[Bugfix][Frontend] Reject stop strings on --tokens-only servers instead of silently ignoring them (#57058)

Signed-off-by: Shimi Bandiel <shimib@google.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

tests/entrypoints/scale_out/token_in_token_out/test_serving_tokens.py

@@ -422,6 +422,31 @@ async def test_stop_string_workflow(client, tokenizer, messages):
     assert generate_res == completions_res
 
 
+@pytest.mark.asyncio
+@pytest.mark.skipif(
+    envs.VLLM_USE_RUST_FRONTEND,
+    reason="--tokens-only is not supported by the Rust frontend",
+)
+@pytest.mark.parametrize("server", [["--tokens-only"]], indirect=True)
+async def test_stop_strings_rejected_when_tokens_only(client):
+    """--tokens-only forces detokenize=False after request validation, so
+    stop strings must be rejected explicitly rather than silently ignored."""
+    sampling_params = {"max_tokens": 5, "stop": ["never"]}
+    payload = {
+        "model": MODEL_NAME,
+        "token_ids": [1, 2, 3],
+        "sampling_params": sampling_params,
+        "stream": False,
+    }
+    resp = await client.post(GEN_ENDPOINT, json=payload)
+    assert resp.status_code == 400
+    assert "stop strings" in resp.json()["error"]["message"]
+
+    sampling_params["stop"] = None
+    resp = await client.post(GEN_ENDPOINT, json=payload)
+    resp.raise_for_status()
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
     "server",

vllm/entrypoints/scale_out/token_in_token_out/serving.py

@@ -134,6 +134,16 @@ class ServingTokens(GenerateBaseServing):
                 f"sampling_params.n must be at most the server's max_num_seqs "
                 f"({max_num_seqs}), got {sampling_params.n}."
             )
+        if self.force_no_detokenize and sampling_params.stop:
+            # SamplingParams rejects stop with detokenize=False at request
+            # validation, but this server forces detokenize=False afterwards,
+            # so the combination must be rejected here or stop strings are
+            # silently never applied.
+            return self.create_error_response(
+                "stop strings are not supported on a --tokens-only server "
+                "because detokenization is disabled. Check stop strings on "
+                "the coordinator, or use stop_token_ids."
+            )
         try:
             msgspec.msgpack.encode(
                 (