Refactor to use a config object to manage arguments to sparse dense matmul.

PiperOrigin-RevId: 751450733

Author: manuadg-google

jax_tpu_embedding/sparsecore/examples/shakespeare/jax_sc_shakespeare_jit.py

@@ -46,6 +46,7 @@
 import orbax.checkpoint as ocp
 import tree
 
+
 np.set_printoptions(threshold=np.inf)
 Nested = embedding.Nested
 
@@ -376,7 +377,7 @@ def train_step_fn(
       mesh: jax.sharding.Mesh,
       model: nn.Module,
       optimizer,
-      feature_specs,
+      config: embedding.SparseDenseMatmulConfig,
       train_state: TrainState,
       preprocessed_inputs,
       emb_variables,
@@ -386,22 +387,15 @@ def train_step_fn(
 
     # Sparse forward pass - embedding lookup.
     with jax.named_scope('sc_forward_pass'):
-      tpu_sparse_dense_matmul = partial(
-          embedding.tpu_sparse_dense_matmul,
-          global_device_count=num_global_devices,
-          feature_specs=feature_specs,
-          sharding_strategy='MOD',
-      )
       tpu_sparse_dense_matmul = shard_map(
-          f=tpu_sparse_dense_matmul,
+          f=embedding.tpu_sparse_dense_matmul,
           mesh=mesh,
-          in_specs=(pd, pe),
+          in_specs=(pd, pe, None),
           out_specs=pd,
           check_rep=False,
       )
       emb_act = tpu_sparse_dense_matmul(
-          preprocessed_inputs,
-          emb_variables,
+          preprocessed_inputs, emb_variables, config
       )
 
     # Dense forward + backward pass.
@@ -429,22 +423,15 @@ def train_step_fn(
 
     # Sparse backward pass - embedding update.
     with jax.named_scope('sc_backward_pass'):
-      tpu_sparse_dense_matmul_grad = partial(
-          embedding.tpu_sparse_dense_matmul_grad,
-          feature_specs=feature_specs,
-          sharding_strategy='MOD',
-      )
       tpu_sparse_dense_matmul_grad = shard_map(
-          f=tpu_sparse_dense_matmul_grad,
+          f=embedding.tpu_sparse_dense_matmul_grad,
           mesh=mesh,
-          in_specs=(pd, pd, pe),
+          in_specs=(pd, pd, pe, None),
           out_specs=pe,
           check_rep=False,
       )
       emb_variables = tpu_sparse_dense_matmul_grad(
-          emb_grad,
-          preprocessed_inputs,
-          emb_variables,
+          emb_grad, preprocessed_inputs, emb_variables, config
       )
 
     train_state = train_state.replace(
@@ -503,16 +490,17 @@ def train_step_fn(
         lambda y: jax.make_array_from_process_local_data(global_sharding, y),
         x,
     )
+    config = embedding.SparseDenseMatmulConfig(
+        global_device_count=num_global_devices,
+        local_device_count=num_local_devices,
+        feature_specs=flax.core.freeze(feature_specs),
+        num_sc_per_device=num_sc_per_device,
+        sharding_strategy='MOD',
+    )
     preprocessed_inputs, stats = map(
         make_global_view,
         embedding.preprocess_sparse_dense_matmul_input(
-            features,
-            feature_weights,
-            feature_specs,
-            local_device_count=global_mesh.local_mesh.size,
-            global_device_count=global_mesh.size,
-            num_sc_per_device=num_sc_per_device,
-            sharding_strategy='MOD',
+            features, feature_weights, config=config
         ),
     )
     fdo_client.record(stats)
@@ -524,7 +512,7 @@ def train_step_fn(
         global_mesh,
         model,
         optimizer,
-        feature_specs,
+        config,
         train_state,
         preprocessed_inputs,
         emb_variables,

jax_tpu_embedding/sparsecore/examples/shakespeare/jax_sc_shakespeare_pmap.py

@@ -11,8 +11,6 @@
 # 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.
-"""An example Shakespeare model that uses the SparseCore embedding API."""
-
 # pylint: disable=g-importing-member
 import collections
 from functools import partial
@@ -42,6 +40,8 @@
 import tree
 
