| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for DeepSeek V3.2 parser engine semantics. |
| 4 | |
| 5 | V3.2 uses the same DSML parameter format as V4 but wraps tool calls in |
| 6 | ``<｜DSML｜function_calls>`` instead of ``<｜DSML｜tool_calls>`` and has |
| 7 | no reasoning (``<think>``/``</think>``) support. |
| 8 | """ |
| 9 | |
| 10 | import json |
| 11 | |
| 12 | import pytest |
| 13 | |
| 14 | from tests.parser.engine.conftest import make_mock_tokenizer |
| 15 | from tests.parser.engine.streaming_helpers import ( |
| 16 | collect_content, |
| 17 | collect_function_name, |
| 18 | collect_tool_arguments, |
| 19 | simulate_tool_streaming, |
| 20 | ) |
| 21 | from vllm.parser.deepseek_v4 import ( |
| 22 | DSML_INVOKE_END, |
| 23 | DSML_INVOKE_NAME_END, |
| 24 | DSML_INVOKE_PREFIX, |
| 25 | ) |
| 26 | from vllm.parser.deepseek_v32 import ( |
| 27 | DSML_FUNC_END, |
| 28 | DSML_FUNC_START, |
| 29 | DeepSeekV32Parser, |
| 30 | ) |
| 31 | from vllm.parser.engine.parser_engine_config import ParserState |
| 32 | |
| 33 | _PARAM_OPEN = '｜DSML｜parameter name="{name}" string="{is_str}">' |
| 34 | _PARAM_CLOSE = "</｜DSML｜parameter>" |
| 35 | |
| 36 | |
| 37 | def _param(name: str, is_str: str, value: str) -> str: |
| 38 | return f"<{_PARAM_OPEN.format(name=name, is_str=is_str)}{value}{_PARAM_CLOSE}" |
| 39 | |
| 40 | |
| 41 | def _invoke(name: str, *params: str) -> str: |
| 42 | body = "\n".join(params) |
| 43 | return ( |
| 44 | f"{DSML_INVOKE_PREFIX}{name}{DSML_INVOKE_NAME_END}\n{body}\n{DSML_INVOKE_END}" |
| 45 | ) |
| 46 | |
| 47 | |
| 48 | def _func_calls(*invocations: str) -> str: |
| 49 | body = "\n".join(invocations) |
| 50 | return f"{DSML_FUNC_START}\n{body}\n{DSML_FUNC_END}" |
| 51 | |
| 52 | |
| 53 | def _make_tool(name, properties): |
| 54 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 55 | ChatCompletionToolsParam, |
| 56 | ) |
| 57 | |
| 58 | return ChatCompletionToolsParam( |
| 59 | type="function", |
| 60 | function={ |
| 61 | "name": name, |
| 62 | "parameters": { |
| 63 | "type": "object", |
| 64 | "properties": properties, |
| 65 | }, |
| 66 | }, |
| 67 | ) |
| 68 | |
| 69 | |
| 70 | @pytest.fixture |
| 71 | def mock_tokenizer(): |
| 72 | return make_mock_tokenizer({}) |
| 73 | |
| 74 | |
| 75 | @pytest.fixture |
| 76 | def mock_request(): |
| 77 | from unittest.mock import MagicMock |
| 78 | |
| 79 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 80 | ChatCompletionRequest, |
| 81 | ) |
| 82 | |
| 83 | req = MagicMock(spec=ChatCompletionRequest) |
| 84 | req.tools = [] |
| 85 | req.tool_choice = "auto" |
| 86 | return req |
| 87 | |
| 88 | |
| 89 | # ── Non-streaming extraction ──────────────────────────────────────── |
| 90 | |
| 91 | |
| 92 | class TestNonStreaming: |
| 93 | def test_no_tool_call(self, mock_tokenizer, mock_request): |
| 94 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 95 | result = parser.extract_tool_calls("Hello world", mock_request) |
| 96 | assert not result.tools_called |
| 97 | assert result.content == "Hello world" |
| 98 | |
| 99 | def test_single_tool(self, mock_tokenizer, mock_request): |
| 100 | text = _func_calls( |
| 101 | _invoke("get_weather", _param("city", "true", "SF")), |
| 102 | ) |
| 103 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 104 | result = parser.extract_tool_calls(text, mock_request) |
| 105 | assert result.tools_called |
| 106 | assert len(result.tool_calls) == 1 |
| 107 | assert result.tool_calls[0].function.name == "get_weather" |
| 108 | args = json.loads(result.tool_calls[0].function.arguments) |
| 109 | assert args == {"city": "SF"} |
| 110 | |
| 111 | def test_parallel_tools(self, mock_tokenizer, mock_request): |
| 112 | text = _func_calls( |
| 113 | _invoke("get_weather", _param("city", "true", "SF")), |
| 114 | _invoke("get_weather", _param("city", "true", "NYC")), |
| 115 | ) |
| 116 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 117 | result = parser.extract_tool_calls(text, mock_request) |
| 118 | assert result.tools_called |
| 119 | assert len(result.tool_calls) == 2 |
