| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from unittest.mock import MagicMock, patch |
| 5 | |
| 6 | import pytest |
| 7 | |
| 8 | from tests.tool_parsers.utils import ( |
| 9 | run_tool_extraction, |
| 10 | run_tool_extraction_streaming, |
| 11 | ) |
| 12 | from vllm.entrypoints.generate.base.protocol import FunctionCall |
| 13 | from vllm.tokenizers import TokenizerLike |
| 14 | from vllm.tool_parsers import ToolParser, ToolParserManager |
| 15 | |
| 16 | # https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/text_prompt_format.md#model-response-format-1 |
| 17 | SIMPLE_FUNCTION_OUTPUT = "get_weather(city='San Francisco', metric='celsius')" |
| 18 | SIMPLE_FUNCTION_CALL = FunctionCall( |
| 19 | name="get_weather", |
| 20 | arguments='{"city": "San Francisco", "metric": "celsius"}', |
| 21 | ) |
| 22 | MORE_TYPES_FUNCTION_OUTPUT = ( |
| 23 | "register_user(name='John Doe', " |
| 24 | "age=37, " |
| 25 | "address={'city': 'San Francisco', 'state': 'CA'}, " |
| 26 | "role=None, " |
| 27 | "passed_test=True, " |
| 28 | "aliases=['John', 'Johnny'])" |
| 29 | ) |
| 30 | MORE_TYPES_FUNCTION_CALL = FunctionCall( |
| 31 | name="register_user", |
| 32 | arguments='{"name": "John Doe", ' |
| 33 | '"age": 37, ' |
| 34 | '"address": {"city": "San Francisco", "state": "CA"}, ' |
| 35 | '"role": null, ' |
| 36 | '"passed_test": true, ' |
| 37 | '"aliases": ["John", "Johnny"]}', |
| 38 | ) |
| 39 | PARAMETERLESS_FUNCTION_OUTPUT = "get_weather()" |
| 40 | PARAMETERLESS_FUNCTION_CALL = FunctionCall( |
| 41 | name="get_weather", |
| 42 | arguments="{}", |
| 43 | ) |
| 44 | EMPTY_DICT_FUNCTION_OUTPUT = "do_something_cool(additional_data={})" |
| 45 | EMPTY_DICT_FUNCTION_CALL = FunctionCall( |
| 46 | name="do_something_cool", |
| 47 | arguments='{"additional_data": {}}', |
| 48 | ) |
| 49 | EMPTY_LIST_FUNCTION_OUTPUT = "do_something_cool(steps=[])" |
| 50 | EMPTY_LIST_FUNCTION_CALL = FunctionCall( |
| 51 | name="do_something_cool", |
| 52 | arguments='{"steps": []}', |
| 53 | ) |
| 54 | ESCAPED_STRING_FUNCTION_OUTPUT = ( |
| 55 | r"get_weather(city='Martha\'s Vineyard', metric='\"cool units\"')" |
| 56 | ) |
| 57 | ESCAPED_STRING_FUNCTION_CALL = FunctionCall( |
| 58 | name="get_weather", |
| 59 | arguments='{"city": "Martha\'s Vineyard", "metric": "\\"cool units\\""}', |
| 60 | ) |
| 61 | |
| 62 | |
| 63 | @pytest.mark.parametrize("streaming", [True, False]) |
| 64 | def test_no_tool_call(streaming: bool, default_tokenizer: TokenizerLike): |
| 65 | tool_parser: ToolParser = ToolParserManager.get_tool_parser("pythonic")( |
| 66 | default_tokenizer |
| 67 | ) |
| 68 | model_output = "How can I help you today?" |
| 69 | |
| 70 | content, tool_calls = run_tool_extraction( |
| 71 | tool_parser, model_output, streaming=streaming |
