diff --git a/recml/layers/linen/sparsecore.py b/recml/layers/linen/sparsecore.py
index cc41957..1f1b660 100644
--- a/recml/layers/linen/sparsecore.py
+++ b/recml/layers/linen/sparsecore.py
@@ -276,10 +276,12 @@ class SparsecorePreprocessor:
     sparsecore_config: The sparsecore config used to create the tables.
     global_batch_size: The global batch size across all devices to partition the
       inputs across.
+    _batch_number: The batch number for preprocessing, incremented on each call.
   """
 
   sparsecore_config: SparsecoreConfig
   global_batch_size: int
+  _batch_number: int = dataclasses.field(init=False, default=0)
 
   def __post_init__(self):
     self.sparsecore_config.init_feature_specs(self.global_batch_size)
@@ -328,6 +330,7 @@ def _to_np(x: Any) -> np.ndarray:
         if weights[key] is not None:
           weights[key] = np.reshape(weights[key], (-1, 1))
 
+    self._batch_number += 1
     csr_inputs, _ = embedding.preprocess_sparse_dense_matmul_input(
         features=features,
         features_weights=weights,
@@ -337,6 +340,7 @@ def _to_np(x: Any) -> np.ndarray:
         num_sc_per_device=self.sparsecore_config.num_sc_per_device,
         sharding_strategy=self.sparsecore_config.sharding_strategy,
         allow_id_dropping=False,
+        batch_number=self._batch_number,
     )
 
     processed_inputs = {