| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | |
| 6 | import openai |
| 7 | import pytest |
| 8 | |
| 9 | from .utils import ( |
| 10 | MESSAGES_ASKING_FOR_PARALLEL_TOOLS, |
| 11 | MESSAGES_WITH_PARALLEL_TOOL_RESPONSE, |
| 12 | SEARCH_TOOL, |
| 13 | SEED, |
| 14 | WEATHER_TOOL, |
| 15 | ServerConfig, |
| 16 | ensure_system_prompt, |
| 17 | ) |
| 18 | |
| 19 | |
| 20 | def apply_parallel_tool_system_prompt( |
| 21 | messages, |
| 22 | server_config: ServerConfig, |
| 23 | ): |
| 24 | if server_config["model"] == "ibm-granite/granite-3.0-8b-instruct": |
| 25 | return ensure_system_prompt(messages, server_config) |
| 26 | return messages |
| 27 | |
| 28 | |
| 29 | # test: getting the model to generate parallel tool calls (streaming/not) |
| 30 | # when requested. NOTE that not all models may support this, so some exclusions |
| 31 | # may be added in the future. e.g. llama 3.1 models are not designed to support |
| 32 | # parallel tool calls. |
| 33 | @pytest.mark.asyncio |
| 34 | async def test_parallel_tool_calls( |
| 35 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 36 | ): |
| 37 | if not server_config.get("supports_parallel", True): |
| 38 | pytest.skip( |
| 39 | "The {} model doesn't support parallel tool calls".format( |
| 40 | server_config["model"] |
| 41 | ) |
| 42 | ) |
| 43 | |
| 44 | models = await client.models.list() |
| 45 | model_name: str = models.data[0].id |
| 46 | messages = apply_parallel_tool_system_prompt( |
| 47 | MESSAGES_ASKING_FOR_PARALLEL_TOOLS, server_config |
| 48 | ) |
| 49 | chat_completion = await client.chat.completions.create( |
| 50 | messages=messages, |
| 51 | temperature=0, |
| 52 | max_completion_tokens=200, |
| 53 | model=model_name, |
| 54 | tools=[WEATHER_TOOL, SEARCH_TOOL], |
| 55 | logprobs=False, |
| 56 | seed=SEED, |
| 57 | ) |
| 58 | |
| 59 | choice = chat_completion.choices[0] |
| 60 | stop_reason = chat_completion.choices[0].finish_reason |
| 61 | non_streamed_tool_calls = chat_completion.choices[0].message.tool_calls |
| 62 | |
| 63 | # make sure 2 tool calls are present |
| 64 | assert choice.message.role == "assistant" |
| 65 | assert non_streamed_tool_calls is not None |
| 66 | assert len(non_streamed_tool_calls) == 2 |
| 67 | |
| 68 | for tool_call in non_streamed_tool_calls: |
| 69 | # make sure the tool includes a function and ID |
| 70 | assert tool_call.type == "function" |
| 71 | assert tool_call.function is not None |
| 72 | assert isinstance(tool_call.id, str) |
| 73 | assert len(tool_call.id) >= 9 |
| 74 | |
| 75 | # make sure the weather tool was called correctly |
| 76 | assert tool_call.function.name == WEATHER_TOOL["function"]["name"] |
| 77 | assert isinstance(tool_call.function.arguments, str) |
| 78 | |
| 79 | parsed_arguments = json.loads(tool_call.function.arguments) |
| 80 | assert isinstance(parsed_arguments, dict) |
| 81 | assert isinstance(parsed_arguments.get("city"), str) |
| 82 | assert isinstance(parsed_arguments.get("state"), str) |
| 83 | |
| 84 | assert stop_reason == "tool_calls" |
| 85 | |
| 86 | # make the same request, streaming |
| 87 | stream = await client.chat.completions.create( |
| 88 | model=model_name, |
| 89 | messages=messages, |
| 90 | temperature=0, |
| 91 | max_completion_tokens=200, |
| 92 | tools=[WEATHER_TOOL, SEARCH_TOOL], |
| 93 | logprobs=False, |
| 94 | seed=SEED, |
| 95 | stream=True, |
| 96 | ) |
| 97 | |
| 98 | role_name: str | None = None |
| 99 | finish_reason_count: int = 0 |
| 100 | |
| 101 | tool_call_names: list[str] = [] |
| 102 | tool_call_args: list[str] = [] |
| 103 | tool_call_idx: int = -1 |
| 104 | tool_call_id_count: int = 0 |
| 105 | |