| 72 | ) |
| 73 | |
| 74 | assert content == model_output |
| 75 | assert len(tool_calls) == 0 |
| 76 | |
| 77 | |
| 78 | TEST_CASES = [ |
| 79 | pytest.param( |
| 80 | True, |
| 81 | f"[{SIMPLE_FUNCTION_OUTPUT}]", |
| 82 | [SIMPLE_FUNCTION_CALL], |
| 83 | id="simple_streaming", |
| 84 | ), |
| 85 | pytest.param( |
| 86 | False, |
| 87 | f"[{SIMPLE_FUNCTION_OUTPUT}]", |
| 88 | [SIMPLE_FUNCTION_CALL], |
| 89 | id="simple_nonstreaming", |
| 90 | ), |
| 91 | pytest.param( |
| 92 | True, |
| 93 | f"[{MORE_TYPES_FUNCTION_OUTPUT}]", |
| 94 | [MORE_TYPES_FUNCTION_CALL], |
| 95 | id="more_types_streaming", |
| 96 | ), |
| 97 | pytest.param( |
| 98 | False, |
| 99 | f"[{MORE_TYPES_FUNCTION_OUTPUT}]", |
| 100 | [MORE_TYPES_FUNCTION_CALL], |
| 101 | id="more_types_nonstreaming", |
| 102 | ), |
| 103 | pytest.param( |
| 104 | True, |
| 105 | f"[{PARAMETERLESS_FUNCTION_OUTPUT}]", |
| 106 | [PARAMETERLESS_FUNCTION_CALL], |
| 107 | id="parameterless_streaming", |
| 108 | ), |
| 109 | pytest.param( |
| 110 | False, |
| 111 | f"[{PARAMETERLESS_FUNCTION_OUTPUT}]", |
| 112 | [PARAMETERLESS_FUNCTION_CALL], |
| 113 | id="parameterless_nonstreaming", |
| 114 | ), |
| 115 | pytest.param( |
| 116 | True, |
| 117 | f"[{EMPTY_DICT_FUNCTION_OUTPUT}]", |
| 118 | [EMPTY_DICT_FUNCTION_CALL], |
| 119 | id="empty_dict_streaming", |
| 120 | ), |
| 121 | pytest.param( |
| 122 | False, |
| 123 | f"[{EMPTY_DICT_FUNCTION_OUTPUT}]", |
| 124 | [EMPTY_DICT_FUNCTION_CALL], |
| 125 | id="empty_dict_nonstreaming", |
| 126 | ), |
| 127 | pytest.param( |
| 128 | True, |
| 129 | f"[{EMPTY_LIST_FUNCTION_OUTPUT}]", |
| 130 | [EMPTY_LIST_FUNCTION_CALL], |
| 131 | id="empty_list_streaming", |
| 132 | ), |
| 133 | pytest.param( |
| 134 | False, |
| 135 | f"[{EMPTY_LIST_FUNCTION_OUTPUT}]", |
| 136 | [EMPTY_LIST_FUNCTION_CALL], |
| 137 | id="empty_list_nonstreaming", |
| 138 | ), |
| 139 | pytest.param( |
| 140 | True, |
| 141 | f"[{ESCAPED_STRING_FUNCTION_OUTPUT}]", |
| 142 | [ESCAPED_STRING_FUNCTION_CALL], |
| 143 | id="escaped_string_streaming", |
| 144 | ), |
| 145 | pytest.param( |
| 146 | False, |
| 147 | f"[{ESCAPED_STRING_FUNCTION_OUTPUT}]", |
| 148 | [ESCAPED_STRING_FUNCTION_CALL], |
| 149 | id="escaped_string_nonstreaming", |
| 150 | ), |
| 151 | pytest.param( |
| 152 | True, |
| 153 | f"[{SIMPLE_FUNCTION_OUTPUT}, {MORE_TYPES_FUNCTION_OUTPUT}]", |
| 154 | [SIMPLE_FUNCTION_CALL, MORE_TYPES_FUNCTION_CALL], |
| 155 | id="parallel_calls_streaming", |
| 156 | ), |
| 157 | pytest.param( |
| 158 | False, |
| 159 | f"[{SIMPLE_FUNCTION_OUTPUT}, {MORE_TYPES_FUNCTION_OUTPUT}]", |
| 160 | [SIMPLE_FUNCTION_CALL, MORE_TYPES_FUNCTION_CALL], |
| 161 | id="parallel_calls_nonstreaming", |
| 162 | ), |
| 163 | ] |
| 164 | |
| 165 | |