| 120 | assert json.loads(result.tool_calls[0].function.arguments) == {"city": "SF"} |
| 121 | assert json.loads(result.tool_calls[1].function.arguments) == {"city": "NYC"} |
| 122 | |
| 123 | def test_content_before_tool_call(self, mock_tokenizer, mock_request): |
| 124 | text = "Let me check. " + _func_calls( |
| 125 | _invoke("search", _param("q", "true", "vllm")), |
| 126 | ) |
| 127 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 128 | result = parser.extract_tool_calls(text, mock_request) |
| 129 | assert result.tools_called |
| 130 | assert result.content is not None |
| 131 | assert "Let me check" in result.content |
| 132 | |
| 133 | def test_non_string_params_json_parsed(self, mock_tokenizer, mock_request): |
| 134 | text = _func_calls( |
| 135 | _invoke( |
| 136 | "toggle", |
| 137 | _param("enabled", "false", "true"), |
| 138 | _param("count", "false", "42"), |
| 139 | ), |
| 140 | ) |
| 141 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 142 | result = parser.extract_tool_calls(text, mock_request) |
| 143 | args = json.loads(result.tool_calls[0].function.arguments) |
| 144 | assert args["enabled"] is True |
| 145 | assert args["count"] == 42 |
| 146 | |
| 147 | def test_wrapper_unwrapping(self, mock_tokenizer, mock_request): |
| 148 | tool = _make_tool("get_weather", {"location": {"type": "string"}}) |
| 149 | mock_request.tools = [tool] |
| 150 | text = _func_calls( |
| 151 | _invoke( |
| 152 | "get_weather", |
| 153 | _param("arguments", "false", '{"location":"Beijing"}'), |
| 154 | ), |
| 155 | ) |
| 156 | parser = DeepSeekV32Parser(mock_tokenizer, tools=[tool]) |
| 157 | result = parser.extract_tool_calls(text, mock_request) |
| 158 | args = json.loads(result.tool_calls[0].function.arguments) |
| 159 | assert args == {"location": "Beijing"} |
| 160 | |
| 161 | |
| 162 | # ── Initial state ──────────────────────────────────────────────────── |
| 163 | |
| 164 | |
| 165 | class TestInitialState: |
| 166 | def test_always_content(self, mock_tokenizer): |
| 167 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 168 | cfg = parser.parser_engine_config |
| 169 | assert cfg.initial_state == ParserState.CONTENT |
| 170 | |
| 171 | def test_ignores_thinking_kwargs(self, mock_tokenizer): |
| 172 | parser = DeepSeekV32Parser( |
| 173 | mock_tokenizer, |
| 174 | chat_template_kwargs={"thinking": True, "enable_thinking": True}, |
| 175 | ) |
| 176 | cfg = parser.parser_engine_config |
| 177 | assert cfg.initial_state == ParserState.CONTENT |
| 178 | |
| 179 | |
| 180 | # ── Streaming ──────────────────────────────────────────────────────── |
| 181 | |
| 182 | |
| 183 | class TestStreaming: |
| 184 | def test_split_parameter_start_is_buffered(self, mock_tokenizer, mock_request): |
| 185 | chunks = [ |
| 186 | DSML_FUNC_START, |
| 187 | f"{DSML_INVOKE_PREFIX}edit{DSML_INVOKE_NAME_END}", |
| 188 | f"<{_PARAM_OPEN.format(name='expr', is_str='true')}a<b><｜DSML｜parameter", |
| 189 | ' name="date" string="true">tomorrow', |
| 190 | _PARAM_CLOSE, |
| 191 | DSML_INVOKE_END, |
| 192 | DSML_FUNC_END, |
| 193 | ] |
| 194 | |
| 195 | results = simulate_tool_streaming( |
| 196 | DeepSeekV32Parser(mock_tokenizer), mock_request, chunks |
| 197 | ) |
| 198 | |
| 199 | arguments = collect_tool_arguments(results) |
| 200 | assert "<｜DSML｜parameter" not in arguments |
| 201 | assert json.loads(arguments) == {"expr": "a<b>", "date": "tomorrow"} |
| 202 | |
| 203 | def test_single_tool_streaming(self, mock_tokenizer, mock_request): |
| 204 | text = _func_calls( |
| 205 | _invoke("get_weather", _param("city", "true", "SF")), |
| 206 | ) |
| 207 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 208 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 209 | assert collect_function_name(results) == "get_weather" |
| 210 | args_json = collect_tool_arguments(results) |
| 211 | assert json.loads(args_json) == {"city": "SF"} |
| 212 | |
| 213 | def test_content_before_tool_streaming(self, mock_tokenizer, mock_request): |
| 214 | text = "Checking... " + _func_calls( |
| 215 | _invoke("fn", _param("k", "true", "v")), |
| 216 | ) |
