
+        elif location == TransformLocation.WEIGHT_OUTPUT:
+            # equivalent to (value.T @ transform_weight).T
+            return _multihead_matmul(transform_weight.T, value)
+
+        elif location == TransformLocation.OUTPUT:
+            return _multihead_matmul(value, transform_weight)
+
+    # similar derivation to torch.nn.Linear, but `y = (x W)`
+    elif module_type == torch.nn.Embedding:
+        if location == TransformLocation.INPUT:
+            return _multihead_matmul(value, transform_weight)
+
+        elif location == TransformLocation.WEIGHT_INPUT:
+            return _multihead_matmul(
+                transform_weight,
+                value,
+            )
+
+        elif location == TransformLocation.WEIGHT_OUTPUT:
+            return _multihead_matmul(value, transform_weight)
+
+        elif location == TransformLocation.OUTPUT:
+            return _multihead_matmul(value, transform_weight)
+
+    raise NotImplementedError(
+        f"Applying transforms to {module_type} {location} is not supported"
+    )
+
+
+def _multihead_matmul(A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
+    """
+    Performs A @ B for last two dims of two matrices A and B that possibly
+    have different shapes, as is the case in multi-headed dimension. If
+    shapes are different, this is equivalent to converting the last two dims
+    of the smaller matrix into a block-diagonal matrix with the same shape as
+    the last two dims of the larger matrix.
+
+    E.g. if A is half the size of B, this function will perform
+    [[A  ]  @ B
+     [  A]]
+
+    If B is a third of the size of A, this function will perform
+    A @ [[B    ]
+         [  B  ]
+         [    B]]
+
+    This function will error out if the shapes are not evenly divisble
+
+    :param A: left-hand tensor
+    :param B: right-hand tensor
+    :return: result
+    """
+    if A.shape[-1] > B.shape[-2]:
+        head_dim = B.shape[-2]
+        num_heads = A.shape[-1] // head_dim
+        A = A.unflatten(-1, (num_heads, head_dim))
+        return (A @ B).flatten(-2, -1)
+    elif A.shape[-1] < B.shape[-2]:
+        head_dim = A.shape[-1]
+        num_heads = B.shape[-2] // head_dim
+        B = B.unflatten(-2, (num_heads, head_dim))
+        return (A @ B).flatten(-3, -2)
+    else:
+        return A @ B

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250715'
+__version__ = version = '0.10.3.a20250716'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250715
+Version: 0.10.3a20250716
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250716}/src/compressed_tensors.egg-info/SOURCES.txt

@@ -84,7 +84,7 @@ src/compressed_tensors/transform/factory/random_hadamard.py
 src/compressed_tensors/transform/utils/__init__.py
 src/compressed_tensors/transform/utils/hadamard.py
 src/compressed_tensors/transform/utils/hadamards.safetensors
-src/compressed_tensors/transform/utils/utils.py
+src/compressed_tensors/transform/utils/matrix.py
 src/compressed_tensors/utils/__init__.py
 src/compressed_tensors/utils/helpers.py
 src/compressed_tensors/utils/internal.py

compressed_tensors-0.10.3a20250716/tests/test_transform/conftest.py

@@ -0,0 +1,115 @@
+# 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.
+
+import pytest
+import torch
+from compressed_tensors.transform import TransformArgs
+
+
+class TransformableModel(torch.nn.Module):
+    def __init__(self, *sizes):
+        super().__init__()
+        self.fcs = torch.nn.ModuleList(
+            [
+                torch.nn.Linear(sizes[index], sizes[index + 1], bias=False)
+                for index in range(0, len(sizes) - 1)
+            ]
+        )
+
+    def forward(self, x):
+        for layer in self.fcs:
+            x = layer(x)
+        return x
+
+
+class MockAttention(torch.nn.Module):
+    def __init__(
+        self, hidden_size: int, num_attention_heads: int, num_key_value_heads: int
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+
+        self.num_key_value_groups = num_attention_heads // num_key_value_heads
+        self.head_dim = hidden_size // num_attention_heads
+        self.scaling = self.head_dim**-0.5
+        assert hidden_size >= num_attention_heads * self.head_dim
+
+        self.q_proj = torch.nn.Linear(
+            hidden_size, num_attention_heads * self.head_dim, bias=False
+        )
+        self.k_proj = torch.nn.Linear(
+            hidden_size, num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = torch.nn.Linear(
+            hidden_size, num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = torch.nn.Linear(
+            num_attention_heads * self.head_dim, hidden_size, bias=False
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, seq_len, hidden_size = hidden_states.shape
+        hidden_shape = (batch_size, seq_len, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        key_states = self.repeat_kv(key_states, self.num_key_value_groups)
+        value_states = self.repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = (
+            torch.matmul(query_states, key_states.transpose(2, 3)) * self.scaling
+        )
+
+        attn_weights = torch.nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32
+        ).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+        attn_output = attn_output.reshape((batch_size, seq_len, -1)).contiguous()
+
+        return self.o_proj(attn_output)
+
+    def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+        batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+        if n_rep == 1:
+            return hidden_states
+        hidden_states = hidden_states[:, :, None, :, :].expand(
+            batch, num_key_value_heads, n_rep, slen, head_dim
+        )
+        return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+@pytest.fixture(scope="function")
+def model_apply():
+    model = TransformableModel(2, 4, 8, 16, 32, 64)
+    apply = [
+        # weight output -> input
+        TransformArgs(targets="fcs.0", location="weight_output"),
+        TransformArgs(targets="fcs.1", location="input", inverse=True),
+        # output -> weight input
+        TransformArgs(targets="fcs.1", location="output"),
+        TransformArgs(targets="fcs.2", location="weight_input", inverse=True),
+        # output -> input
+        TransformArgs(targets="fcs.2", location="output"),
+        TransformArgs(targets="fcs.3", location="input", inverse=True),
+        # weight output -> weight input
+        TransformArgs(targets="fcs.3", location="weight_output"),
+        TransformArgs(targets="fcs.4", location="weight_input", inverse=True),
+    ]
+
+    return model, apply

compressed_tensors-0.10.3a20250716/tests/test_transform/factory/test_correctness.py