| 106 | async for chunk in stream: |
| 107 | # if there's a finish reason make sure it's tools |
| 108 | if chunk.choices[0].finish_reason: |
| 109 | finish_reason_count += 1 |
| 110 | assert chunk.choices[0].finish_reason == "tool_calls" |
| 111 | |
| 112 | # if a role is being streamed make sure it wasn't already set to |
| 113 | # something else |
| 114 | if chunk.choices[0].delta.role: |
| 115 | assert not role_name or role_name == "assistant" |
| 116 | role_name = "assistant" |
| 117 | |
| 118 | # a chunk may carry >1 tool-call delta at a parallel-call boundary |
| 119 | streamed_tool_calls = chunk.choices[0].delta.tool_calls |
| 120 | |
| 121 | for tool_call in streamed_tool_calls or []: |
| 122 | # deltas arrive in non-decreasing index order |
| 123 | assert tool_call.index >= tool_call_idx |
| 124 | |
| 125 | # if a new tool is being called, set up empty arguments |
| 126 | if tool_call.index != tool_call_idx: |
| 127 | tool_call_idx = tool_call.index |
| 128 | tool_call_args.append("") |
| 129 | |
| 130 | # if a tool call ID is streamed, make sure one hasn't been already |
| 131 | if tool_call.id: |
| 132 | tool_call_id_count += 1 |
| 133 | assert isinstance(tool_call.id, str) and (len(tool_call.id) >= 9) |
| 134 | |
| 135 | # if parts of the function start being streamed |
| 136 | if tool_call.function: |
| 137 | # if the function name is defined, set it. it should be streamed |
| 138 | # IN ENTIRETY, exactly one time. |
| 139 | if tool_call.function.name: |
| 140 | assert isinstance(tool_call.function.name, str) |
| 141 | tool_call_names.append(tool_call.function.name) |
| 142 | |
| 143 | if tool_call.function.arguments: |
| 144 | # make sure they're a string and then add them to the list |
| 145 | assert isinstance(tool_call.function.arguments, str) |
| 146 | |
| 147 | tool_call_args[tool_call.index] += tool_call.function.arguments |
| 148 | |
| 149 | assert finish_reason_count == 1 |
| 150 | assert role_name == "assistant" |
| 151 | |
| 152 | assert len(non_streamed_tool_calls) == len(tool_call_names) == len(tool_call_args) |
| 153 | |
| 154 | for i in range(2): |
| 155 | assert non_streamed_tool_calls[i].function.name == tool_call_names[i] |
| 156 | streamed_args = json.loads(tool_call_args[i]) |
| 157 | non_streamed_args = json.loads(non_streamed_tool_calls[i].function.arguments) |
| 158 | assert streamed_args == non_streamed_args |
| 159 | |
| 160 | |
| 161 | # test: providing parallel tool calls back to the model to get a response |
| 162 | # (streaming/not) |
| 163 | @pytest.mark.asyncio |
| 164 | async def test_parallel_tool_calls_with_results( |
| 165 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 166 | ): |
| 167 | if not server_config.get("supports_parallel", True): |
| 168 | pytest.skip( |
| 169 | "The {} model doesn't support parallel tool calls".format( |
| 170 | server_config["model"] |
| 171 | ) |
| 172 | ) |
| 173 | |
| 174 | models = await client.models.list() |
| 175 | model_name: str = models.data[0].id |
| 176 | messages = apply_parallel_tool_system_prompt( |
| 177 | MESSAGES_WITH_PARALLEL_TOOL_RESPONSE, server_config |
| 178 | ) |
| 179 | chat_completion = await client.chat.completions.create( |
| 180 | messages=messages, |
| 181 | temperature=0, |
| 182 | max_completion_tokens=200, |
| 183 | model=model_name, |
| 184 | tools=[WEATHER_TOOL, SEARCH_TOOL], |
| 185 | logprobs=False, |
| 186 | seed=SEED, |
| 187 | ) |
| 188 | |
| 189 | choice = chat_completion.choices[0] |
| 190 | |
| 191 | assert choice.finish_reason != "tool_calls" # "stop" or "length" |
| 192 | assert choice.message.role == "assistant" |
| 193 | assert choice.message.tool_calls is None or len(choice.message.tool_calls) == 0 |
