[BugFix] --max-model-len=-1 causes over-limit requests to hang and starve the entire service (#39102)

Signed-off-by: triangle14 <y1019026570@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>

tests/v1/e2e/general/test_context_length.py

@@ -15,6 +15,7 @@ import pytest
 
 from tests.conftest import VllmRunner
 from tests.utils import create_new_process_for_each_test
+from vllm.exceptions import VLLMValidationError
 
 
 @create_new_process_for_each_test()
@@ -61,3 +62,42 @@ def test_decoder_max_context_length_validation(
             with pytest.raises(ValueError) as excinfo:
                 vllm_model.generate_greedy(prompt_ids, max_tokens)
             assert expected_msg in str(excinfo.value)
+
+
+@create_new_process_for_each_test()
+@pytest.mark.parametrize("model", ["JackFram/llama-160m"])
+def test_auto_fit_max_model_len_rejects_oversized_input(
+    model: str,
+    vllm_runner: type[VllmRunner],
+) -> None:
+    """When max_model_len='auto' and KV cache memory is very limited,
+    the engine auto-fits max_model_len to a small value. The frontend
+    must see this reduced value and reject prompts that exceed it,
+    rather than accepting them and hanging."""
+
+    # Use a tiny KV cache budget to force auto-fit to a very small
+    # max_model_len (e.g. ~16 tokens).
+    kv_cache_bytes = 1_000_000  # 1 MB
+
+    with vllm_runner(
+        model_name=model,
+        max_model_len=-1,
+        max_num_seqs=1,
+        enforce_eager=True,
+        kv_cache_memory_bytes=kv_cache_bytes,
+        load_format="dummy",
+    ) as vllm_model:
+        auto_fitted_len = (
+            vllm_model.llm.llm_engine.vllm_config.model_config.max_model_len
+        )
+        # Sanity check: auto-fit should have reduced it well below the
+        # model's native context length.
+        assert auto_fitted_len < 2048, (
+            f"Expected auto-fit to reduce max_model_len significantly, "
+            f"but got {auto_fitted_len}"
+        )
+
+        # A prompt longer than the auto-fitted length must be rejected.
+        oversized_prompt = [[43] * (auto_fitted_len + 10)]
+        with pytest.raises(VLLMValidationError, match="Please reduce the length"):
+            vllm_model.generate_greedy(oversized_prompt, max_tokens=4)

tests/v1/engine/test_engine_core_client.py

@@ -114,7 +114,7 @@ def test_mp_client_uses_env_timeout(monkeypatch: pytest.MonkeyPatch):
             return 1
 
         def recv_multipart(self):
-            return (b"\x00\x00", b"ready")
+            return (b"\x00\x00", b"")
 
     class DummySocket:
         def send_multipart(self, _msg, *, copy: bool = False, track: bool = False):

vllm/v1/engine/__init__.py

@@ -63,6 +63,20 @@ class FinishReason(enum.IntEnum):
         return FINISH_REASON_STRINGS[self.value]
 
 
+class EngineCoreReadyResponse(
+    msgspec.Struct,
+    array_like=True,  # type: ignore[call-arg]
+    omit_defaults=True,  # type: ignore[call-arg]
+):
+    """Sent from EngineCore to the frontend during the ready handshake.
+
+    Contains post-initialization config that may differ from the original
+    values (e.g. max_model_len after KV cache auto-fitting).
+    """
+
+    max_model_len: int | None = None
+
+
 class EngineCoreRequest(
     msgspec.Struct,
     array_like=True,  # type: ignore[call-arg]

vllm/v1/engine/core.py

@@ -52,6 +52,7 @@ from vllm.v1.engine import (
     EEPNotificationType,
     EngineCoreOutput,
     EngineCoreOutputs,
+    EngineCoreReadyResponse,
     EngineCoreRequest,
     EngineCoreRequestType,
     FinishReason,
@@ -1385,11 +1386,15 @@ class EngineCoreProc(EngineCore):
 
             # Register sockets with poller.
             poller = zmq.Poller()
+            ready_response = EngineCoreReadyResponse(
+                max_model_len=self.vllm_config.model_config.max_model_len,
+            )
+            ready_payload = msgspec.msgpack.encode(ready_response)
             for input_socket in input_sockets:
                 # Send initial message to each input socket - this is required
                 # before the front-end ROUTER socket can send input messages
                 # back to us.
-                input_socket.send(b"")
+                input_socket.send(ready_payload)
                 poller.register(input_socket, zmq.POLLIN)
 
             if coord_socket is not None:

vllm/v1/engine/core_client.py

@@ -35,6 +35,7 @@ from vllm.v1.engine import (
     EEP_NOTIFICATION_CALL_ID,
     EEPNotificationType,
     EngineCoreOutputs,
+    EngineCoreReadyResponse,
     EngineCoreRequest,
     EngineCoreRequestType,
     PauseMode,
@@ -456,6 +457,22 @@ class ElasticScalingCache:
     pending_notifications: dict[EEPNotificationType, set[int]]
 
 
+_ready_response_decoder = msgspec.msgpack.Decoder(EngineCoreReadyResponse)
+
+
+def _apply_ready_response(payload: bytes, vllm_config: VllmConfig) -> None:
+    """Decode an EngineCoreReadyResponse and sync any post-initialization
+    config changes (e.g. auto-fitted max_model_len) back to the frontend."""
+    if not payload:
+        return
+    response = _ready_response_decoder.decode(payload)
+    if response.max_model_len is not None:
+        vllm_config.model_config.max_model_len = min(
+            vllm_config.model_config.max_model_len,
+            response.max_model_len,
+        )
+
+
 class MPClient(EngineCoreClient):
     """
     MPClient: base client for multi-proc EngineCore.
@@ -589,8 +606,9 @@ class MPClient(EngineCoreClient):
                         f"timeout, set the environment variable: "
                         f"VLLM_ENGINE_READY_TIMEOUT_S=<seconds>"
                     )
-                identity, _ = sync_input_socket.recv_multipart()
+                identity, payload = sync_input_socket.recv_multipart()
                 identities.remove(identity)
+                _apply_ready_response(payload, vllm_config)
 
             self.core_engine: EngineIdentity = self.core_engines[0]
             self.utility_results: dict[int, AnyFuture] = {}
@@ -1582,8 +1600,9 @@ class DPLBAsyncMPClient(DPAsyncMPClient):
                     f"timeout, set the environment variable: "
                     f"VLLM_ENGINE_READY_TIMEOUT_S=<seconds>"
                 )
-            identity, _ = sync_input_socket.recv_multipart()
+            identity, payload = sync_input_socket.recv_multipart()
             new_engine_identities.discard(identity)
+            _apply_ready_response(payload, self.vllm_config)
 
         # NOTE(yongji): Before we schedule any requests on the new workers,
         # we should wait for them to switch to the new setup.