[Transform] apply_transform_config

setup.py

@@ -88,7 +88,7 @@ def _setup_packages() -> List:
     )
 
 def _setup_install_requires() -> List:
-    return ["torch>=1.7.0", "transformers", "pydantic>=2.0"]
+    return ["torch>=1.7.0", "transformers", "pydantic>=2.0", "frozendict"]
 
 def _setup_extras() -> Dict:
     return {

src/compressed_tensors/quantization/lifecycle/apply.py

@@ -152,11 +152,7 @@ def apply_quantization_config(
     # list of submodules to ignore
     ignored_submodules = defaultdict(list)
     # mark appropriate layers for quantization by setting their quantization schemes
-    for name, submodule in iter_named_quantizable_modules(
-        model,
-        include_children=True,
-        include_attn=True,
-    ):  # child modules and attention modules
+    for name, submodule in model.named_modules():  # child modules and attention modules
         # potentially fix module name to remove FSDP wrapper prefix
         name = fix_fsdp_module_name(name)
         if matches := find_name_or_class_matches(name, submodule, config.ignore):

src/compressed_tensors/transform/__init__.py

@@ -23,3 +23,4 @@
 from .factory.hadamard import *
 from .factory.matrix_multiply import *
 from .factory.random_hadamard import *
+from .apply import *

src/compressed_tensors/transform/apply.py

@@ -0,0 +1,32 @@
+# 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 torch
+from compressed_tensors.transform import TransformConfig, TransformFactory
+
+
+__all__ = ["apply_transform_config"]
+
+
+def apply_transform_config(model: torch.nn.Module, config: TransformConfig):
+    """
+    Apply a transform config to a model. Weight transforms are fused into weights, while
+    activation transforms are attached as submodules and trigger via pytorch hooks
+
+    :param model: model to apply config to
+    :param config: transform config to apply
+    """
+    for name, scheme in config.config_groups.items():
+        factory = TransformFactory.from_scheme(scheme, name=name)
+        factory.apply_to_model(model)

src/compressed_tensors/transform/factory/base.py

@@ -26,6 +26,7 @@
 )
 from compressed_tensors.utils import (
     align_module_device,
+    delete_offload_module,
     has_offloaded_params,
     patch_attr,
     register_offload_module,
@@ -99,7 +100,7 @@ def _apply_to_module(self, module: Module, args: TransformArgs):
         # create transform as submodule
         transform_name = f"{self.name}_{args.location.value}"
         transform = self.create_transform(module, args)
-        register_offload_module(module, transform_name, transform)  # (1)
+        register_offload_module(module, transform_name, transform)
 
         # register input transformation hook
         if args.location == TransformLocation.INPUT:
@@ -118,6 +119,7 @@ def input_hook(_, args):
             assert isinstance(module, torch.nn.Linear)
             assert module.bias is None
 
+            # fuse transform into weight
             with torch.no_grad(), align_module_device(module):
                 update_offload_parameter(module, "weight", transform(module.weight))
 
@@ -128,6 +130,9 @@ def input_hook(_, args):
                     raise ValueError("Offloaded training is not supported")
                 P.register_parametrization(module, "weight", transform)
 
+            # transform is no longer needed (unfusing is not supported)
+            delete_offload_module(module, transform_name)
+
         # register output transformation hook
         elif args.location == TransformLocation.OUTPUT:
 
@@ -140,9 +145,6 @@ def output_hook(_, _input, output):
         else:
             raise NotImplementedError()
 
-        # (1) even in the `weight` cases, this submodule attachment is needed in order
-        # to support saving in the frozen state
-
 
 class TransformBase(Module, ABC):
     """

src/compressed_tensors/transform/factory/hadamard.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional
+from typing import Optional, Union
 
 import torch
 from compressed_tensors.transform import TransformArgs, TransformScheme
@@ -22,7 +22,7 @@
     apply_transform_weight,
     get_matrix_size,
 )
-from compressed_tensors.utils import get_offloaded_device
+from compressed_tensors.utils import get_execution_device, get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
 from torch import Tensor, device, dtype
 from torch.nn import Linear, Module, Parameter
@@ -41,6 +41,7 @@ class HadamardFactory(TransformFactory):
     def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
         super().__init__(name, scheme, seed)
         self.weights = ParameterizedDefaultDict(self._create_weight)
+        self.perms = ParameterizedDefaultDict(self._create_permutation)
 
     def create_transform(self, module: Module, args: TransformArgs):
         """
@@ -54,26 +55,46 @@ def create_transform(self, module: Module, args: TransformArgs):
         size = get_matrix_size(module, args.location)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
+        exec_device = get_execution_device(module)
 
