[Misc] add return token strs for tokenize (#18941)

Signed-off-by: reidliu41 <reid201711@gmail.com>
Co-authored-by: reidliu41 <reid201711@gmail.com>

tests/entrypoints/openai/test_tokenization.py

@@ -76,11 +76,11 @@ async def test_tokenize_completions(
                                  })
         response.raise_for_status()
 
-        assert response.json() == {
-            "tokens": tokens,
-            "count": len(tokens),
-            "max_model_len": 8192
-        }
+        result = response.json()
+        assert result["tokens"] == tokens
+        assert result["count"] == len(tokens)
+        assert result["max_model_len"] == 8192
+        assert result["token_strs"] is None
 
 
 @pytest.mark.asyncio
@@ -138,11 +138,11 @@ async def test_tokenize_chat(
                                          })
                 response.raise_for_status()
 
-                assert response.json() == {
-                    "tokens": tokens,
-                    "count": len(tokens),
-                    "max_model_len": 8192
-                }
+                result = response.json()
+                assert result["tokens"] == tokens
+                assert result["count"] == len(tokens)
+                assert result["max_model_len"] == 8192
+                assert result["token_strs"] is None
 
 
 @pytest.mark.asyncio
@@ -215,11 +215,46 @@ async def test_tokenize_chat_with_tools(
                 )
                 response.raise_for_status()
 
-                assert response.json() == {
-                    "tokens": tokens,
-                    "count": len(tokens),
-                    "max_model_len": 8192,
-                }
+                result = response.json()
+                assert result["tokens"] == tokens
+                assert result["count"] == len(tokens)
+                assert result["max_model_len"] == 8192
+                assert result["token_strs"] is None
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model_name, tokenizer_name",
+    [(MODEL_NAME, MODEL_NAME), ("zephyr-lora2", "zephyr-lora2")],
+    indirect=["tokenizer_name"],
+)
+async def test_tokenize_with_return_token_strs(
+    server: RemoteOpenAIServer,
+    model_name: str,
+    tokenizer_name: str,
+):
+    tokenizer = get_tokenizer(tokenizer_name=tokenizer_name,
+                              tokenizer_mode="fast")
+
+    prompt = "This is a token_strs test prompt! vllm1"
+    response = requests.post(
+        server.url_for("tokenize"),
+        json={
+            "prompt": prompt,
+            "model": model_name,
+            "return_token_strs": True
+        },
+    )
+    response.raise_for_status()
+
+    tokens = tokenizer.encode(prompt, add_special_tokens=True)
+    tokens_str = tokenizer.convert_ids_to_tokens(tokens)
+
+    result = response.json()
+    assert result["tokens"] == tokens
+    assert result["count"] == len(tokens)
+    assert result["max_model_len"] == 8192
+    assert result["token_strs"] == tokens_str
 
 
 @pytest.mark.asyncio

vllm/entrypoints/openai/protocol.py

@@ -1563,6 +1563,11 @@ class TokenizeCompletionRequest(OpenAIBaseModel):
             "If true (the default), special tokens (e.g. BOS) will be added to "
             "the prompt."),
     )
+    return_token_strs: Optional[bool] = Field(
+        default=False,
+        description=("If true, also return the token strings "
+                     "corresponding to the token ids."),
+    )
 
 
 class TokenizeChatRequest(OpenAIBaseModel):
@@ -1576,6 +1581,11 @@ class TokenizeChatRequest(OpenAIBaseModel):
          "This is a parameter used by chat template in tokenizer config of the "
          "model."),
     )
+    return_token_strs: Optional[bool] = Field(
+        default=False,
+        description=("If true, also return the token strings "
+                     "corresponding to the token ids."),
+    )
     continue_final_message: bool = Field(
         default=False,
         description=
@@ -1633,6 +1643,7 @@ class TokenizeResponse(OpenAIBaseModel):
     count: int
     max_model_len: int
     tokens: list[int]
+    token_strs: Optional[list[str]] = None
 
 
 class DetokenizeRequest(OpenAIBaseModel):

vllm/entrypoints/openai/serving_tokenization.py

@@ -110,7 +110,12 @@ class OpenAIServingTokenization(OpenAIServing):
                           dict) and "prompt_token_ids" in engine_prompt:
                 input_ids.extend(engine_prompt["prompt_token_ids"])
 
+        token_strs = None
+        if request.return_token_strs:
+            token_strs = tokenizer.convert_ids_to_tokens(input_ids)
+
         return TokenizeResponse(tokens=input_ids,
+                                token_strs=token_strs,
                                 count=len(input_ids),
                                 max_model_len=self.max_model_len)