Simplify dataparallel approach (#191)

.vscode/cspell.json

@@ -51,6 +51,7 @@
         "jaxtyping",
         "kaiming",
         "keepdim",
+        "logit",
         "lognormal",
         "loguniform",
         "loguniformvalues",
@@ -73,6 +74,7 @@
         "neox",
         "nonlinerity",
         "numel",
+        "onebit",
         "openwebtext",
         "optim",
         "penality",
@@ -116,6 +118,7 @@
         "venv",
         "virtualenv",
         "virtualenvs",
-        "wandb"
+        "wandb",
+        "zoadam"
     ]
 }

pyproject.toml

@@ -31,7 +31,7 @@
             pytest-integration=">=0.2.3"
             pytest-timeout=">=2.2.0"
             pytest-xdist="^3.5.0"
-            ruff=">=0.1.4"
+            ruff=">=0.1.14"
             syrupy=">=4.6.0"
 
         [tool.poetry.group.demos.dependencies]

sparse_autoencoder/activation_resampler/activation_resampler.py

@@ -7,6 +7,7 @@
 from pydantic import Field, NonNegativeInt, PositiveInt, validate_call
 import torch
 from torch import Tensor
+from torch.nn.parallel import DataParallel
 from torch.utils.data import DataLoader
 
 from sparse_autoencoder.activation_resampler.utils.component_slice_tensor import (
@@ -17,7 +18,6 @@
 from sparse_autoencoder.loss.abstract_loss import AbstractLoss
 from sparse_autoencoder.tensor_types import Axis
 from sparse_autoencoder.train.utils.get_model_device import get_model_device
-from sparse_autoencoder.utils.data_parallel import DataParallelWithModelAttributes
 
 
 @dataclass
@@ -207,7 +207,7 @@ def _get_dead_neuron_indices(
     def compute_loss_and_get_activations(
         self,
         store: ActivationStore,
-        autoencoder: SparseAutoencoder | DataParallelWithModelAttributes[SparseAutoencoder],
+        autoencoder: SparseAutoencoder | DataParallel[SparseAutoencoder],
         loss_fn: AbstractLoss,
         train_batch_size: int,
     ) -> LossInputActivationsTuple:
@@ -440,7 +440,7 @@ def renormalize_and_scale(
     def resample_dead_neurons(
         self,
         activation_store: ActivationStore,
-        autoencoder: SparseAutoencoder | DataParallelWithModelAttributes[SparseAutoencoder],
+        autoencoder: SparseAutoencoder | DataParallel[SparseAutoencoder],
         loss_fn: AbstractLoss,
         train_batch_size: int,
     ) -> list[ParameterUpdateResults]:
@@ -530,7 +530,7 @@ def step_resampler(
         self,
         batch_neuron_activity: Int64[Tensor, Axis.names(Axis.COMPONENT, Axis.LEARNT_FEATURE)],
         activation_store: ActivationStore,
-        autoencoder: SparseAutoencoder | DataParallelWithModelAttributes[SparseAutoencoder],
+        autoencoder: SparseAutoencoder | DataParallel[SparseAutoencoder],
         loss_fn: AbstractLoss,
         train_batch_size: int,
     ) -> list[ParameterUpdateResults] | None:

sparse_autoencoder/autoencoder/model.py

@@ -276,7 +276,7 @@ def load(
             The loaded model.
         """
         # Load the file
-        serialized_state = torch.load(file_path)
+        serialized_state = torch.load(file_path, map_location=torch.device("cpu"))
         state = SparseAutoencoderState.model_validate(serialized_state)
 
         # Initialise the model

sparse_autoencoder/metrics/validate/tests/test_model_reconstruction_score.py

@@ -1,15 +1,13 @@
 """Test the model reconstruction score metric."""
 
-from jaxtyping import Float
 import pytest
 from syrupy.session import SnapshotSession
 import torch
-from torch import Tensor
+from torch import Tensor, tensor
 
 from sparse_autoencoder.metrics.utils.find_metric_result import find_metric_result
 from sparse_autoencoder.metrics.validate.abstract_validate_metric import ValidationMetricData
 from sparse_autoencoder.metrics.validate.model_reconstruction_score import ModelReconstructionScore
-from sparse_autoencoder.tensor_types import Axis
 
