Fix stop condition for <|eom_id|> (#1766)

python/sglang/srt/hf_transformers_utils.py

@@ -163,6 +163,15 @@ def get_tokenizer(
             "Using a slow tokenizer. This might cause a significant "
             "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
+
     return tokenizer