| 194 | assert choice.message.content is not None |
| 195 | assert "98" in choice.message.content # Dallas temp in tool response |
| 196 | assert "78" in choice.message.content # Orlando temp in tool response |
| 197 | |
| 198 | stream = await client.chat.completions.create( |
| 199 | messages=messages, |
| 200 | temperature=0, |
| 201 | max_completion_tokens=200, |
| 202 | model=model_name, |
| 203 | tools=[WEATHER_TOOL, SEARCH_TOOL], |
| 204 | logprobs=False, |
| 205 | seed=SEED, |
| 206 | stream=True, |
| 207 | ) |
| 208 | |
| 209 | chunks: list[str] = [] |
| 210 | finish_reason_count = 0 |
| 211 | role_sent: bool = False |
| 212 | |
| 213 | async for chunk in stream: |
| 214 | delta = chunk.choices[0].delta |
| 215 | |
| 216 | if delta.role: |
| 217 | assert not role_sent |
| 218 | assert delta.role == "assistant" |
| 219 | role_sent = True |
| 220 | |
| 221 | if delta.content: |
| 222 | chunks.append(delta.content) |
| 223 | |
| 224 | if chunk.choices[0].finish_reason is not None: |
| 225 | finish_reason_count += 1 |
| 226 | assert chunk.choices[0].finish_reason == choice.finish_reason |
| 227 | |
| 228 | assert not delta.tool_calls or len(delta.tool_calls) == 0 |
| 229 | |
| 230 | assert role_sent |
| 231 | assert finish_reason_count == 1 |
| 232 | assert len(chunks) |
| 233 | assert "".join(chunks) == choice.message.content |
| 234 | |
| 235 | |
| 236 | @pytest.mark.asyncio |
| 237 | async def test_parallel_tool_calls_false( |
| 238 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 239 | ): |
| 240 | """Ensure only one tool call is returned when parallel_tool_calls is False.""" |
| 241 | models = await client.models.list() |
| 242 | model_name: str = models.data[0].id |
| 243 | messages = apply_parallel_tool_system_prompt( |
| 244 | MESSAGES_ASKING_FOR_PARALLEL_TOOLS, server_config |
| 245 | ) |
| 246 | chat_completion = await client.chat.completions.create( |
| 247 | messages=messages, |
| 248 | temperature=0, |
| 249 | max_completion_tokens=200, |
| 250 | model=model_name, |
| 251 | tools=[WEATHER_TOOL, SEARCH_TOOL], |
| 252 | logprobs=False, |
| 253 | seed=SEED, |
| 254 | parallel_tool_calls=False, |
| 255 | ) |
| 256 | |
| 257 | stop_reason = chat_completion.choices[0].finish_reason |
| 258 | non_streamed_tool_calls = chat_completion.choices[0].message.tool_calls |
| 259 | |
| 260 | # make sure only 1 tool call is present |
| 261 | assert len(non_streamed_tool_calls) == 1 |
| 262 | assert stop_reason == "tool_calls" |
| 263 | |
| 264 | # make the same request, streaming |
| 265 | stream = await client.chat.completions.create( |
| 266 | model=model_name, |
| 267 | messages=messages, |
| 268 | temperature=0, |
| 269 | max_completion_tokens=200, |
| 270 | tools=[WEATHER_TOOL, SEARCH_TOOL], |
| 271 | logprobs=False, |
| 272 | seed=SEED, |
| 273 | parallel_tool_calls=False, |
| 274 | stream=True, |
| 275 | ) |
| 276 | |
| 277 | finish_reason_count: int = 0 |
| 278 | tool_call_id_count: int = 0 |
| 279 | |
| 280 | async for chunk in stream: |
| 281 | # if there's a finish reason make sure it's tools |
| 282 | if chunk.choices[0].finish_reason: |
| 283 | finish_reason_count += 1 |
| 284 | assert chunk.choices[0].finish_reason == "tool_calls" |
| 285 | |
| 286 | streamed_tool_calls = chunk.choices[0].delta.tool_calls |
| 287 | if streamed_tool_calls and len(streamed_tool_calls) > 0: |
| 288 | tool_call = streamed_tool_calls[0] |
| 289 | if tool_call.id: |
| 290 | tool_call_id_count += 1 |
| 291 | |
| 292 | # make sure only 1 streaming tool call is present |
| 293 | assert tool_call_id_count == 1 |
| 294 | assert finish_reason_count == 1 |