| 217 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 218 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 219 | content = collect_content(results) |
| 220 | assert "Checking" in content |
| 221 | |
| 222 | def test_parallel_tools_streaming(self, mock_tokenizer, mock_request): |
| 223 | text = _func_calls( |
| 224 | _invoke("fn_a", _param("x", "true", "1")), |
| 225 | _invoke("fn_b", _param("y", "true", "2")), |
| 226 | ) |
| 227 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 228 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 229 | |
| 230 | names = [] |
| 231 | for delta, _ in results: |
| 232 | if delta and delta.tool_calls: |
| 233 | for tc in delta.tool_calls: |
| 234 | if tc.function and tc.function.name: |
| 235 | names.append(tc.function.name) |
| 236 | assert "fn_a" in names |
| 237 | assert "fn_b" in names |
| 238 | |
| 239 | def test_no_tool_content_only(self, mock_tokenizer, mock_request): |
| 240 | text = "Just some text, no tools." |
| 241 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 242 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 243 | content = collect_content(results) |
| 244 | assert "Just some text" in content |
| 245 | args = collect_tool_arguments(results) |
| 246 | assert args == "" |
| 247 | |
| 248 | def test_streaming_wrapper_unwrap_consistency(self, mock_tokenizer, mock_request): |
| 249 | tool = _make_tool("get_weather", {"location": {"type": "string"}}) |
| 250 | mock_request.tools = [tool] |
| 251 | parser = DeepSeekV32Parser(mock_tokenizer, tools=[tool]) |
| 252 | |
| 253 | chunks = [ |
| 254 | DSML_FUNC_START, |
| 255 | _invoke( |
| 256 | "get_weather", |
| 257 | _param("arguments", "false", '{"location": "NYC"}'), |
| 258 | ), |
| 259 | DSML_FUNC_END, |
| 260 | ] |
| 261 | |
| 262 | results = simulate_tool_streaming(parser, mock_request, chunks) |
| 263 | streamed_args = collect_tool_arguments(results) |
| 264 | |
| 265 | final_delta, _ = results[-1] |
| 266 | finish_delta = parser.finish_streaming() |
| 267 | extracted = parser._build_extracted_result(final_delta, finish_delta) |
| 268 | |
| 269 | assert extracted.tools_called is True |
| 270 | assert len(extracted.tool_calls) == 1 |
| 271 | final_args = extracted.tool_calls[0].function.arguments |
| 272 | assert json.loads(final_args) == {"location": "NYC"} |
| 273 | assert '"arguments"' not in streamed_args |
| 274 | assert final_args.startswith(streamed_args) |
| 275 | |
| 276 | def test_missing_invoke_end(self, mock_tokenizer, mock_request): |
| 277 | text = ( |
| 278 | f"{DSML_FUNC_START}\n" |
| 279 | f"{DSML_INVOKE_PREFIX}fn{DSML_INVOKE_NAME_END}\n" |
| 280 | f"{_param('k', 'true', 'v')}\n" |
| 281 | f"{DSML_FUNC_END}" |
| 282 | ) |
| 283 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 284 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 285 | assert collect_function_name(results) == "fn" |
| 286 | args = json.loads(collect_tool_arguments(results)) |
| 287 | assert args == {"k": "v"} |
| 288 | |
| 289 | |
| 290 | class TestMissingFunctionCallsWrapper: |
| 291 | """An invoke without the ``<｜DSML｜function_calls>`` wrapper is still a |
| 292 | tool call (V3.2 counterpart of #48931).""" |
| 293 | |
| 294 | def test_non_streaming(self, mock_tokenizer, mock_request): |
| 295 | text = _invoke("fn", _param("k", "true", "v")) |
| 296 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 297 | result = parser.extract_tool_calls(text, mock_request) |
| 298 | |
| 299 | assert result.tools_called is True |
| 300 | assert result.tool_calls[0].function.name == "fn" |
| 301 | assert json.loads(result.tool_calls[0].function.arguments) == {"k": "v"} |
| 302 | assert result.content is None |
| 303 | |
| 304 | def test_streaming(self, mock_tokenizer, mock_request): |
| 305 | text = "Sure.\n" + _invoke("fn", _param("k", "true", "v")) |
| 306 | parser = DeepSeekV32Parser(mock_tokenizer) |
| 307 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 308 | |
| 309 | assert collect_function_name(results) == "fn" |
| 310 | assert json.loads(collect_tool_arguments(results)) == {"k": "v"} |
| 311 | assert "DSML" not in collect_content(results) |