implement apply, use in tests

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

Author: kylesayrs

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 +1,25 @@
+# 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):
+    for name, scheme in config.config_groups.items():
+        factory = TransformFactory.from_scheme(scheme, name=name)
+        factory.apply_to_model(model)

tests/test_transform/conftest.py

@@ -0,0 +1,52 @@
+# 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([])
+        self.fcs.append(torch.nn.Linear(sizes[0], sizes[1], bias=False))
+        for index in range(1, len(sizes) - 1):
+            self.fcs.append(torch.nn.Linear(sizes[index], sizes[index + 1], bias=False))
+
+    def forward(self, x):
+        for layer in self.fcs:
+            x = layer(x)
+        return x
+
+
+@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

tests/test_transform/factory/test_correctness.py

@@ -16,38 +16,27 @@
 import torch
 from compressed_tensors.transform import (
     TransformArgs,
+    TransformConfig,
     TransformFactory,
     TransformScheme,
+    apply_transform_config,
 )
 from compressed_tensors.utils import force_cpu_offload
 from tests.testing_utils import requires_accelerate, requires_gpu
 
 
-def all_schemes():
+def scheme_kwargs():
     all_types = TransformFactory.registered_names()
-    base = [TransformScheme(type=type) for type in all_types]
-    randomized = [TransformScheme(type=type, randomize=True) for type in all_types]
+    base = [{"type": type} for type in all_types]
+    randomized = [{"type": type, "randomize": True} for type in all_types]
     return base + randomized
 
 
-class TransformableModel(torch.nn.Module):
-    def __init__(self, *sizes):
-        super().__init__()
-        self.fcs = torch.nn.ModuleList([])
-        self.fcs.append(torch.nn.Linear(sizes[0], sizes[1], bias=False))
-        for index in range(1, len(sizes) - 1):
-            self.fcs.append(torch.nn.Linear(sizes[index], sizes[index + 1], bias=False))
-
-    def forward(self, x):
-        for layer in self.fcs:
-            x = layer(x)
-        return x
-
-
-@pytest.mark.parametrize("scheme", all_schemes())
-def test_correctness_linear(scheme):
+@pytest.mark.parametrize("scheme_kwargs", scheme_kwargs())
+def test_correctness_linear(scheme_kwargs):
     size = (4, 8)
     module = torch.nn.Linear(*size, bias=True)
+    scheme = TransformScheme(**scheme_kwargs)
     factory = TransformFactory.from_scheme(scheme, name="")
 
     input_tfm = factory.create_transform(
@@ -71,44 +60,37 @@ def test_correctness_linear(scheme):
     assert torch.allclose(true_output, output, atol=1e-5, rtol=0.0)
 
 
-@pytest.mark.parametrize("scheme", all_schemes())
-def test_correctness_model(scheme, offload=False):
+@pytest.mark.parametrize("scheme_kwargs", scheme_kwargs())
+def test_correctness_model(scheme_kwargs, model_apply, offload=False):
     # load model
-    model = TransformableModel(2, 4, 8, 16, 32, 64)
+    model = model_apply[0]
     if offload:
         model = force_cpu_offload(model, torch.device("cuda"))
 
-    # create factory
-    scheme.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),
-    ]
-    factory = TransformFactory.from_scheme(scheme, name="")
-
-    # create inputs
+    # get output
     input = torch.rand((17, model.fcs[0].in_features))
     if offload:
         input = input.to(torch.device("cuda"))
+    true_output = model(input)
+
+    # apply transforms
+    config = TransformConfig(
+        config_groups={
+            "": TransformScheme(
+                **scheme_kwargs,
+                apply=model_apply[1],
+            )
+        }
+    )
+    apply_transform_config(model, config)
 
     # compare outputs
-    true_output = model(input)
-    factory.apply_to_model(model)
     output = model(input)
     assert torch.allclose(true_output, output, atol=1e-5, rtol=0.0)
 
