[Transform] Serialize with tied weights (neuralmagic#370)

* add utilities

Signed-off-by: Kyle Sayers <[email protected]>

* add tests

Signed-off-by: Kyle Sayers <[email protected]>

* add additional tests

Signed-off-by: Kyle Sayers <[email protected]>

* add utils and tests

Signed-off-by: Kyle Sayers <[email protected]>

* Implement transform factories

Signed-off-by: Kyle Sayers <[email protected]>

* add permutations

Signed-off-by: Kyle Sayers <[email protected]>

* add delete_offload_module

Signed-off-by: Kyle Sayers <[email protected]>

* key inverses by weight

Signed-off-by: Kyle Sayers <[email protected]>

* fix tests

Signed-off-by: Kyle Sayers <[email protected]>

* standardize random hadamard

Signed-off-by: Kyle Sayers <[email protected]>

* prepend input hooks

Signed-off-by: Kyle Sayers <[email protected]>

* apply sqrt division first

Signed-off-by: Kyle Sayers <[email protected]>

* use divided hadamards

Signed-off-by: Kyle Sayers <[email protected]>

* fix typo

Signed-off-by: Kyle Sayers <[email protected]>

* add random option

Signed-off-by: Kyle Sayers <[email protected]>

* use random seeds, rename matrix multiply

Signed-off-by: Kyle Sayers <[email protected]>

* add deterministic generation to random matrix

Signed-off-by: Kyle Sayers <[email protected]>

* fix perm math

Signed-off-by: Kyle Sayers <[email protected]>

* update docstrings

Signed-off-by: Kyle Sayers <[email protected]>

* update docstrings

Signed-off-by: Kyle Sayers <[email protected]>

* cleanup

Signed-off-by: Kyle Sayers <[email protected]>

* cleanup 2

Signed-off-by: Kyle Sayers <[email protected]>

* make seed optional

Signed-off-by: Kyle Sayers <[email protected]>

* remove iterable check and missing return value

Signed-off-by: Kyle Sayers <[email protected]>

* Remove unrelated changes

* simplify code

Signed-off-by: Kyle Sayers <[email protected]>

* implement apply, use in tests

Signed-off-by: Kyle Sayers <[email protected]>

* use hadamards database file

Signed-off-by: Kyle Sayers <[email protected]>

* try manifest

Signed-off-by: Kyle Sayers <[email protected]>

* try setup, update hadamards list

Signed-off-by: Kyle Sayers <[email protected]>

* fix setup

Signed-off-by: Kyle Sayers <[email protected]>

* add docstrings, cleanup

Signed-off-by: Kyle Sayers <[email protected]>

* fix setup, thank you @dbarbuzzi

Signed-off-by: Kyle Sayers <[email protected]>

* remove numpy, add tests

Signed-off-by: Kyle Sayers <[email protected]>

* solidify dtype, add gpu tests

Signed-off-by: Kyle Sayers <[email protected]>

* fix docstring

Signed-off-by: Kyle Sayers <[email protected]>

* add device option

Signed-off-by: Kyle Sayers <[email protected]>

* construct on execution device, cache on offload device

Signed-off-by: Kyle Sayers <[email protected]>

* save construction device changes for later

Signed-off-by: Kyle Sayers <[email protected]>

* construct on execution device, cache on offload device

* cite nja sloane

Signed-off-by: Kyle Sayers <[email protected]>

* remove dreg

Signed-off-by: Kyle Sayers <[email protected]>

* put on device via safe_open

Signed-off-by: Kyle Sayers <[email protected]>

* nits and docstrings

Signed-off-by: Kyle Sayers <[email protected]>

* update docstring

Signed-off-by: Kyle Sayers <[email protected]>

* Merge

* merge with construct: construct in float32

Signed-off-by: Kyle Sayers <[email protected]>

* construct with same dtype, constructing on fp32 found no difference

Signed-off-by: Kyle Sayers <[email protected]>

* remove unnecessary imports

Signed-off-by: Kyle Sayers <[email protected]>

* bugfixes (neuralmagic#375)

Signed-off-by: Brian Dellabetta <[email protected]>

* use factory_kwargs

Signed-off-by: Kyle Sayers <[email protected]>

* add frozen dict to deps

Signed-off-by: Kyle Sayers <[email protected]>

* fix style

Signed-off-by: Kyle Sayers <[email protected]>

* merge

Signed-off-by: Kyle Sayers <[email protected]>

* use delete_offload_module

Signed-off-by: Kyle Sayers <[email protected]>

* add docstrign

Signed-off-by: Kyle Sayers <[email protected]>

* use parametrize

Signed-off-by: Kyle Sayers <[email protected]>

* populate _dynamic_tied_weights_keys

Signed-off-by: Kyle Sayers <[email protected]>

* ensure serializable

Signed-off-by: Kyle Sayers <[email protected]>

* remove extra space

Signed-off-by: Kyle Sayers <[email protected]>

* apply style

Signed-off-by: Kyle Sayers <[email protected]>

* merge dregs

* skip offloading tests until transformers changes land

Signed-off-by: Kyle Sayers <[email protected]>

* use set

Signed-off-by: Kyle Sayers <[email protected]>

---------

Signed-off-by: Kyle Sayers <[email protected]>
Signed-off-by: Brian Dellabetta <[email protected]>
Co-authored-by: Brian Dellabetta <[email protected]>

