keras-rs-nightly 0.0.1.dev2025043003__py3-none-any.whl → 0.2.2.dev202506100336__py3-none-any.whl

keras_rs/src/layers/embedding/jax/embedding_lookup.py

@@ -0,0 +1,255 @@
+"""Defines a differentiable `embedding_lookup` function.
+
+Implementation details for use in JAX models.
+"""
+
+import functools
+from typing import Any, Mapping, TypeAlias
+
+import jax
+import numpy as np
+from jax.experimental import layout
+from jax_tpu_embedding.sparsecore.lib.nn import embedding
+from jax_tpu_embedding.sparsecore.lib.nn import embedding_spec
+from jax_tpu_embedding.sparsecore.utils import utils as jte_utils
+
+from keras_rs.src.layers.embedding.jax import embedding_utils
+from keras_rs.src.types import Nested
+
+ShardedCooMatrix = embedding_utils.ShardedCooMatrix
+shard_map = jax.experimental.shard_map.shard_map  # type: ignore[attr-defined]
+
+ArrayLike: TypeAlias = jax.Array | np.ndarray[Any, Any]
+JaxLayout: TypeAlias = jax.sharding.NamedSharding | layout.Layout
+
+
+class EmbeddingLookupConfiguration:
+    """Feature, mesh and sharding configrmation for lookups."""
+
+    mesh: jax.sharding.Mesh
+    feature_specs: embedding.Nested[embedding_spec.FeatureSpec]
+    table_sharding_strategy: str
+    num_sc_per_device: int
+    samples_partition: jax.sharding.PartitionSpec
+    table_partition: jax.sharding.PartitionSpec
+    table_layout: JaxLayout
+
+    def __init__(
+        self,
+        feature_specs: embedding.Nested[embedding_spec.FeatureSpec],
+        mesh: jax.sharding.Mesh | None = None,
+        table_sharding_strategy: str = "MOD",
+        num_sc_per_device: int | None = None,
+        sharding_axis: str = "sparsecore_sharding",
+        samples_partition: jax.sharding.PartitionSpec | None = None,
+        samples_layout: JaxLayout | None = None,
+        table_partition: jax.sharding.PartitionSpec | None = None,
+        table_layout: JaxLayout | None = None,
+    ):
+        self.mesh = mesh or jax.sharding.Mesh(jax.devices(), sharding_axis)
+        self.feature_specs = feature_specs
+        self.table_sharding_strategy = table_sharding_strategy
+        self.num_sc_per_device = (
+            num_sc_per_device
+            if num_sc_per_device is not None
+            else jte_utils.num_sparsecores_per_device()
+        )
+        self.samples_partition = (
+            samples_partition
+            or jax.sharding.PartitionSpec(
+                sharding_axis  # type: ignore[no-untyped-call]
+            )
+        )
+        self.samples_layout = samples_layout or jax.sharding.NamedSharding(
+            self.mesh, self.samples_partition
+        )
+        self.table_partition = table_partition or jax.sharding.PartitionSpec(
+            sharding_axis,
+            None,  # type: ignore[no-untyped-call]
+        )
+        self.table_layout = table_layout or jax.sharding.NamedSharding(
+            self.mesh, self.table_partition
+        )
+
+
+# Embedding lookup function with custom gradient.
+@functools.partial(jax.custom_vjp, nondiff_argnums=(0,))
+def embedding_lookup(
+    config: EmbeddingLookupConfiguration,
+    lookups: Mapping[str, ShardedCooMatrix],
+    tables: Nested[jax.Array],
+    step: jax.Array | None = None,
+) -> Nested[jax.Array]:
+    """Embedding lookup function with custom gradient.
+
+    Args:
+      config: Embedding lookup configuration.
+      lookups: Embedding lookup stacked/sharded inputs.
+      tables: Embedding lookup stacked/sharded tables.
+      step: Current training step number.
+    """
+    del step  # Only used in backward pass.
+
+    # Decompose COO matrices.