-        weight = self.weights[size, dtype, device]
-        return HadamardTransform(weight, args)
+        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)
 
-    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        data = deterministic_hadamard_matrix(size, dtype, device)
-        data = data.to(dtype=dtype, device=device)
+    def _create_weight(
+        self,
+        size: int,
+        dtype: dtype,
+        device: device,
+        construct_device: device,
+    ) -> Parameter:
+        # construct on execution device, cache on offload device
+        data = deterministic_hadamard_matrix(size, dtype, construct_device)
+        data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 
+    def _create_permutation(self, weight: Parameter) -> Parameter:
+        data = torch.randperm(weight.size(0), generator=self.generator)
+        return Parameter(data, requires_grad=False)
+
 
 class HadamardTransform(TransformBase):
-    def __init__(self, weight: Parameter, args: TransformArgs):
+    def __init__(
+        self, weight: Parameter, perm: Union[Parameter, None], args: TransformArgs
+    ):
         super().__init__()
         self.weight = weight
+        self.perm = perm
         self.args = args
 
     def forward(self, value: Tensor) -> Tensor:
-        if not self.args.inverse:
-            weight = self.weight
-        else:
-            weight = self.weight.T
+        weight = self.weight
+
+        if self.perm is not None:
+            weight = weight[self.perm][:, self.perm]
+
+        if self.args.inverse:
+            weight = weight.T
 
         return apply_transform_weight(weight, value, self.args.location)

src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -62,6 +62,7 @@ def create_transform(self, module: Module, args: TransformArgs):
         return RandomMatrixTransform(weight, args)
 
     def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
+        # TODO: verify that weight is invertable (has non-zero determinant)
         data = torch.rand(
             (size, size), generator=self.generator, dtype=dtype, device=device
         )

src/compressed_tensors/transform/factory/random_hadamard.py

@@ -28,7 +28,14 @@ class RandomHadamardFactory(HadamardFactory):
     :param seed: random seed used to transform weight randomization
     """
 
-    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        data = random_hadamard_matrix(size, dtype, device, self.generator)
-        data = data.to(dtype=dtype, device=device)
+    def _create_weight(
+        self,
+        size: int,
+        dtype: dtype,
+        device: device,
+        construct_device: device,
+    ) -> Parameter:
+        # construct on execution device, cache on offload device
+        data = random_hadamard_matrix(size, dtype, construct_device, self.generator)
+        data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)

src/compressed_tensors/transform/transform_config.py

@@ -49,7 +49,7 @@ class TransformConfig(BaseModel):
                     inverse=True,
                 ),
             ],
-            randomize_modules=True,
+            randomize=True,
         ),
         "u": TransformScheme(
             type="hadamard",
@@ -62,7 +62,7 @@ class TransformConfig(BaseModel):
                     targets=["Linear"], location="output", inverse=True  # non-mergable
                 ),
             ],
-            randomize_modules=True,
+            randomize=True,
         ),
     }
 )

src/compressed_tensors/transform/transform_scheme.py

@@ -31,13 +31,12 @@ class TransformScheme(BaseModel):
         (see `Transforms.registered_names()`)
     :param apply: list of TransformationArgs containing the information about the
         modules that should be targeted by the specified transform
-    :param randomize_modules: True if unique transforms should be applied to each
-        unique module targeted by `apply`, otherwise reuse transform weights where
-        applicable
+    :param randomize: True if uniquely randomized transform weights should be used,
+        otherwise use identical transform weights where applicable
     :param requires_grad: True if weights include gradients for training
     """
 
     type: str
     apply: List[TransformArgs] = Field(default_factory=list)
-    randomize_modules: bool = Field(default=False)
+    randomize: bool = Field(default=False)
     requires_grad: bool = Field(default=False)

src/compressed_tensors/utils/helpers.py

@@ -15,10 +15,11 @@
 import contextlib
 import warnings
 from functools import wraps
-from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Mapping, Optional
 
 import numpy
 import torch
+from frozendict import frozendict
 from transformers import AutoConfig
 
