diff --git a/optformer/embed_then_regress/icl_transformer.py b/optformer/embed_then_regress/icl_transformer.py
index 43b082c..6a04eba 100644
--- a/optformer/embed_then_regress/icl_transformer.py
+++ b/optformer/embed_then_regress/icl_transformer.py
@@ -14,9 +14,11 @@
 
 """Transformer model for ICL regression."""
 
+import dataclasses
 import functools
 from typing import Callable
 from flax import linen as nn
+from flax import struct
 import jax
 import jax.numpy as jnp
 import jaxtyping as jt
@@ -86,14 +88,12 @@ def __call__(
     return x
 
 
-class EmbeddingCache(dict[str, AnyTensor]):
+@struct.dataclass
+class EmbeddingCache:
+  """Cache for storing previously computed embeddings."""
 
-  def get_or_set(self, key: str, fn: Callable[[], AnyTensor]):
-    value = self.get(key)
-    if value is None:
-      value = fn()
-      self.update({key: value})
-    return value
+  x_emb: jt.Float[jax.Array, 'L E'] | None = None
+  metadata_emb: jt.Float[jax.Array, 'E'] | None = None
 
 
 class ICLTransformer(nn.Module):
@@ -206,36 +206,33 @@ def infer(
       x_targ: jt.Int[jax.Array, 'Q T'],  # Q is fixed to avoid re-jitting.
       metadata: jt.Int[jax.Array, 'T'],
       mask: jt.Bool[jax.Array, 'L'],
-      cache: EmbeddingCache | None = None,  # For caching embeddings.
+      cache: EmbeddingCache,  # For caching embeddings.
   ) -> tuple[
       jt.Float[jax.Array, 'L'],
       jt.Float[jax.Array, 'L'],
-      dict[str, jax.Array],
+      EmbeddingCache,
   ]:
     """Friendly for inference, no batch dimension."""
-    if cache is None:
-      cache = EmbeddingCache()
-
-    # [L, E]
-    x_pad_emb = cache.get_or_set('x_pad_emb', lambda: self.embed(x_padded))
+    if cache.x_emb is None:
+      cache = dataclasses.replace(cache, x_emb=self.embed(x_padded))
+    x_pad_emb = cache.x_emb  # [L, E]
     x_targ_emb = self.embed(x_targ)  # [Q, E]
-    L, E = x_pad_emb.shape  # pylint: disable=invalid-name
 
     # Combine target and historical (padded) embeddings.
     target_index = jnp.sum(mask, dtype=jnp.int32)  # [1]
-    padded_target_emb = jnp.zeros((L, E), dtype=x_targ_emb.dtype)
+    padded_target_emb = jnp.zeros_like(x_pad_emb)
     padded_target_emb = jax.lax.dynamic_update_slice_in_dim(
         padded_target_emb, x_targ_emb, start_index=target_index, axis=0
     )
     w_mask = jnp.expand_dims(mask, axis=-1)  # [L, 1]
-    x_emb = x_pad_emb * w_mask + padded_target_emb * (1 - w_mask)
+    x_emb = x_pad_emb * w_mask + padded_target_emb * (1 - w_mask)  # [L, E]
 
     if self.use_metadata:  # Attach metadata embeddings too.
-      metadata_emb = cache.get_or_set(
-          'metadata_emb', lambda: self.embed(metadata)
-      )
+      if cache.metadata_emb is None:
+        cache = dataclasses.replace(cache, metadata_emb=self.embed(metadata))
+      metadata_emb = cache.metadata_emb  # [E]
       metadata_emb = jnp.expand_dims(metadata_emb, axis=0)  # [1, E]
-      metadata_emb = jnp.repeat(metadata_emb, L, axis=0)  # [L, E]
+      metadata_emb = jnp.repeat(metadata_emb, x_emb.shape[0], axis=0)  # [L, E]
       x_emb = jnp.concatenate((x_emb, metadata_emb), axis=-1)  # [L, 2E]
 
     mean, std = self.__call__(