 
+"""An example Shakespeare model that uses the SparseCore embedding API."""
+
 np.set_printoptions(threshold=np.inf)
 Nested = embedding.Nested
 
@@ -260,10 +260,9 @@ def run_model():
   )
 
   def train_step_fn(
-      global_device_count: int,
       model: nn.Module,
       optimizer,
-      feature_specs,
+      config,
       train_state: TrainState,
       preprocessed_inputs,
       emb_variables: Mapping[str, embedding.EmbeddingVariables],
@@ -276,9 +275,7 @@ def train_step_fn(
     with jax.named_scope('sc_forward_pass'):
       tpu_sparse_dense_matmul = partial(
           embedding.tpu_sparse_dense_matmul,
-          global_device_count=global_device_count,
-          feature_specs=feature_specs,
-          sharding_strategy='MOD',
+          config=config,
       )
       emb_act = tpu_sparse_dense_matmul(
           preprocessed_inputs,
@@ -312,8 +309,7 @@ def train_step_fn(
     with jax.named_scope('sc_backward_pass'):
       tpu_sparse_dense_matmul_grad = partial(
           embedding.tpu_sparse_dense_matmul_grad,
-          feature_specs=feature_specs,
-          sharding_strategy='MOD',
+          config=config,
       )
       emb_variables = tpu_sparse_dense_matmul_grad(
           emb_grad,
@@ -337,14 +333,16 @@ def train_step_fn(
   vlog1('Replicating train_state')
   train_state = flax_utils.replicate(train_state, local_devices)
   parameter_overview.log_parameter_overview(train_state.params)
+
+  config = embedding.SparseDenseMatmulConfig(
+      feature_specs=feature_specs,
+      local_device_count=global_mesh.local_mesh.size,
+      global_device_count=global_mesh.size,
+      num_sc_per_device=num_sc_per_device,
+      has_leading_dimension=True,
+  )
   p_train_step_fn = jax.pmap(
-      partial(
-          train_step_fn,
-          num_global_devices,
-          model,
-          optimizer,
-          feature_specs,
-      ),
+      partial(train_step_fn, model, optimizer, config),
       axis_name='batch',
   )
 
@@ -376,16 +374,16 @@ def train_step_fn(
     )
 
     # Preprocess the inputs.
-    preprocessed_inputs, _ = embedding.preprocess_sparse_dense_matmul_input(
-        features,
-        feature_weights,
-        feature_specs,
+    config = embedding.SparseDenseMatmulConfig(
+        feature_specs=flax.core.freeze(feature_specs),
         local_device_count=global_mesh.local_mesh.size,
         global_device_count=global_mesh.size,
         num_sc_per_device=num_sc_per_device,
-        sharding_strategy='MOD',
         has_leading_dimension=True,
     )
+    preprocessed_inputs, _ = embedding.preprocess_sparse_dense_matmul_input(
+        features, feature_weights, config
+    )
 
     # TODO(patn): This (local_slice)will go away once the input processor is
     # updated to only produce local batches.

jax_tpu_embedding/sparsecore/lib/flax/embed.py

@@ -88,6 +88,14 @@ def __post_init__(self):
         self.mesh.devices.item(0)
     )
 
+    self.config = embedding.SparseDenseMatmulConfig(
+        feature_specs=self.feature_specs,
+        local_device_count=self.mesh.local_mesh.size,
+        global_device_count=self.mesh.size,
+        num_sc_per_device=self.num_sc_per_device,
+        sharding_strategy=self.table_sharding_strategy,
+    )
+
     super().__post_init__()
 
   def setup(self):
@@ -143,11 +151,7 @@ def preprocess_inputs(
     return embedding.preprocess_sparse_dense_matmul_input(
         features,
         features_weights,
-        self.feature_specs,
-        self.mesh.local_mesh.size,
-        self.mesh.size,
-        num_sc_per_device=self.num_sc_per_device,
-        sharding_strategy=self.table_sharding_strategy,
+        self.config,
     )[0]
 
