Fix missing additional_stop_token_ids (#1769)

python/sglang/srt/hf_transformers_utils.py

@@ -164,14 +164,7 @@ def get_tokenizer(
             "slowdown. Consider using a fast tokenizer instead."
         )
 
-    # Special handling for stop token <|eom_id|> generated by llama 3 tool use.
-    if "<|eom_id|>" in tokenizer.get_added_vocab():
-        tokenizer.additional_stop_token_ids = set(
-            [tokenizer.get_added_vocab()["<|eom_id|>"]]
-        )
-    else:
-        tokenizer.additional_stop_token_ids = None
-
+    handle_additional_stop_token_ids(tokenizer)
     return tokenizer
 
 
@@ -190,4 +183,16 @@ def get_processor(
         tokenizer_revision=tokenizer_revision,
         **kwargs,
     )
+
+    handle_additional_stop_token_ids(processor.tokenizer)
     return processor
+
+
+def handle_additional_stop_token_ids(tokenizer):
+    # Special handling for stop token <|eom_id|> generated by llama 3 tool use.
+    if "<|eom_id|>" in tokenizer.get_added_vocab():
+        tokenizer.additional_stop_token_ids = set(
+            [tokenizer.get_added_vocab()["<|eom_id|>"]]
+        )
+    else:
+        tokenizer.additional_stop_token_ids = None

python/sglang/srt/sampling/sampling_params.py

@@ -135,7 +135,7 @@ class SamplingParams:
             self.stop_str_max_len = stop_str_max_len
 
         # Process stop token ids
-        if tokenizer.additional_stop_token_ids:
+        if tokenizer and tokenizer.additional_stop_token_ids:
             self.stop_token_ids.update(tokenizer.additional_stop_token_ids)
 
     def to_srt_kwargs(self):