Add DLRM-V2 with sparsecore.

PiperOrigin-RevId: 745051634

Author: Hilly12

recml/core/data/iterator.py

@@ -57,16 +57,15 @@ def __next__(self) -> clu_data.Element:
     if self._prefetched_batch is not None:
       batch = self._prefetched_batch
       self._prefetched_batch = None
-      return batch
-
-    batch = next(self._iterator)
-    if self._postprocessor is not None:
-      batch = self._postprocessor(batch)
+    else:
+      batch = next(self._iterator)
+      if self._postprocessor is not None:
+        batch = self._postprocessor(batch)
 
     def _maybe_to_numpy(
-        x: tf.Tensor | tf.SparseTensor | tf.RaggedTensor,
+        x: tf.Tensor | tf.SparseTensor | tf.RaggedTensor | np.ndarray,
     ) -> np.ndarray | tf.SparseTensor | tf.RaggedTensor:
-      if isinstance(x, (tf.SparseTensor, tf.RaggedTensor)):
+      if isinstance(x, (tf.SparseTensor, tf.RaggedTensor, np.ndarray)):
         return x
       if hasattr(x, "_numpy"):
         numpy = x._numpy()  # pylint: disable=protected-access
@@ -83,13 +82,16 @@ def _maybe_to_numpy(
   @property
   def element_spec(self) -> clu_data.ElementSpec:
     if self._element_spec is not None:
-      batch = self._element_spec
-    else:
-      batch = self.__next__()
-      self._prefetched_batch = batch
+      return self._element_spec
+
+    batch = next(self._iterator)
+    if self._postprocessor is not None:
+      batch = self._postprocessor(batch)
+
+    self._prefetched_batch = batch
 
     def _to_element_spec(
-        x: np.ndarray | tf.SparseTensor | tf.RaggedTensor,
+        x: tf.Tensor | tf.SparseTensor | tf.RaggedTensor | np.ndarray,
     ) -> clu_data.ArraySpec:
       if isinstance(x, tf.SparseTensor):
         return clu_data.ArraySpec(
@@ -101,6 +103,10 @@ def _to_element_spec(
             dtype=x.dtype.as_numpy_dtype,  # pylint: disable=attribute-error
             shape=tuple(x.shape.as_list()),  # pylint: disable=attribute-error
         )
+      if isinstance(x, tf.Tensor):
+        return clu_data.ArraySpec(
+            dtype=x.dtype.as_numpy_dtype, shape=tuple(x.shape.as_list())
+        )
       return clu_data.ArraySpec(dtype=x.dtype, shape=tuple(x.shape))
 
     element_spec = tf.nest.map_structure(_to_element_spec, batch)

recml/core/ops/embedding_ops.py

@@ -0,0 +1,109 @@
+# Copyright 2024 RecML authors <[email protected]>.
+#
+# 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.
+"""Embedding lookup ops."""
+
+from collections.abc import Mapping, Sequence
+import dataclasses
+import functools
+
+import jax
+from jax.experimental import shard_map
+from jax_tpu_embedding.sparsecore.lib.nn import embedding
+from jax_tpu_embedding.sparsecore.lib.nn import embedding_spec
+
+
+@dataclasses.dataclass
+class SparsecoreParams:
+  """Embedding parameters."""
+
+  feature_specs: embedding.Nested[embedding_spec.FeatureSpec]
+  abstract_mesh: jax.sharding.AbstractMesh
+  data_axes: Sequence[str | None]
+  embedding_axes: Sequence[str | None]
+  sharding_strategy: str
+
+
+@functools.partial(jax.custom_vjp, nondiff_argnums=(0,))
+def sparsecore_lookup(
+    sparsecore_params: SparsecoreParams,
+    tables: Mapping[str, tuple[jax.Array, ...]],
+    csr_inputs: tuple[jax.Array, ...],
+):
+  return shard_map.shard_map(
+      functools.partial(
+          embedding.tpu_sparse_dense_matmul,
+          global_device_count=sparsecore_params.abstract_mesh.size,
+          feature_specs=sparsecore_params.feature_specs,
+          sharding_strategy=sparsecore_params.sharding_strategy,
+      ),
+      mesh=sparsecore_params.abstract_mesh,
+      in_specs=(
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.embedding_axes),
+      ),
+      out_specs=jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+      check_rep=False,
+  )(*csr_inputs, tables)
+
+
+def _emb_lookup_fwd(
+    sparsecore_params: SparsecoreParams,
+    tables: Mapping[str, tuple[jax.Array, ...]],
+    csr_inputs: tuple[jax.Array, ...],
+):
+  out = sparsecore_lookup(sparsecore_params, tables, csr_inputs)
+  return out, (tables, csr_inputs)
+
+
+def _emb_lookup_bwd(
+    sparsecore_params: SparsecoreParams,
+    res: tuple[Mapping[str, tuple[jax.Array, ...]], tuple[jax.Array, ...]],
+    gradients: embedding.Nested[jax.Array],
+) -> tuple[embedding.Nested[jax.Array], None]:
+  """Backward pass for embedding lookup."""
+  (tables, csr_inputs) = res
+
+  emb_table_grads = shard_map.shard_map(
+      functools.partial(
+          embedding.tpu_sparse_dense_matmul_grad,
+          feature_specs=sparsecore_params.feature_specs,
+          sharding_strategy=sparsecore_params.sharding_strategy,
+      ),
+      mesh=sparsecore_params.abstract_mesh,
+      in_specs=(
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+          jax.sharding.PartitionSpec(*sparsecore_params.embedding_axes),
+      ),
+      out_specs=jax.sharding.PartitionSpec(*sparsecore_params.data_axes),
+      check_rep=False,
+  )(gradients, *csr_inputs, tables)
+
+  # `tpu_sparse_dense_matmul_grad` returns a general mapping (usually a dict).
+  # It may not be the same type as the embedding table (e.g. FrozenDict).
+  # Here we use flatten / unflatten to ensure the types are the same.
+  emb_table_grads = jax.tree.unflatten(
+      jax.tree.structure(tables), jax.tree.leaves(emb_table_grads)
+  )
+
+  return emb_table_grads, None
+
+
+sparsecore_lookup.defvjp(_emb_lookup_fwd, _emb_lookup_bwd)

