add test

Signed-off-by: sasha0552 <[email protected]>

tests/prefix_caching/test_prefix_caching.py

@@ -5,13 +5,22 @@
 import pytest
 
 from tests.kernels.utils import override_backend_env_variable
+from vllm import SamplingParams, TokensPrompt
 
 from ..models.utils import check_outputs_equal
 
 MODELS = [
     "facebook/opt-125m",
 ]
 
+UNSTABLE_PROMPT_SEQUENCE = [
+    ([0] * 588) + ([1] * 1332) + ([2] * 30) + ([3] * 1),
+    ([0] * 588) + ([1] * 1332) + ([4] * 3) + ([5] * 50),
+    ([0] * 588) + ([1] * 1332) + ([2] * 30) + ([6] * 95),
+    ([0] * 588) + ([1] * 1332) + ([4] * 3) + ([7] * 174),
+    ([0] * 588) + ([8] * 1539),
+]
+
 
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
@@ -57,3 +66,22 @@ def test_mixed_requests(
         name_0="hf",
         name_1="vllm",
     )
+
+
+@pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
+def test_unstable_prompt_sequence(
+    vllm_runner,
+    backend: str,
+    monkeypatch,
+) -> None:
+    override_backend_env_variable(monkeypatch, backend)
+
+    with vllm_runner(
+            enable_chunked_prefill=True,
+            enable_prefix_caching=True,
+            max_model_len=4096,
+            model="Qwen/Qwen2.5-0.5B-Instruct",
+    ) as vllm_model:
+        for prompt in UNSTABLE_PROMPT_SEQUENCE:
+            vllm_model.generate(TokensPrompt(prompt_token_ids=prompt),
+                                SamplingParams(max_tokens=1))