 
@@ -373,11 +374,23 @@ class ParameterizedDefaultDict(dict):
 
     def __init__(self, default_factory: Callable[[Any], Any]):
         self.default_factory = default_factory
+        self._factory_kwargs = frozendict()
 
-    def __missing__(self, key):
+    def __missing__(self, key: Any) -> Any:
         if isinstance(key, tuple):
-            value = self.default_factory(*key)
+            value = self.default_factory(*key, **self._factory_kwargs)
         else:
-            value = self.default_factory(key)
+            value = self.default_factory(key, **self._factory_kwargs)
         self[key] = value
         return value
+
+    def get(self, *args, factory_kwargs: Mapping = frozendict()) -> Any:
+        """
+        Similar to `__getitem__`, but allows passing kwargs to factory function
+
+        :param \\*args: args whose tuple will value will be treated as key
+        :param factory_kwargs: keyword arguments to pass to `default_factory`
+        :return: dictionary entry for given key
+        """
+        with patch_attr(self, "_factory_kwargs", factory_kwargs):
+            return self[args]

tests/test_transform/factory/test_correctness.py

@@ -19,6 +19,7 @@
     TransformConfig,
     TransformFactory,
     TransformScheme,
+    apply_transform_config,
 )
 from compressed_tensors.utils import offloaded_dispatch
 from tests.testing_utils import requires_accelerate, requires_gpu
@@ -81,9 +82,7 @@ def test_correctness_model(scheme_kwargs, model_apply, offload=False):
             )
         }
     )
-    for name, scheme in config.config_groups.items():
-        factory = TransformFactory.from_scheme(scheme, name=name)
-        factory.apply_to_model(model)
+    apply_transform_config(model, config)
 
     # compare outputs
     output = model(input)

tests/test_transform/factory/test_memory.py

@@ -22,6 +22,7 @@
     TransformConfig,
     TransformFactory,
     TransformScheme,
+    apply_transform_config,
 )
 from compressed_tensors.utils import align_modules, offloaded_dispatch
 from tests.test_transform.conftest import TransformableModel
@@ -54,9 +55,7 @@ def test_memory_sharing(scheme_kwargs, offload=False):
             )
         }
     )
-    for name, scheme in config.config_groups.items():
-        factory = TransformFactory.from_scheme(scheme, name=name)
-        factory.apply_to_model(model)
+    apply_transform_config(model, config)
 
     # check that memory is shared when onloaded
     with align_modules(model.modules()):

tests/test_transform/test_transform_scheme.py

@@ -24,7 +24,7 @@ def test_basic_scheme():
         type="hadamard",
         apply=[basic_args],
     )
-    assert not scheme.randomize_modules
+    assert not scheme.randomize
     assert scheme.type == "hadamard"
     assert len(scheme.apply) == 1
     assert isinstance(scheme.apply[0], TransformArgs)
@@ -43,10 +43,10 @@ def test_multiple_groups_global():
     scheme = TransformScheme(
         type="hadamard",
         apply=[embedding_args, linear_args],
-        randomize_modules=True,
+        randomize=True,
     )
 
-    assert scheme.randomize_modules
+    assert scheme.randomize
     assert scheme.type == "hadamard"
     assert len(scheme.apply) == 2
     assert isinstance(scheme.apply[0], TransformArgs)
@@ -69,6 +69,6 @@ def test_multiple_groups():
         apply=apply,
     )
 
-    assert not scheme.randomize_modules
+    assert not scheme.randomize
     assert scheme.type == "hadamard"
     assert len(scheme.apply) == 20