fixed decoder_start_token_id for T5 (#1552)

* fixed decoder_start_token_id for T5, pad_token as default for decoder_start --------- Co-authored-by: Ehsan Jahanbakhsh <[email protected]>
OpenNMT · Nov 17, 2023 · c5f46a3 · c5f46a3
1 parent 120746e
commit c5f46a3
Showing 1 changed file with 6 additions and 1 deletion.
diff --git a/python/ctranslate2/converters/transformers.py b/python/ctranslate2/converters/transformers.py
@@ -1024,7 +1024,12 @@ def set_config(self, config, model, tokenizer):
         config.bos_token = tokenizer.pad_token
         config.eos_token = tokenizer.eos_token
         config.unk_token = tokenizer.unk_token
-        config.decoder_start_token = tokenizer.pad_token
+        if hasattr(model.config, "decoder_start_token_id"):
+            config.decoder_start_token = tokenizer.convert_ids_to_tokens(
+                model.config.decoder_start_token_id
+            )
+        else:
+            config.decoder_start_token = tokenizer.pad_token
 
     def set_stack(self, spec, module, is_decoder=False):
         self.set_layer_norm(spec.layer_norm, module.final_layer_norm)