[Bugfix][Frontend] Guard against bad token ids (#9634)

Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>

tests/entrypoints/openai/test_completion.py

@@ -157,15 +157,15 @@ async def test_added_lora_tokens(client: openai.AsyncOpenAI):
 @pytest.mark.asyncio
 async def test_added_lora_tokens_base_model(client: openai.AsyncOpenAI):
     # test using token IDs
-    completion = await client.completions.create(
-        model=MODEL_NAME,
-        prompt=[0, 0, 32000, 32001, 32002],
-        echo=True,
-        max_tokens=5,
-        temperature=0.0,
-    )
-    # Added tokens should not appear in tokenized prompt
-    assert "vllm" not in completion.choices[0].text
+    with pytest.raises(openai.BadRequestError, match="out of vocabulary"):
+        # Added tokens should be rejected by the base model
+        await client.completions.create(
+            model=MODEL_NAME,
+            prompt=[0, 0, 32000, 32001, 32002],
+            echo=True,
+            max_tokens=5,
+            temperature=0.0,
+        )
 
 
 @pytest.mark.asyncio

tests/entrypoints/openai/test_prompt_validation.py

@@ -20,3 +20,18 @@ async def test_empty_prompt():
                                             prompt="",
                                             max_tokens=5,
                                             temperature=0.0)
+
+
+@pytest.mark.asyncio
+async def test_out_of_vocab_token_ids():
+    model_name = "gpt2"
+    server_args = ["--enforce-eager"]
+    with RemoteOpenAIServer(model_name, server_args) as remote_server:
+        client = remote_server.get_async_client()
+
+        with pytest.raises(openai.BadRequestError,
+                           match=re.compile('.*out of vocabulary.*')):
+            await client.completions.create(model=model_name,
+                                            prompt=[999999],
+                                            max_tokens=5,
+                                            temperature=0.0)