recml/core/training/jax.py

@@ -26,6 +26,7 @@
 from clu import periodic_actions
 import clu.metrics as clu_metrics
 from flax import struct
+import flax.linen as nn
 import jax
 import jax.numpy as jnp
 import keras
@@ -67,43 +68,85 @@ class JaxState(struct.PyTreeNode, Generic[MetaT]):
     step: A counter of the current step of the job. It starts at zero and it is
       incremented by 1 on a call to `state.update(...)`. This should be a Jax
       array and not a Python integer.
-    apply: A function that can be used to apply the forward pass of the model.
-      For Flax models this is usually set to `model.apply`.
     params: A pytree of trainable variables that will be updated by `tx` and
       used in `apply`.
     tx: An optax gradient transformation that will be used to update the
       parameters contained in `params` on a call to `state.update(...)`.
     opt_state: The optimizer state for `tx`. This is usually created by calling
       `tx.init(params)`.
+    _apply: An optional function that can be used to apply the forward pass of
+      the model. For Flax models this is usually set to `model.apply` while for
+      Haiku models this is usually set to `transform.apply`.
+    _model: An optional reference to a stateless Flax model for convenience.
     mutable: A pytree of mutable variables that are used by `apply`.
     meta: Arbitrary metadata that is recorded on the state. This can be useful
       for tracking additional references in the state.
   """
 
   step: jax.Array
-  apply: Callable[..., Any] = struct.field(pytree_node=False)
   params: PyTree = struct.field(pytree_node=True)
   tx: optax.GradientTransformation = struct.field(pytree_node=False)
   opt_state: optax.OptState = struct.field(pytree_node=True)
   mutable: PyTree = struct.field(pytree_node=True, default_factory=dict)
   meta: MetaT = struct.field(pytree_node=False, default_factory=dict)
+  _apply: Callable[..., Any] | None = struct.field(
+      pytree_node=False, default_factory=None
+  )
+  _model: nn.Module | None = struct.field(pytree_node=False, default=None)
+
+  @property
+  def model(self) -> nn.Module:
+    """Returns a reference to the model used to create the state."""
+    if self._model is None:
+      raise ValueError("No Flax `model` is set on the state.")
+    return self._model
+
+  def apply(self, *args, **kwargs) -> Any:
+    """Applies the forward pass of the model."""
+    if self._apply is None:
+      raise ValueError("No `apply` function is set on the state.")
+    return self._apply(*args, **kwargs)
 
   @classmethod
   def create(
       cls,
       *,
-      apply: Callable[..., Any],
+      apply: Callable[..., Any] | None = None,
+      model: nn.Module | None = None,
       params: PyTree,
       tx: optax.GradientTransformation,
       **kwargs,
   ) -> Self:
-    """Creates a new instance from a Jax apply function and Optax optimizer."""
+    """Creates a new instance from a Jax model / apply fn and Optax optimizer.
+
+    Args:
+      apply: A function that can be used to apply the forward pass of the model.
+        For Flax models this is usually set to `model.apply`. This cannot be set
+        along with `model`.
+      model: A reference to a stateless Flax model. This cannot be set along
+        with `apply`. When set the `apply` attribute of the state will be set to
+        `model.apply`.
+      params: A pytree of trainable variables that will be updated by `tx` and
+        used in `apply`.
+      tx: An optax gradient transformation that will be used to update the
+        parameters contained in `params` on a call to `state.update(...)`.
+      **kwargs: Other updates to set on the new state.
+
+    Returns:
+      An new instance of the state.
+    """
+    if apply is not None and model is not None:
+      raise ValueError("Only one of `apply` or `model` can be provided.")
+    elif model is not None:
+      apply = model.apply
+
     return cls(
         step=jnp.zeros([], dtype=jnp.int32),
-        apply=apply,
         params=params,
         tx=tx,
         opt_state=tx.init(params),
+        _apply=apply,
+        _model=model,
         **kwargs,
     )
 