Author: Etelis

src/compressed_tensors/transform/factory/base.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 
 from abc import ABC, abstractmethod
-from typing import Optional
+from collections import defaultdict
+from typing import List, Optional, Tuple, Set
 
 import torch
 import torch.nn.utils.parametrize as P
@@ -49,10 +50,13 @@ class TransformFactory(RegistryMixin, ABC):
     :param seed: random seed used to transform weight randomization
     """
 
+    transforms: List["TransformBase"]
+
     def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
         self.name = name
         self.scheme = scheme
         self.generator = torch.Generator()
+        self.transforms = list()
         if seed is not None:
             self.generator.manual_seed(seed)
 
@@ -90,16 +94,26 @@ def apply_to_model(self, model: Module):
             for _, module in match_named_modules(model, arg.targets, arg.ignore):
                 self._apply_to_module(module, arg)
 
+        self._update_tied_weights()
+
     def _apply_to_module(self, module: Module, args: TransformArgs):
         """
         Create transforms and apply them to the module
 
         :param module: target module to apply transforms to
         :param args: defines how the transform will be applied to the target module
         """
+        if has_offloaded_params(module):
+            if module._hf_hook.place_submodules:
+                raise NotImplementedError(
+                    "Applying transforms to offloaded submodules with "
+                    "`place_submodules=True` is not supported"
+                )
+
         # create transform as submodule
         transform_name = f"{self.name}_{args.location}"
         transform = self.create_transform(module, args)
+        self.transforms.append(transform)
         register_offload_module(module, transform_name, transform)
 
         # register input transformation hook
@@ -128,8 +142,9 @@ 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)
+            else:
+                # transform is no longer needed (unfusing is not supported)
+                delete_offload_module(module, transform_name)
 
         # register output transformation hook
         elif args.location == TransformLocation.OUTPUT:
@@ -143,6 +158,31 @@ def output_hook(_, _input, output):
         else:
             raise NotImplementedError()
 
+    def _update_tied_weights(self):
+        """
+        Populate the `_dynamic_tied_weights_keys` attribute of transforms,
+        which is used by transformers to detect and remove shared pointers
+        during saving
+        """
+        # map from data_ptrs to keys
+        ptr_to_keys: dict[int, List[Tuple[TransformBase, str]]] = defaultdict(list)
+        for transform in self.transforms:
+            for name, param in transform.named_parameters(recurse=False):
+                # NOTE: previously asserted that parent._hf_hook.place_submodules=False
+                if has_offloaded_params(transform):
+                    param = transform._hf_hook.weights_map[name]
+                ptr_to_keys[param.data_ptr()].append((transform, name))
+
+        # populate `_dynamic_tied_weights_keys` if there is more than one key
+        # and ensure that they share tensors
+        for shared_keys in ptr_to_keys.values():
+            if len(shared_keys) > 1:
+                tensor = getattr(shared_keys[0][0], shared_keys[0][1])
+
+                for transform, name in shared_keys:
+                    transform._dynamic_tied_weights_keys.add(name)
+                    setattr(transform, name, tensor)
+
 
 class TransformBase(InternalModule, ABC):
     """
@@ -151,6 +191,11 @@ class TransformBase(InternalModule, ABC):
 
     args: TransformArgs
     weight: Parameter
+    _dynamic_tied_weights_keys: Set[str]
+
+    def __init__(self):
+        super().__init__()
+        self._dynamic_tied_weights_keys = set()
 
     @abstractmethod
     def forward(self, value: Tensor) -> Tensor:

src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -70,6 +70,7 @@ def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
 
     def _create_inverse(self, weight: Parameter) -> Parameter:
         data = high_precision_invert(weight.data)
+        data = data.contiguous()  # ensure proper serialization
         return Parameter(data, requires_grad=False)
 
 

tests/test_transform/conftest.py

@@ -14,12 +14,13 @@
 
 import pytest
 import torch
-from compressed_tensors.transform import TransformArgs
+from compressed_tensors.transform import TransformArgs, TransformFactory
+from transformers import PretrainedConfig, PreTrainedModel
 
 
-class TransformableModel(torch.nn.Module):
+class TransformableModel(PreTrainedModel):
     def __init__(self, *sizes):
-        super().__init__()
+        super().__init__(config=PretrainedConfig())
         self.fcs = torch.nn.ModuleList(
             [
                 torch.nn.Linear(sizes[index], sizes[index + 1], bias=False)

tests/test_transform/factory/test_correctness.py

@@ -27,13 +27,13 @@
 
 
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (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):
+def test_correctness_linear(type, randomize, 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)
+    scheme = TransformScheme(type=type, randomize=randomize, head_dim=head_dim)
     factory = TransformFactory.from_scheme(scheme, name="")
 
     input_tfm = factory.create_transform(
@@ -58,10 +58,10 @@ def test_correctness_linear(type, randomized, head_dim, input_batch_size):
 