   def __call__(self, embedding_lookups: EmbeddingLookups) -> Nested[jax.Array]:
@@ -198,34 +202,22 @@ def _emb_lookup(
   pt = embedding_layer.embedding_table_partition
   pd = embedding_layer.data_partition
   return shard_map(
-      functools.partial(
-          embedding.tpu_sparse_dense_matmul,
-          global_device_count=embedding_layer.mesh.size,
-          feature_specs=embedding_layer.feature_specs,
-          sharding_strategy=embedding_layer.table_sharding_strategy,
-      ),
+      embedding.tpu_sparse_dense_matmul,
       mesh=embedding_layer.mesh,
-      in_specs=(pd, pt),
+      in_specs=(pd, pt, None),
       out_specs=pd,
       check_rep=False,
-  )(
-      embedding_lookups,
-      emb_table,
-  )
+  )(embedding_lookups, emb_table, embedding_layer.config)
 
 
 def _emb_lookup_fwd(
     embedding_layer: SparseCoreEmbed,
     embedding_lookups: EmbeddingLookups,
     emb_table: Mapping[str, tuple[jax.Array, ...]],
 ):
-  return _emb_lookup(
-      embedding_layer,
-      embedding_lookups,
-      emb_table,
-  ), (
-      embedding_lookups,
-      emb_table,
+  return (
+      _emb_lookup(embedding_layer, embedding_lookups, emb_table),
+      (embedding_lookups, emb_table),
   )
 
 
@@ -236,20 +228,12 @@ def _emb_lookup_bwd(embedding_layer, res, gradients):
   pt = embedding_layer.embedding_table_partition
   pd = embedding_layer.data_partition
   emb_table_grads = shard_map(
-      functools.partial(
-          embedding.tpu_sparse_dense_matmul_grad,
-          feature_specs=embedding_layer.feature_specs,
-          sharding_strategy=embedding_layer.table_sharding_strategy,
-      ),
+      embedding.tpu_sparse_dense_matmul_grad,
       mesh=embedding_layer.mesh,
-      in_specs=(pd, pd, pt),
+      in_specs=(pd, pd, pt, None),
       out_specs=pt,
       check_rep=False,
-  )(
-      gradients,
-      embedding_lookups,
-      emb_table,
-  )
+  )(gradients, embedding_lookups, emb_table, config=embedding_layer.config)
 
   # 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).
@@ -258,10 +242,7 @@ def _emb_lookup_bwd(embedding_layer, res, gradients):
       jax.tree.structure(emb_table), jax.tree.leaves(emb_table_grads)
   )
 
-  return (
-      None,
-      emb_table_grads,
-  )
+  return (None, emb_table_grads)
 
 
 _emb_lookup.defvjp(_emb_lookup_fwd, _emb_lookup_bwd)

jax_tpu_embedding/sparsecore/lib/flax/tests/autograd_test.py

@@ -129,7 +129,7 @@ def test_shakespeare_model_loss_convergence(self):
     )
 
     # Initialize the model.
-    def process_inputs(feature_batch):
+    def process_inputs(feature_batch, config):
       features = np.reshape(feature_batch, (-1, 1))
       feature_weights = np.ones(features.shape, dtype=np.float32)
 
@@ -144,15 +144,18 @@ def process_inputs(feature_batch):
           *embedding.preprocess_sparse_dense_matmul_input(
               features,
               feature_weights,
-              feature_specs,
-              mesh.local_mesh.size,
-              mesh.size,
-              num_sc_per_device=num_sc_per_device,
-              sharding_strategy='MOD',
+              config,
           )[0]
       )
 
-    first_model_input = process_inputs(feature_batches[0])
+    config = embedding.SparseDenseMatmulConfig(
+        feature_specs=feature_specs,
+        local_device_count=mesh.local_mesh.size,
+        global_device_count=mesh.size,
+        num_sc_per_device=num_sc_per_device,
+        sharding_strategy='MOD',
+    )
+    first_model_input = process_inputs(feature_batches[0], config)
     params = model.init(jax.random.key(42), first_model_input)
 
     # Create optimizer.
@@ -204,7 +207,7 @@ def forward_pass(params, embedding_lookups, labels):
       # ------------------------------------------------------------------------
       # Step 1: SC input processing.
       # ------------------------------------------------------------------------
-      processed_input_tensor = process_inputs(features)
+      processed_input_tensor = process_inputs(features, config)
 
       # ------------------------------------------------------------------------
       # Step 2: run model.