| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """End-to-end tests for the vLLM RL pause/resume lifecycle.""" |
| 4 | |
| 5 | import os |
| 6 | import threading |
| 7 | from typing import Any |
| 8 | from unittest.mock import patch |
| 9 | |
| 10 | import pytest |
| 11 | import requests |
| 12 | |
| 13 | from tests.entrypoints.serve.dev.rlhf.conftest import ( |
| 14 | cached_tokens, |
| 15 | completion_with_cache_details, |
| 16 | gen, |
| 17 | golden_output, |
| 18 | is_paused, |
| 19 | ok, |
| 20 | pause, |
| 21 | resume, |
| 22 | server, |
| 23 | start_stream, |
| 24 | ) |
| 25 | |
| 26 | |
| 27 | @pytest.fixture(scope="module", params=[False, True], ids=["MRV1", "MRV2"]) |
| 28 | def use_v2(request): |
| 29 | return request.param |
| 30 | |
| 31 | |
| 32 | @pytest.fixture(scope="module") |
| 33 | def server_url(use_v2): |
| 34 | env_vars = { |
| 35 | "VLLM_USE_V2_MODEL_RUNNER": "1" if use_v2 else "0", |
| 36 | } |
| 37 | |
| 38 | with ( |
| 39 | patch.dict(os.environ, env_vars), |
| 40 | server( |
| 41 | extra_args=[ |
| 42 | "--enable-prefix-caching", |
| 43 | "--enable-prompt-tokens-details", |
| 44 | ] |
| 45 | ) as url, |
| 46 | ): |
| 47 | yield url |
| 48 | |
| 49 | |
| 50 | @pytest.fixture(autouse=True) |
| 51 | def restore_unpaused_state(server_url): |
| 52 | assert resume(server_url) == 200 |
| 53 | yield |
| 54 | assert resume(server_url) == 200 |
| 55 | |
| 56 | |
| 57 | class TestPauseResume: |
| 58 | def test_state_and_idempotency_across_cycles(self, server_url): |
| 59 | assert not is_paused(server_url) |
| 60 | |
| 61 | assert resume(server_url) == 200 |
| 62 | assert resume(server_url) == 200 |
| 63 | assert not is_paused(server_url) |
| 64 | |
| 65 | for _ in range(2): |
| 66 | for mode in ("abort", "wait", "keep"): |
| 67 | assert pause(server_url, mode=mode) == 200 |
| 68 | assert pause(server_url, mode=mode) == 200 |
| 69 | assert is_paused(server_url) |
| 70 | |
| 71 | assert resume(server_url) == 200 |
| 72 | assert resume(server_url) == 200 |
| 73 | assert not is_paused(server_url) |
| 74 | |
| 75 | def test_invalid_mode_preserves_state(self, server_url): |
| 76 | for paused in (False, True): |
| 77 | if paused: |
| 78 | assert pause(server_url) == 200 |
| 79 | assert is_paused(server_url) is paused |
| 80 | |
| 81 | response = requests.post( |
| 82 | f"{server_url}/pause", |
| 83 | params={"mode": "invalid"}, |
| 84 | timeout=10, |
| 85 | ) |
| 86 | assert response.status_code == 400 |
| 87 | assert response.json()["error"]["param"] == "query.mode" |
| 88 | assert is_paused(server_url) is paused |
| 89 | |
| 90 | assert resume(server_url) == 200 |
| 91 | |
| 92 | @pytest.mark.parametrize( |
| 93 | ("mode", "max_tokens", "inflight_finish_reason"), |
| 94 | [ |
| 95 | pytest.param("abort", 256, "abort", id="abort"), |
| 96 | pytest.param("wait", 256, "length", id="wait"), |
| 97 | pytest.param("keep", 256, "length", id="keep"), |
| 98 | ], |
| 99 | ) |
| 100 | def test_mode_request_lifecycle( |
| 101 | self, |
| 102 | server_url, |
| 103 | mode, |
| 104 | max_tokens, |
| 105 | inflight_finish_reason, |
| 106 | ): |
| 107 | inflight, inflight_thread = start_stream(server_url, max_tokens) |
| 108 | new_result: dict[str, Any] = {} |
| 109 | new_done = threading.Event() |
| 110 | |
| 111 | def _new_request(): |
| 112 | new_result["response"] = gen(server_url, max_tokens=4, timeout=60) |
| 113 | new_done.set() |
| 114 | |
| 115 | new_thread = threading.Thread(target=_new_request) |
| 116 | try: |
| 117 | assert pause(server_url, mode=mode) == 200 |
| 118 | assert is_paused(server_url) |
| 119 | |
| 120 | if mode in ("abort", "wait"): |
| 121 | assert inflight.done.is_set() |
| 122 | else: |
| 123 | chunks_after_pause = len(inflight.chunks) |
| 124 | assert not inflight.done.wait(timeout=5) |
| 125 | assert len(inflight.chunks) == chunks_after_pause, ( |
| 126 | "in-flight request continued generating in keep mode" |
| 127 | ) |
| 128 | |
| 129 | new_thread.start() |
| 130 | assert not new_done.wait(timeout=0.3), ( |
| 131 | "new request completed while generation was paused" |
| 132 | ) |
| 133 | finally: |
| 134 | assert resume(server_url) == 200 |
| 135 | inflight_thread.join(timeout=30) |
| 136 | if new_thread.ident is not None: |
| 137 | new_thread.join(timeout=30) |
| 138 | |
| 139 | assert not inflight_thread.is_alive() |
| 140 | assert inflight.error is None |
| 141 | assert inflight.finish_reason == inflight_finish_reason |
| 142 | assert not new_thread.is_alive() |
| 143 | assert ok(new_result.get("response")) |
| 144 | |
| 145 | def test_clear_cache_preserves_output_and_controls_prefix_cache(self, server_url): |
| 146 | prompt = ( |
| 147 | "Paris is the capital of France. Berlin is the capital of Germany. " |
| 148 | ) * 20 |
| 149 | |
| 150 | assert pause(server_url, clear_cache=True) == 200 |
| 151 | assert resume(server_url) == 200 |
| 152 | baseline = completion_with_cache_details(server_url, prompt) |
| 153 | warmed = completion_with_cache_details(server_url, prompt) |
| 154 | |
| 155 | assert cached_tokens(baseline) == 0 |
| 156 | assert cached_tokens(warmed) > 0 |
| 157 | |
| 158 | assert pause(server_url, clear_cache=False) == 200 |
| 159 | assert resume(server_url) == 200 |
| 160 | preserved = completion_with_cache_details(server_url, prompt) |
| 161 | assert golden_output(preserved) == golden_output(baseline) |
| 162 | assert cached_tokens(preserved) > 0 |
| 163 | |
| 164 | assert pause(server_url, clear_cache=True) == 200 |
| 165 | assert resume(server_url) == 200 |
| 166 | cleared = completion_with_cache_details(server_url, prompt) |
| 167 | assert golden_output(cleared) == golden_output(baseline) |
| 168 | assert cached_tokens(cleared) == 0 |