 
 def test_model_reconstruction_score_empty_data() -> None:
@@ -19,9 +17,9 @@ def test_model_reconstruction_score_empty_data() -> None:
     is provided (i.e., at the end of training or in similar scenarios).
     """
     data = ValidationMetricData(
-        source_model_loss=Float[Tensor, Axis.ITEMS]([]),
-        source_model_loss_with_reconstruction=Float[Tensor, Axis.ITEMS]([]),
-        source_model_loss_with_zero_ablation=Float[Tensor, Axis.ITEMS]([]),
+        source_model_loss=tensor([]),
+        source_model_loss_with_reconstruction=tensor([]),
+        source_model_loss_with_zero_ablation=tensor([]),
     )
     metric = ModelReconstructionScore()
     result = metric.calculate(data)
@@ -41,13 +39,9 @@ def test_model_reconstruction_score_empty_data() -> None:
         ),
         (
             ValidationMetricData(
-                source_model_loss=Float[Tensor, Axis.ITEMS]([[0.5], [1.5], [2.5]]),
-                source_model_loss_with_reconstruction=Float[Tensor, Axis.ITEMS](
-                    [[1.5], [2.5], [3.5]]
-                ),
-                source_model_loss_with_zero_ablation=Float[Tensor, Axis.ITEMS](
-                    [[8.0], [7.0], [4.0]]
-                ),
+                source_model_loss=tensor([[0.5], [1.5], [2.5]]),
+                source_model_loss_with_reconstruction=tensor([[1.5], [2.5], [3.5]]),
+                source_model_loss_with_zero_ablation=tensor([[8.0], [7.0], [4.0]]),
             ),
             0.79,
         ),

sparse_autoencoder/optimizer/adam_with_reset.py

@@ -42,7 +42,7 @@ def __init__(  # (extending existing implementation)
         lr: float | Float[Tensor, Axis.names(Axis.SINGLE_ITEM)] = 1e-3,
         betas: tuple[float, float] = (0.9, 0.999),
         eps: float = 1e-8,
-        weight_decay: float = 0,
+        weight_decay: float = 0.0,
         *,
         amsgrad: bool = False,
         foreach: bool | None = None,

sparse_autoencoder/source_model/replace_activations_hook.py

@@ -1,33 +1,36 @@
 """Replace activations hook."""
 from typing import TYPE_CHECKING
 
+from jaxtyping import Float
 from torch import Tensor
+from torch.nn.parallel import DataParallel
 from transformer_lens.hook_points import HookPoint
 
 from sparse_autoencoder.autoencoder.model import SparseAutoencoder
-from sparse_autoencoder.utils.data_parallel import DataParallelWithModelAttributes
 
 
 if TYPE_CHECKING:
     from sparse_autoencoder.tensor_types import Axis
-from jaxtyping import Float
 
 
 def replace_activations_hook(
     value: Tensor,
     hook: HookPoint,  # noqa: ARG001
-    sparse_autoencoder: SparseAutoencoder | DataParallelWithModelAttributes[SparseAutoencoder],
+    sparse_autoencoder: SparseAutoencoder | DataParallel[SparseAutoencoder],
     component_idx: int | None = None,
+    n_components: int | None = None,
 ) -> Tensor:
     """Replace activations hook.
 
+    This should be pre-initialised with `functools.partial`.
+
     Args:
         value: The activations to replace.
         hook: The hook point.
-        sparse_autoencoder: The sparse autoencoder. This should be pre-initialised with
-            `functools.partial`.
+        sparse_autoencoder: The sparse autoencoder.
         component_idx: The component index to replace the activations with, if just replacing
             activations for a single component. Requires the model to have a component axis.
+        n_components: The number of components that the SAE is trained on.
 
     Returns:
         Replaced activations.
@@ -43,11 +46,8 @@ def replace_activations_hook(
     )
 
     if component_idx is not None:
-        if sparse_autoencoder.config.n_components is None:
-            error_message = (
-                "Cannot replace for a specific component, if the model does not have a "
-                "component axis."
-            )
+        if n_components is None:
+            error_message = "The number of model components must be set if component_idx is set."
             raise RuntimeError(error_message)
 
         # The approach here is to run a forward pass with dummy values for all components other than
@@ -56,7 +56,7 @@ def replace_activations_hook(
         # components.
         expanded_shape = [
             squashed_value.shape[0],
-            sparse_autoencoder.config.n_components,
+            n_components,
             squashed_value.shape[-1],
         ]
         expanded = squashed_value.unsqueeze(1).expand(*expanded_shape)

sparse_autoencoder/source_model/tests/test_replace_activations_hook.py

@@ -66,7 +66,12 @@ def test_hook_replaces_activations_2_components() -> None:
         fwd_hooks=[
             (
                 "blocks.0.hook_mlp_out",
-                partial(replace_activations_hook, sparse_autoencoder=autoencoder, component_idx=1),
+                partial(
+                    replace_activations_hook,
+                    sparse_autoencoder=autoencoder,
+                    component_idx=1,
+                    n_components=2,
+                ),
             )
         ],
     )

sparse_autoencoder/train/pipeline.py

@@ -9,6 +9,7 @@
 from pydantic import NonNegativeInt, PositiveInt, validate_call
 import torch
 from torch import Tensor
+from torch.nn.parallel import DataParallel
 from torch.optim.lr_scheduler import LRScheduler
 from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
@@ -32,7 +33,6 @@
 from sparse_autoencoder.source_model.zero_ablate_hook import zero_ablate_hook
 from sparse_autoencoder.tensor_types import Axis
 from sparse_autoencoder.train.utils.get_model_device import get_model_device
-from sparse_autoencoder.utils.data_parallel import DataParallelWithModelAttributes
 
 
 if TYPE_CHECKING:
@@ -51,9 +51,15 @@ class Pipeline:
     activation_resampler: ActivationResampler | None
     """Activation resampler to use."""
 