recml/core/training/optax_factory.py

@@ -29,10 +29,10 @@ def _default_weight_decay_mask(params: optax.Params) -> optax.Params:
 
 
 def _regex_mask(regex: str) -> Callable[[optax.Params], optax.Params]:
-  """Returns a weight decay mask that applies to parameters matching a regex."""
+  """Returns a mask that applies to parameters matching a regex."""
 
   def _matches_regex(path: tuple[str, ...], _: Any) -> bool:
-    key = "/".join([jax.tree_util.keystr((k,), simple=True) for k in path])
+    key = '/'.join([jax.tree_util.keystr((k,), simple=True) for k in path])
     return re.fullmatch(regex, key) is not None
 
   def _mask(params: optax.Params) -> optax.Params:
@@ -54,6 +54,8 @@ class OptimizerFactory(types.Factory[optax.GradientTransformation]):
       magnitude of the gradients during optimization. Defaults to None.
     weight_decay_mask: The weight decay mask to use when applying weight decay.
       Defaults applying weight decay to all non-1D parameters.
+    freeze_mask: Optional mask to freeze parameters during optimization.
+      Defaults to None.
 
   Example usage:
 
@@ -78,6 +80,7 @@ class OptimizerFactory(types.Factory[optax.GradientTransformation]):
   weight_decay_mask: str | Callable[[optax.Params], optax.Params] = (
       _default_weight_decay_mask
   )
+  freeze_mask: str | Callable[[optax.Params], optax.Params] | None = None
 
   def make(self) -> optax.GradientTransformation:
     if self.grad_clip_norm is not None:
@@ -99,13 +102,30 @@ def make(self) -> optax.GradientTransformation:
     else:
       weight_decay = optax.identity()
 
-    return optax.chain(*[
+    tx = optax.chain(*[
         apply_clipping,
         self.scaling,
         weight_decay,
         lr_scaling,
     ])
 
+    if self.freeze_mask is not None:
+      if isinstance(self.freeze_mask, str):
+        mask = _regex_mask(self.freeze_mask)
+      else:
+        mask = self.freeze_mask
+
+      def _param_labels(params: optax.Params) -> optax.Params:
+        return jax.tree.map(
+            lambda p: 'frozen' if mask(p) else 'trainable', params
+        )
+
+      tx = optax.multi_transform(
+          transforms={'trainable': tx, 'frozen': optax.set_to_zero()},
+          param_labels=_param_labels,
+      )
+    return tx
+
 
 class AdamFactory(types.Factory[optax.GradientTransformation]):
   """Adam optimizer factory.
@@ -121,6 +141,8 @@ class AdamFactory(types.Factory[optax.GradientTransformation]):
       magnitude of the gradients during optimization. Defaults to None.
     weight_decay_mask: The weight decay mask to use when applying weight decay.
       Defaults applying weight decay to all non-1D parameters.
+    freeze_mask: Optional mask to freeze parameters during optimization.
+      Defaults to None.
 
   Example usage:
   ```
@@ -143,6 +165,7 @@ class AdamFactory(types.Factory[optax.GradientTransformation]):
   weight_decay_mask: str | Callable[[optax.Params], optax.Params] = (
       _default_weight_decay_mask
   )
+  freeze_mask: str | Callable[[optax.Params], optax.Params] | None = None
 
   def make(self) -> optax.GradientTransformation:
     return OptimizerFactory(
@@ -164,6 +187,8 @@ class AdagradFactory(types.Factory[optax.GradientTransformation]):
     eps: The epsilon coefficient for the Adagrad optimizer. Defaults to 1e-7.
     grad_clip_norm: Optional gradient clipping norm to limit the maximum
       magnitude of the gradients during optimization. Defaults to None.
+    freeze_mask: Optional mask to freeze parameters during optimization.
+      Defaults to None.
 
   Example usage:
   ```
@@ -175,6 +200,7 @@ class AdagradFactory(types.Factory[optax.GradientTransformation]):
   initial_accumulator_value: float = 0.1
   eps: float = 1e-7
   grad_clip_norm: float | None = None
+  freeze_mask: str | Callable[[optax.Params], optax.Params] | None = None
 
   def make(self) -> optax.GradientTransformation:
     return OptimizerFactory(