[Bugfix] Validate routed-expert prompt offsets before engine submission (#56844)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Bhushan Asati <bhushanasati25@gmail.com>
Co-authored-by: Megha Agarwal <19240983+meghaagr13@users.noreply.github.com>
This commit is contained in:
aoshen02
2026-09-15 09:17:41 +08:00
committed by GitHub
co-authored by Bhushan Asati Megha Agarwal
parent ef5f7cd119
commit 84e5f3edac
4 changed files with 83 additions and 10 deletions
+30
View File
@@ -12,6 +12,7 @@ import subprocess
import sys
from collections.abc import Callable
from typing import Protocol
from unittest.mock import Mock
import pytest
from pydantic import ValidationError
@@ -27,10 +28,39 @@ from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.exceptions import VLLMValidationError
from vllm.sampling_params import BeamSearchParams
from vllm.tokenizers import deepseek_v4_encoding, deepseek_v32_encoding
from vllm.v1.engine.input_processor import InputProcessor
pytestmark = [pytest.mark.cpu_test, pytest.mark.skip_global_cleanup]
@pytest.mark.parametrize("offset", [-1, 0, 2, 3, 4, 2**127])
def test_routed_experts_offset_validated_before_engine_submission(offset):
"""Reject invalid offsets even when routed-expert output is disabled."""
processor = Mock(spec=InputProcessor)
processor.tokenizer = None
processor.generation_config_fields = {}
processor.renderer = Mock()
processor.renderer.get_eos_token_id.return_value = None
processor.vllm_config = Mock()
params = SamplingParams(routed_experts_prompt_start=offset)
prompt = {"type": "token", "prompt_token_ids": [1, 2, 3]}
if not 0 <= offset <= 3:
with pytest.raises(VLLMValidationError, match="routed_experts_prompt_start"):
InputProcessor.process_inputs(
processor, "invalid", prompt, params, ("generate",)
)
params.routed_experts_prompt_start = 0
request = InputProcessor.process_inputs(
processor, "valid", prompt, params, ("generate",)
)
assert (
request.sampling_params.routed_experts_prompt_start
== params.routed_experts_prompt_start
)
# --- Stop strings: public requests cap the number of stop strings ---------