works with latest flax version 0.5.2 now

min_dalle/models/dalle_bart_decoder_flax.py

@@ -35,7 +35,7 @@ class DecoderSelfAttentionFlax(AttentionFlax):
     ) -> Tuple[jnp.ndarray, Tuple[jnp.ndarray, jnp.ndarray]]:
         shape_split = decoder_state.shape[:2] + (self.head_count, -1)
         keys_state = lax.dynamic_update_slice(
-            keys_state, 
+            keys_state,
             self.k_proj(decoder_state).reshape(shape_split), 
             state_index
         )
@@ -205,6 +205,7 @@ class DalleBartDecoderFlax(nn.Module):
         params: dict
     ) -> jnp.ndarray:
         attention_mask = jnp.not_equal(text_tokens, 1)
+        encoder_state = encoder_state.astype(jnp.float16)
         
         def sample_next_image_token(
             state: SampleState,
@@ -247,8 +248,8 @@ class DalleBartDecoderFlax(nn.Module):
         initial_state = SampleState(
             prev_token = self.start_token,
             prng_key = prng_key,
-            keys_state = jnp.zeros(state_shape),
-            values_state = jnp.zeros(state_shape)
+            keys_state = jnp.zeros(state_shape, dtype=jnp.float16),
+            values_state = jnp.zeros(state_shape, dtype=jnp.float16)
         )
 
         _, image_tokens = lax.scan(