[Bugfix][Core] Fix tekken edge case for mistral tokenizer

tests/models/decoder_only/language/test_mistral.py

@@ -5,6 +5,7 @@
 import pytest
 
 from vllm import SamplingParams
+from vllm import LLM
 
 from ...utils import check_logprobs_close
 
@@ -16,6 +17,10 @@
 ]
 
 SAMPLING_PARAMS = SamplingParams(max_tokens=512, temperature=0.0, logprobs=5)
+SYMBOLIC_LANG_PROMPTS = [
+    "勇敢な船乗りについての詩を書く",  # japanese
+    "寫一首關於勇敢的水手的詩",  # chinese
+]
 
 # for function calling
 TOOLS = [{
@@ -77,7 +82,7 @@ def test_models(
     with hf_runner(model, dtype=dtype) as hf_model:
         hf_outputs = hf_model.generate_greedy_logprobs_limit(
             example_prompts, max_tokens, num_logprobs)
-
+            
     with vllm_runner(model, dtype=dtype,
                      tokenizer_mode="mistral") as vllm_model:
         vllm_outputs = vllm_model.generate_greedy_logprobs(
@@ -131,6 +136,29 @@ def test_mistral_format(
     )
 
 
+@pytest.mark.parametrize("model", MODELS[1:])
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [64])
+@pytest.mark.parametrize("num_logprobs", [5])
+@pytest.mark.parametrize("prompt", SYMBOLIC_LANG_PROMPTS)
+def test_mistral_symbolic_languages(
+    vllm_runner,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+    prompt: str,
+) -> None:
+    prompt = "hi"
+    msg = {"role": "user", "content": prompt}
+    llm = LLM(model=model,
+              tokenizer_mode="mistral",
+              config_format="mistral",
+              load_format="mistral")
+    outputs = llm.chat([msg], sampling_params=SAMPLING_PARAMS)
+    assert "�" not in outputs[0].outputs[0].text.strip()
+
+
 @pytest.mark.parametrize("dtype", ["bfloat16"])
 @pytest.mark.parametrize("model", MODELS[1:])  # v1 can't do func calling
 def test_mistral_function_calling(

vllm/transformers_utils/tokenizers/mistral.py

@@ -175,10 +175,29 @@ def apply_chat_template(self,
 
     def convert_tokens_to_string(self, tokens: List[str]) -> str:
         if isinstance(self.tokenizer, Tekkenizer):
-            return "".join(t for t in tokens
-                           if t not in self.tokenizer._all_special_tokens)
+            tokens = [
+                t for t in tokens
+                if t not in self.tokenizer._all_special_tokens
+            ]
+
+            if any(isinstance(t, bytes) for t in tokens):
+                # we need to encode and decode all tokens again
+                shift = self.tokenizer.num_special_tokens
+                byte_tokens = [
+                    t.encode("utf-8") if not isinstance(t, bytes) else t
+                    for t in tokens
+                ]
+                ids = [
+                    self.tokenizer._tekken_token2id_nospecial[t] + shift
+                    for t in byte_tokens
+                ]
+                decoded = self.tokenizer.decode(ids)
+            else:
+                decoded = "".join(tokens)
         else:
-            return self.tokenizer.decode(tokens)  # type: ignore[arg-type]
+            decoded = self.tokenizer.decode(tokens)  # type: ignore[arg-type]
+
+        return decoded
 
     def decode(self, ids: Union[List[int], int]) -> str:
         if isinstance(ids, int):
@@ -200,4 +219,11 @@ def convert_ids_to_tokens(
                               self.tokenizer)
 
         tokens = [self.tokenizer.id_to_piece(id) for id in ids]
+
+        if any(t.strip() == "�" for t in tokens):
+            # if any stripped decoded token is undefined
+            # because it's invalid unicode then pass bytes
+            # See: https://github.com/vllm-project/vllm/pull/8640
+            tokens = [self.tokenizer.id_to_byte_piece(id) for id in ids]
+
         return tokens