-    autoencoder: SparseAutoencoder | DataParallelWithModelAttributes[SparseAutoencoder]
+    autoencoder: SparseAutoencoder | DataParallel[SparseAutoencoder]
     """Sparse autoencoder to train."""
 
+    n_input_features: int
+    """Number of input features in the sparse autoencoder."""
+
+    n_learned_features: int
+    """Number of learned features in the sparse autoencoder."""
+
     cache_names: list[str]
     """Names of the cache hook points to use in the source model."""
 
@@ -81,7 +87,7 @@ class Pipeline:
     source_dataset: SourceDataset
     """Source dataset to generate activation data from (tokenized prompts)."""
 
-    source_model: HookedTransformer | DataParallelWithModelAttributes[HookedTransformer]
+    source_model: HookedTransformer | DataParallel[HookedTransformer]
     """Source model to get activations from."""
 
     total_activations_trained_on: int = 0
@@ -97,13 +103,15 @@ def n_components(self) -> int:
     def __init__(
         self,
         activation_resampler: ActivationResampler | None,
-        autoencoder: SparseAutoencoder | DataParallelWithModelAttributes[SparseAutoencoder],
+        autoencoder: SparseAutoencoder | DataParallel[SparseAutoencoder],
         cache_names: list[str],
         layer: NonNegativeInt,
         loss: AbstractLoss,
         optimizer: AbstractOptimizerWithReset,
         source_dataset: SourceDataset,
-        source_model: HookedTransformer | DataParallelWithModelAttributes[HookedTransformer],
+        source_model: HookedTransformer | DataParallel[HookedTransformer],
+        n_input_features: int,
+        n_learned_features: int,
         run_name: str = "sparse_autoencoder",
         checkpoint_directory: Path = DEFAULT_CHECKPOINT_DIRECTORY,
         lr_scheduler: LRScheduler | None = None,
@@ -124,6 +132,8 @@ def __init__(
             optimizer: Optimizer to use.
             source_dataset: Source dataset to get data from.
             source_model: Source model to get activations from.
+            n_input_features: Number of input features in the sparse autoencoder.
+            n_learned_features: Number of learned features in the sparse autoencoder.
             run_name: Name of the run for saving checkpoints.
             checkpoint_directory: Directory to save checkpoints to.
             lr_scheduler: Learning rate scheduler to use.
@@ -146,6 +156,8 @@ def __init__(
         self.source_data_batch_size = source_data_batch_size
         self.source_dataset = source_dataset
         self.source_model = source_model
+        self.n_input_features = n_input_features
+        self.n_learned_features = n_learned_features
 
         # Create a stateful iterator
         source_dataloader = source_dataset.get_dataloader(
@@ -175,9 +187,10 @@ def generate_activations(self, store_size: PositiveInt) -> TensorActivationStore
             raise ValueError(error_message)
 
         # Setup the store
-        n_neurons: int = self.autoencoder.config.n_input_features
         source_model_device: torch.device = get_model_device(self.source_model)
-        store = TensorActivationStore(store_size, n_neurons, n_components=self.n_components)
+        store = TensorActivationStore(
+            store_size, self.n_input_features, n_components=self.n_components
+        )
 
         # Add the hook to the model (will automatically store the activations every time the model
         # runs)
@@ -225,9 +238,9 @@ def train_autoencoder(
         learned_activations_fired_count: Int64[
             Tensor, Axis.names(Axis.COMPONENT, Axis.LEARNT_FEATURE)
         ] = torch.zeros(
-            (self.n_components, self.autoencoder.config.n_learned_features),
+            (self.n_components, self.n_learned_features),
             dtype=torch.int64,
-            device=autoencoder_device,
+            device=torch.device("cpu"),
         )
 
         for store_batch in activations_dataloader:
@@ -260,7 +273,7 @@ def train_autoencoder(
             # Store count of how many neurons have fired
             with torch.no_grad():
                 fired = learned_activations > 0
-                learned_activations_fired_count.add_(fired.sum(dim=0))
+                learned_activations_fired_count.add_(fired.sum(dim=0).cpu())
 
             # Backwards pass
             total_loss.backward()
@@ -358,6 +371,7 @@ def validate_sae(self, validation_n_activations: PositiveInt) -> None:
                     replace_activations_hook,
                     sparse_autoencoder=self.autoencoder,
                     component_idx=component_idx,
+                    n_components=self.n_components,
                 )
 
                 with torch.no_grad():