 
 @requires_gpu
 @requires_accelerate()
-@pytest.mark.parametrize("scheme", all_schemes())
-def test_correctness_model_offload(scheme):
-    test_correctness_model(scheme, offload=True)
+@pytest.mark.parametrize("scheme_kwargs", scheme_kwargs())
+def test_correctness_model_offload(scheme_kwargs, model_apply):
+    test_correctness_model(scheme_kwargs, model_apply, offload=True)

tests/test_transform/factory/test_memory.py

@@ -19,53 +19,43 @@
 from compressed_tensors.transform import (
     TransformArgs,
     TransformBase,
+    TransformConfig,
     TransformFactory,
     TransformScheme,
+    apply_transform_config,
 )
 from compressed_tensors.utils import align_modules, force_cpu_offload
+from tests.test_transform.conftest import TransformableModel
 from tests.testing_utils import requires_accelerate, requires_gpu
 
 
-def all_schemes():
+def scheme_kwargs():
     all_types = TransformFactory.registered_names()
-    base = [TransformScheme(type=type) for type in all_types]
-    randomized = [TransformScheme(type=type, randomize=True) for type in all_types]
+    base = [{"type": type} for type in all_types]
+    randomized = [{"type": type, "randomize": True} for type in all_types]
     return base + randomized
 
 
-class TransformableModel(torch.nn.Module):
-    def __init__(self, *sizes):
-        super().__init__()
-        self.fcs = torch.nn.ModuleList([])
-        self.fcs.append(torch.nn.Linear(sizes[0], sizes[1], bias=False))
-        for index in range(1, len(sizes) - 1):
-            self.fcs.append(torch.nn.Linear(sizes[index], sizes[index + 1], bias=False))
-
-    def forward(self, x):
-        for layer in self.fcs:
-            x = layer(x)
-        return x
-
-
-@pytest.mark.parametrize("scheme", all_schemes())
-def test_memory_sharing(scheme, offload=False):
-    # load scheme and factory
-    scheme = TransformScheme(
-        type="hadamard",
-        apply=[
-            TransformArgs(targets="Linear", location="input"),
-            TransformArgs(targets="Linear", location="output"),
-        ],
-    )
-    factory = TransformFactory.from_scheme(scheme, name="")
-
+@pytest.mark.parametrize("scheme_kwargs", scheme_kwargs())
+def test_memory_sharing(scheme_kwargs, offload=False):
     # load model (maybe with offloading)
     model = TransformableModel(2, 2, 4, 4, 8, 8)
     if offload:
         force_cpu_offload(model, torch.device("cuda"))
 
     # add transforms to model
-    factory.apply_to_model(model)
+    config = TransformConfig(
+        config_groups={
+            "": TransformScheme(
+                **scheme_kwargs,
+                apply=[
+                    TransformArgs(targets="Linear", location="input"),
+                    TransformArgs(targets="Linear", location="output"),
+                ],
+            )
+        }
+    )
+    apply_transform_config(model, config)
 
     # check that memory is shared when onloaded
     with align_modules(model.modules()):
@@ -97,12 +87,12 @@ def test_memory_sharing(scheme, offload=False):
 
 @requires_gpu
 @requires_accelerate()
-@pytest.mark.parametrize("scheme", all_schemes())
-def test_memory_sharing_offload(scheme):
-    test_memory_sharing(scheme, offload=True)
+@pytest.mark.parametrize("scheme_kwargs", scheme_kwargs())
+def test_memory_sharing_offload(scheme_kwargs):
+    test_memory_sharing(scheme_kwargs, offload=True)
 
 
-@pytest.mark.parametrize("scheme", all_schemes())
-def test_memory_sharing_training(scheme):
-    scheme.requires_grad = True
-    test_memory_sharing(scheme, offload=False)
+@pytest.mark.parametrize("scheme_kwargs", scheme_kwargs())
+def test_memory_sharing_training(scheme_kwargs):
+    scheme_kwargs["requires_grad"] = True
+    test_memory_sharing(scheme_kwargs, offload=False)