change to jnp.bfloat16

Author: carlesoctav

tpu_inference/runner/compilation_manager.py

@@ -114,7 +114,7 @@ def _precompile_pooling(self) -> None:
 
         for num_tokens in self.runner.num_tokens_paddings:
             hidden_states = self._create_dummy_tensor(
-                (num_tokens, hidden_size), t2j_dtype(dtype), sharding=hidden_sharding)
+                (num_tokens, hidden_size), jnp.bfloat16, sharding=hidden_sharding)
 
             for num_reqs in self.runner.num_reqs_paddings:
                 if num_reqs == 0 or num_reqs > num_tokens: