Make cache_size optional

PiperOrigin-RevId: 701019861

Author: Conchylicultor

gemma/transformer.py

@@ -63,7 +63,7 @@ class TransformerConfig:
   use_post_attn_norm: bool
   use_post_ffw_norm: bool
   attention_types: Iterable[modules.AttentionType]
-  max_cache_length: int = 1024
+  max_cache_length: int | None = 1024
   query_pre_attn_norm: QueryPreAttentionNormalisation = (
       QueryPreAttentionNormalisation.BY_ONE_OVER_SQRT_HEAD_DIM
   )
@@ -83,7 +83,7 @@ def query_pre_attn_scalar(self) -> float:
 
   @classmethod
   def from_params(
-      cls, params: params_lib.Params, cache_size: int = 1024
+      cls, params: params_lib.Params, cache_size: int | None = 1024
   ) -> 'TransformerConfig':
     """Creates a TransformerConfig from loaded parameters.
 
@@ -116,7 +116,7 @@ def from_params(
     )
 
   @classmethod
-  def gemma_2b(cls, cache_size: int):
+  def gemma_2b(cls, cache_size: int | None):
     return cls(
         num_layers=_NUM_LAYERS_GEMMA_2B,
         num_embed=256128,
@@ -133,7 +133,7 @@ def gemma_2b(cls, cache_size: int):
     )
 
   @classmethod
-  def gemma_7b(cls, cache_size: int):
+  def gemma_7b(cls, cache_size: int | None):
     return cls(
         num_layers=_NUM_LAYERS_GEMMA_7B,
         num_embed=256128,
@@ -150,7 +150,7 @@ def gemma_7b(cls, cache_size: int):
     )
 
   @classmethod
-  def gemma2_2b(cls, cache_size: int):
+  def gemma2_2b(cls, cache_size: int | None):
     return cls(
         num_layers=_NUM_LAYERS_GEMMA2_2B,
         num_embed=256128,
@@ -174,7 +174,7 @@ def gemma2_2b(cls, cache_size: int):
     )
 
   @classmethod
-  def gemma2_9b(cls, cache_size: int):
+  def gemma2_9b(cls, cache_size: int | None):
     return cls(
         num_layers=_NUM_LAYERS_GEMMA2_9B,
         num_embed=256128,
@@ -199,7 +199,7 @@ def gemma2_9b(cls, cache_size: int):
     )
 
   @classmethod
-  def gemma2_27b(cls, cache_size: int):
+  def gemma2_27b(cls, cache_size: int | None):
     return cls(
         num_layers=_NUM_LAYERS_GEMMA2_27B,
         num_embed=256128,
@@ -229,6 +229,8 @@ def init_cache(
       dtype: jnp.dtype = jnp.bfloat16,
   ) -> Cache:
     """Initializes a new Transformer cache."""
+    if self.max_cache_length is None:
+      raise ValueError('max_cache_length must be set to initialize cache.')
     cache = {
         f'layer_{i}': modules.Attention.init_cache(
             self.max_cache_length,