diff --git a/library/train_util.py b/library/train_util.py
index 8a73445e..c5e903bb 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3761,8 +3761,9 @@ def pool_workaround(
     # get hidden states for EOS token
     pooled_output = last_hidden_state[torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device), eos_token_index]
 
-    # apply projection
-    pooled_output = text_encoder.text_projection(pooled_output)
+    # apply projection: projection may be of different dtype than last_hidden_state
+    pooled_output = text_encoder.text_projection(pooled_output.to(text_encoder.text_projection.weight.dtype))
+    pooled_output = pooled_output.to(last_hidden_state.dtype)
 
     return pooled_output