| 166 | @pytest.mark.parametrize("streaming, model_output, expected_tool_calls", TEST_CASES) |
| 167 | def test_tool_call( |
| 168 | streaming: bool, |
| 169 | model_output: str, |
| 170 | expected_tool_calls: list[FunctionCall], |
| 171 | default_tokenizer: TokenizerLike, |
| 172 | ): |
| 173 | tool_parser: ToolParser = ToolParserManager.get_tool_parser("pythonic")( |
| 174 | default_tokenizer |
| 175 | ) |
| 176 | |
| 177 | content, tool_calls = run_tool_extraction( |
| 178 | tool_parser, model_output, streaming=streaming |
| 179 | ) |
| 180 | |
| 181 | assert content is None |
| 182 | assert len(tool_calls) == len(expected_tool_calls) |
| 183 | for actual, expected in zip(tool_calls, expected_tool_calls): |
| 184 | assert actual.type == "function" |
| 185 | assert actual.function == expected |
| 186 | |
| 187 | |
| 188 | def test_streaming_tool_call_with_large_steps(default_tokenizer: TokenizerLike): |
| 189 | tool_parser: ToolParser = ToolParserManager.get_tool_parser("pythonic")( |
| 190 | default_tokenizer |
| 191 | ) |
| 192 | model_output_deltas = [ |
| 193 | "[get_weather(city='San", |
| 194 | " Francisco', metric='celsius'), " |
| 195 | f"{PARAMETERLESS_FUNCTION_OUTPUT}, " |
| 196 | f"{EMPTY_LIST_FUNCTION_OUTPUT}]", |
| 197 | ] |
| 198 | |
| 199 | reconstructor = run_tool_extraction_streaming( |
| 200 | tool_parser, model_output_deltas, assert_one_tool_per_delta=False |
| 201 | ) |
| 202 | |
| 203 | assert reconstructor.other_content == "" |
| 204 | assert len(reconstructor.tool_calls) == 3 |
| 205 | assert reconstructor.tool_calls[0].function == SIMPLE_FUNCTION_CALL |
| 206 | assert reconstructor.tool_calls[1].function == PARAMETERLESS_FUNCTION_CALL |
| 207 | assert reconstructor.tool_calls[2].function == EMPTY_LIST_FUNCTION_CALL |
| 208 | |
| 209 | |
| 210 | @pytest.mark.parametrize("streaming", [False]) |
| 211 | def test_regex_timeout_handling(streaming: bool, default_tokenizer: TokenizerLike): |
| 212 | """Test regex timeout is handled gracefully.""" |
| 213 | tool_parser: ToolParser = ToolParserManager.get_tool_parser("pythonic")( |
| 214 | default_tokenizer |
| 215 | ) |
| 216 | |
| 217 | fake_problematic_input = "hello world[A(A=" + "\t)A(A=,\t" * 2 |
| 218 | |
| 219 | # create a mock regex that raises TimeoutError |
| 220 | mock_regex = MagicMock() |
| 221 | mock_regex.match.side_effect = TimeoutError("Regex timeout") |
| 222 | |
| 223 | with patch.object(tool_parser, "TOOL_CALL_REGEX", mock_regex): |
| 224 | content, tool_calls = run_tool_extraction( |
| 225 | tool_parser, fake_problematic_input, streaming=streaming |
| 226 | ) |
| 227 | |
| 228 | # should treat as regular text when regex times out |
| 229 | assert content == fake_problematic_input |
| 230 | assert len(tool_calls) == 0 |
| 231 | mock_regex.match.assert_called_once() |