 
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (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):
+def test_correctness_embedding(type, randomize, embed_loc, linear_loc):
     model = torch.nn.Sequential(
         torch.nn.Embedding(2, 4),
         torch.nn.Linear(4, 8, bias=False),
@@ -74,7 +74,7 @@ def test_correctness_embedding(type, randomized, embed_loc, linear_loc):
         config_groups={
             "": TransformScheme(
                 type=type,
-                randomized=randomized,
+                randomize=randomize,
                 apply=[
                     TransformArgs(targets="Embedding", location=embed_loc),
                     TransformArgs(targets="Linear", location=linear_loc, inverse=True),
@@ -90,10 +90,10 @@ def test_correctness_embedding(type, randomized, embed_loc, linear_loc):
 
 
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (True, False))
 @pytest.mark.parametrize("input_batch_size", (1, 5, 17))
 def test_correctness_model(
-    type, randomized, input_batch_size, model_apply, offload=False
+    type, randomize, input_batch_size, model_apply, offload=False
 ):
     # load model
     model = model_apply[0]
@@ -109,7 +109,7 @@ def test_correctness_model(
     # apply transforms
     config = TransformConfig(
         config_groups={
-            "": TransformScheme(type=type, randomized=randomized, apply=model_apply[1])
+            "": TransformScheme(type=type, randomize=randomize, apply=model_apply[1])
         }
     )
     apply_transform_config(model, config)
@@ -122,19 +122,17 @@ def test_correctness_model(
 @requires_gpu
 @requires_accelerate()
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (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
-    )
+def test_correctness_model_offload(type, randomize, input_batch_size, model_apply):
+    test_correctness_model(type, randomize, input_batch_size, model_apply, offload=True)
 
 
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (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):
+def test_correctness_attention_heads(type, randomize, head_dim, input_batch_size):
     hidden_size = 64
     num_attention_heads = 8
 
@@ -151,7 +149,7 @@ def test_correctness_attention_heads(type, randomized, head_dim, input_batch_siz
         config_groups={
             "": TransformScheme(
                 type=type,
-                randomized=randomized,
+                randomize=randomize,
                 head_dim=head_dim,
                 apply=[
                     TransformArgs(targets="v_proj", location="weight_output"),

tests/test_transform/factory/test_memory.py

@@ -29,9 +29,9 @@
 
 
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (True, False))
 @pytest.mark.parametrize("requires_grad", (True, False))
-def test_memory_sharing(type, randomized, requires_grad, offload=False):
+def test_memory_sharing(type, randomize, requires_grad, offload=False):
     # load model (maybe with offloading)
     model = TransformableModel(2, 2, 4, 4, 8, 8)
     if offload:
@@ -42,7 +42,7 @@ def test_memory_sharing(type, randomized, requires_grad, offload=False):
         config_groups={
             "": TransformScheme(
                 type=type,
-                randomzied=randomized,
+                randomzied=randomize,
                 requires_grad=requires_grad,
                 apply=[
                     TransformArgs(targets="Linear", location="input"),
@@ -84,9 +84,9 @@ def test_memory_sharing(type, randomized, requires_grad, offload=False):
 @requires_gpu
 @requires_accelerate()
 @pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
-@pytest.mark.parametrize("randomized", (True, False))
+@pytest.mark.parametrize("randomize", (True, False))
 def test_memory_sharing_offload(
     type,
-    randomized,
+    randomize,
 ):
-    test_memory_sharing(type, randomized, requires_grad=False, offload=True)
+    test_memory_sharing(type, randomize, requires_grad=False, offload=True)

tests/test_transform/factory/test_serialization.py

@@ -0,0 +1,54 @@
+# 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 (
+    TransformConfig,
+    TransformScheme,
+    apply_transform_config,
+)
+from compressed_tensors.utils import offloaded_dispatch
+from tests.testing_utils import requires_accelerate, requires_gpu
+
+
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomize", (True, False))
+def test_serialization(type, randomize, model_apply, tmp_path, offload=False):
+    # get model, maybe offload
+    model, apply = model_apply
+    if offload:
+        offloaded_dispatch(model, torch.device("cuda"))
+
+    # apply transforms to model
+    config = TransformConfig(
+        config_groups={"": TransformScheme(type=type, randomize=randomize, apply=apply)}
+    )
+    apply_transform_config(model, config)
+
+    # save model
+    model.save_pretrained(tmp_path)
+
+    # TODO: reload model
+
+
+@pytest.mark.skip(reason="Requires changes in upstream transformers")
+# https://github.com/huggingface/transformers/pull/39280
+# https://github.com/huggingface/transformers/pull/39263
+@requires_gpu
+@requires_accelerate()
+@pytest.mark.parametrize("type", ("hadamard", "random-hadamard"))
+@pytest.mark.parametrize("randomize", (True, False))
+def test_serialization_offload(type, randomize, model_apply, tmp_path):
+    test_serialization(type, randomize, model_apply, tmp_path, offload=True)