+    row_pointers_raw = {}
+    embedding_ids_raw = {}
+    sample_ids_raw = {}
+    gains_raw = {}
+    for table_name, coo in lookups.items():
+        row_pointers_raw[table_name] = coo.shard_ends
+        embedding_ids_raw[table_name] = coo.col_ids
+        sample_ids_raw[table_name] = coo.row_ids
+        gains_raw[table_name] = coo.values
+
+    sparse_dense_matmul_input = embedding.SparseDenseMatmulInput(
+        row_pointers_raw,
+        embedding_ids_raw,
+        sample_ids_raw,
+        gains_raw,
+    )
+
+    pd = config.samples_partition
+    pt = config.table_partition
+    sharded_matmul = jax.jit(
+        shard_map(
+            functools.partial(
+                embedding.tpu_sparse_dense_matmul,
+                global_device_count=config.mesh.shape[pd[0]],
+                feature_specs=config.feature_specs,
+                sharding_strategy=config.table_sharding_strategy,
+            ),
+            mesh=config.mesh,
+            in_specs=(pd, pt),
+            out_specs=pd,
+            check_rep=False,
+        ),
+    )
+
+    activations: Nested[jax.Array] = sharded_matmul(
+        sparse_dense_matmul_input,
+        tables,
+    )
+
+    return activations
+
+
+def embedding_lookup_fwd(
+    config: EmbeddingLookupConfiguration,
+    lookups: Mapping[str, ShardedCooMatrix],
+    table: Nested[jax.Array],
+    step: jax.Array | None = None,
+) -> tuple[
+    Nested[jax.Array],
+    tuple[Nested[ShardedCooMatrix], Nested[jax.Array], jax.Array | None],
+]:
+    """Forward pass for embedding lookup."""
+    return embedding_lookup(config, lookups, table, step), (
+        lookups,
+        table,
+        step,
+    )
+
+
+def embedding_lookup_bwd(
+    config: EmbeddingLookupConfiguration,
+    res: tuple[
+        Mapping[str, ShardedCooMatrix],  # Lookups.
+        Mapping[str, Nested[jax.Array]],  # Tables.
+        jax.Array | None,  # Step.
+    ],
+    gradients: Nested[jax.Array],
+) -> tuple[None, Nested[jax.Array], jax.Array | None]:
+    """Backward pass for embedding lookup.
+
+    Args:
+      config: Embedding lookup configuration.
+      res: Tuple of embedding lookup (inputs, tables, step).
+      gradients: Embedding lookup gradients.
+
+    Returns:
+      A tuple of gradients (None, table_grads, step + 1).
+    """
+    lookups, tables, step = res
+
+    # Decompose COO matrices.
+    row_pointers_raw = {}
+    embedding_ids_raw = {}
+    sample_ids_raw = {}
+    gains_raw = {}
+    for table_name, coo in lookups.items():
+        row_pointers_raw[table_name] = coo.shard_ends
+        embedding_ids_raw[table_name] = coo.col_ids
+        sample_ids_raw[table_name] = coo.row_ids
+        gains_raw[table_name] = coo.values
+
+    sparse_dense_matmul_input = embedding.SparseDenseMatmulInput(
+        row_pointers_raw,
+        embedding_ids_raw,
+        sample_ids_raw,
+        gains_raw,
+    )
+
+    pt = config.table_partition
+    pd = config.samples_partition
+    # Replicate step count.
+    preplicate = jax.sharding.PartitionSpec()  # type: ignore[no-untyped-call]
+
+    def grad_func(
+        gradients: Nested[jax.Array],
+        sparse_input: embedding.SparseDenseMatmulInput,
+        tables: Mapping[str, embedding.EmbeddingVariables],
+        step: jax.Array | None,
+    ) -> Mapping[str, embedding.EmbeddingVariables]:
+        output: Mapping[str, embedding.EmbeddingVariables] = (
+            embedding.tpu_sparse_dense_matmul_grad(
+                gradients,
+                sparse_input,
+                tables,
+                feature_specs=config.feature_specs,
+                sharding_strategy=config.table_sharding_strategy,
+                step=step,
+            )
+        )
+        return output
+
+    # activation_layout = jax.sharding.NamedSharding(config.mesh, pd)
+    # step_layout = jax.sharding.NamedSharding(config.mesh, preplicate)
+    sharded_matmul_grad = jax.jit(
+        shard_map(
+            grad_func,
+            mesh=config.mesh,
+            in_specs=(pd, pd, pt, preplicate),
+            out_specs=pt,
+            check_rep=False,
+        ),
+        #   in_shardings=(
+        #       activation_layout,
+        #       config.samples_layout,
+        #       config.table_layout,
+        #       step_layout,
+        #   ),
+        # out_shardings=(config.table_layout),
+    )
+
+    table_grads = sharded_matmul_grad(
+        gradients,
+        sparse_dense_matmul_input,
+        tables,
+        step,
+    )
+
+    # 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.
+    table_grads = jax.tree.unflatten(
+        jax.tree.structure(tables), jax.tree.leaves(table_grads)
+    )
+
+    return (
+        None,
+        table_grads,
+        step + 1 if step is not None else None,  # Incremented step count.
+    )
+
+
+embedding_lookup.defvjp(embedding_lookup_fwd, embedding_lookup_bwd)