@@ -0,0 +1,168 @@
+# 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.
+
+import pytest
+import torch
+from compressed_tensors.transform import (
+    TransformArgs,
+    TransformConfig,
+    TransformFactory,
+    TransformScheme,
+    apply_transform_config,
+)
+from compressed_tensors.utils import offloaded_dispatch
+from tests.test_transform.conftest import MockAttention
+from tests.testing_utils import requires_accelerate, requires_gpu
+
+
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("head_dim", (None, 2, 4))
+@pytest.mark.parametrize("input_batch_size", (1, 5, 17))
+def test_correctness_linear(type, randomized, head_dim, input_batch_size):
+    size = (4, 8)
+    module = torch.nn.Linear(*size, bias=False)
+    scheme = TransformScheme(type=type, randomized=randomized, head_dim=head_dim)
+    factory = TransformFactory.from_scheme(scheme, name="")
+
+    input_tfm = factory.create_transform(
+        module, TransformArgs(targets="Linear", location="input", inverse=True)
+    )
+    w_in_tfm = factory.create_transform(
+        module, TransformArgs(targets="Linear", location="weight_input")
+    )
+    w_out_tfm = factory.create_transform(
+        module, TransformArgs(targets="Linear", location="weight_output")
+    )
+    output_tfm = factory.create_transform(
+        module, TransformArgs(targets="Linear", location="output", inverse=True)
+    )
+
+    input = torch.rand((input_batch_size, 5, size[0]))
+    true_output = input @ module.weight.T
+    input_transformed = input_tfm(input)
+    weight_transformed = w_out_tfm(w_in_tfm(module.weight))
+    output = output_tfm(input_transformed @ weight_transformed.T)
+    assert torch.allclose(true_output, output, atol=1e-5, rtol=0.0)
+
+
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("embed_loc", ("weight_output", "output"))
+@pytest.mark.parametrize("linear_loc", ("input", "weight_input"))
+def test_correctness_embedding(type, randomized, embed_loc, linear_loc):
+    model = torch.nn.Sequential(
+        torch.nn.Embedding(2, 4),
+        torch.nn.Linear(4, 8, bias=False),
+    )
+
+    input = torch.randint(high=1, low=0, size=(17, 5, 2))
+    true_output = model(input)
+
+    config = TransformConfig(
+        config_groups={
+            "": TransformScheme(
+                type=type,
+                randomized=randomized,
+                apply=[
+                    TransformArgs(targets="Embedding", location=embed_loc),
+                    TransformArgs(targets="Linear", location=linear_loc, inverse=True),
+                ],
+            )
+        }
+    )
+    apply_transform_config(model, config)
+
+    # compare outputs
+    output = model(input)
+    assert torch.allclose(true_output, output, atol=1e-5, rtol=0.0)
+
+
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("input_batch_size", (1, 5, 17))
+def test_correctness_model(
+    type, randomized, input_batch_size, model_apply, offload=False
+):
+    # load model
+    model = model_apply[0]
+    if offload:
+        model = offloaded_dispatch(model, torch.device("cuda"))
+
+    # get output
+    input = torch.rand((input_batch_size, 5, model.fcs[0].in_features))
+    if offload:
+        input = input.to(torch.device("cuda"))
+    true_output = model(input)
+
+    # apply transforms
+    config = TransformConfig(
+        config_groups={
+            "": TransformScheme(type=type, randomized=randomized, apply=model_apply[1])
+        }
+    )
+    apply_transform_config(model, config)
+
+    # compare outputs
+    output = model(input)
+    assert torch.allclose(true_output, output, atol=1e-5, rtol=0.0)
+
+
+@requires_gpu
+@requires_accelerate()
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("input_batch_size", (1, 5, 17))
+def test_correctness_model_offload(type, randomized, input_batch_size, model_apply):
+    test_correctness_model(
+        type, randomized, input_batch_size, model_apply, offload=True
+    )
+
+
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("head_dim", (4, 8))
+@pytest.mark.parametrize("input_batch_size", (1, 5, 17))
+def test_correctness_attention_heads(type, randomized, head_dim, input_batch_size):
+    hidden_size = 64
+    num_attention_heads = 8
+
+    attention = MockAttention(
+        hidden_size=hidden_size,
+        num_attention_heads=num_attention_heads,
+        num_key_value_heads=head_dim,
+    )
+
+    input = torch.rand(input_batch_size, 5, hidden_size)
+    true_output = attention(input)
+
+    config = TransformConfig(
+        config_groups={
+            "": TransformScheme(
+                type=type,
+                randomized=randomized,
+                head_dim=head_dim,
+                apply=[
+                    TransformArgs(targets="v_proj", location="weight_output"),
+                    TransformArgs(
+                        targets="o_proj", location="weight_input", inverse=True
+                    ),
+                ],
+            )
+        }
+    )
+    apply_transform_config(attention, config)
+
+    output = attention(input)
+    assert torch.allclose(true_output, output, atol=1e-5, rtol=0.0)