| 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 | import pytest_asyncio |
| 9 | from typing_extensions import TypedDict |
| 10 | |
| 11 | from tests.utils import RemoteOpenAIServer |
| 12 | from vllm.tool_parsers.abstract_tool_parser import ToolParser |
| 13 | from vllm.tool_parsers.granite4_tool_parser import Granite4ToolParser |
| 14 | from vllm.tool_parsers.hermes_tool_parser import Hermes2ProToolParser |
| 15 | from vllm.transformers_utils.repo_utils import hf_api |
| 16 | |
| 17 | LORA_MODEL = "minpeter/LoRA-Llama-3.2-1B-tool-vllm-ci" |
| 18 | |
| 19 | TOOLS = [ |
| 20 | { |
| 21 | "type": "function", |
| 22 | "function": { |
| 23 | "name": "get_current_weather", |
| 24 | "description": "Get the current weather in a given location", |
| 25 | "parameters": { |
| 26 | "type": "object", |
| 27 | "properties": { |
| 28 | "location": { |
| 29 | "type": "string", |
| 30 | "description": "The city and state, e.g. San Francisco, CA", |
| 31 | }, |
| 32 | "unit": { |
| 33 | "type": "string", |
| 34 | "enum": ["celsius", "fahrenheit"], |
| 35 | }, |
| 36 | }, |
| 37 | "required": ["location"], |
| 38 | }, |
| 39 | }, |
| 40 | } |
| 41 | ] |
| 42 | |
| 43 | |
| 44 | class ServerConfig(TypedDict, total=False): |
| 45 | model: str |
| 46 | arguments: list[str] |
| 47 | model_arg: str |
| 48 | tool_parser: type[ToolParser] |
| 49 | |
| 50 | |
| 51 | CONFIGS: dict[str, ServerConfig] = { |
| 52 | "llama": { |
| 53 | "model": "meta-llama/Llama-3.2-1B-Instruct", |
| 54 | "arguments": [ |
| 55 | "--enforce-eager", |
| 56 | "--enable-auto-tool-choice", |
| 57 | "--tool-call-parser", |
| 58 | "hermes", |
| 59 | "--enable-lora", |
| 60 | "--lora-modules", |
| 61 | f"{LORA_MODEL}={LORA_MODEL}", |
| 62 | "--tokenizer", |
| 63 | f"{LORA_MODEL}", |
| 64 | ], |
| 65 | "model_arg": LORA_MODEL, |
| 66 | "tool_parser": Hermes2ProToolParser, |
| 67 | }, |
| 68 | "granite4": { |
| 69 | "model": "ibm-granite/granite-4.0-h-tiny", |
| 70 | "arguments": [ |
| 71 | "--enforce-eager", |
| 72 | "--enable-auto-tool-choice", |
| 73 | "--tool-call-parser", |
| 74 | "granite4", |
| 75 | "--tokenizer", |
| 76 | "ibm-granite/granite-4.0-h-tiny", |
| 77 | "--max-model-len", |
| 78 | "4096", |
| 79 | "--max-num-seqs", |
| 80 | "2", |
| 81 | ], |
| 82 | "model_arg": "ibm-granite/granite-4.0-h-tiny", |
| 83 | "tool_parser": Granite4ToolParser, |
| 84 | }, |
| 85 | } |
| 86 | |
| 87 | |
| 88 | # for each server config, download the model and return the config |
| 89 | @pytest.fixture(scope="session", params=CONFIGS.keys()) |
| 90 | def server_config(request): |
| 91 | config = CONFIGS[request.param] |
| 92 | |
| 93 | # download model and tokenizer using transformers |
| 94 | hf_api().snapshot_download(config["model"]) |
| 95 | yield CONFIGS[request.param] |
| 96 | |
| 97 | |
| 98 | @pytest.fixture(scope="module") |
| 99 | def server(request, server_config: ServerConfig): |
| 100 | model = server_config["model"] |
| 101 | args_for_model = server_config["arguments"] |
| 102 | with RemoteOpenAIServer(model, args_for_model, max_wait_seconds=480) as server: |
| 103 | yield server |
| 104 | |
| 105 | |
| 106 | @pytest_asyncio.fixture |
| 107 | async def client(server: RemoteOpenAIServer): |
| 108 | async with server.get_async_client() as async_client: |
| 109 | yield async_client |
| 110 | |
| 111 | |
| 112 | PRODUCT_TOOLS = [ |
| 113 | { |
| 114 | "type": "function", |
| 115 | "function": { |
| 116 | "name": "get_product_info", |
| 117 | "description": "Get detailed information of a product based on its " |
| 118 | "product ID.", |
| 119 | "parameters": { |
| 120 | "type": "object", |
| 121 | "properties": { |
| 122 | "inserted": { |
| 123 | "type": "boolean", |
| 124 | "description": "inserted.", |
| 125 | }, |
| 126 | "product_id": { |
| 127 | "type": "integer", |
| 128 | "description": "The product ID of the product.", |
| 129 | }, |
| 130 | }, |
| 131 | "required": ["product_id", "inserted"], |
| 132 | }, |
| 133 | }, |
| 134 | } |
| 135 | ] |
| 136 | |
| 137 | MESSAGES = [{"role": "user", "content": "What's the weather like in Boston?"}] |
| 138 | |
| 139 | PRODUCT_MESSAGES = [ |
| 140 | { |
| 141 | "role": "user", |
| 142 | "content": "Hi! Do you have any detailed information about the product id " |
| 143 | "7355608 and inserted true?", |
| 144 | } |
| 145 | ] |
| 146 | |
| 147 | |
| 148 | @pytest.mark.asyncio |
| 149 | async def test_non_streaming_tool_call( |
| 150 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 151 | ): |
| 152 | """Test tool call in non-streaming mode.""" |
| 153 | response = await client.chat.completions.create( |
| 154 | model=server_config["model_arg"], |
| 155 | messages=MESSAGES, |
| 156 | tools=TOOLS, |
| 157 | tool_choice="auto", |
| 158 | temperature=0.0, |
| 159 | ) |
| 160 | |
| 161 | assert response.choices |
| 162 | choice = response.choices[0] |
| 163 | message = choice.message |
| 164 | |
| 165 | assert choice.finish_reason == "tool_calls" |
| 166 | assert message.tool_calls is not None |
| 167 | |
| 168 | tool_call = message.tool_calls[0] |
| 169 | assert tool_call.type == "function" |
| 170 | assert tool_call.function.name == "get_current_weather" |
| 171 | |
| 172 | arguments = json.loads(tool_call.function.arguments) |
| 173 | assert "location" in arguments |
| 174 | assert "Boston" in arguments["location"] |
| 175 | print("\n[Non-Streaming Test Passed]") |
| 176 | print(f"Tool Call: {tool_call.function.name}") |
| 177 | print(f"Arguments: {arguments}") |
| 178 | |
| 179 | |
| 180 | @pytest.mark.asyncio |
| 181 | async def test_streaming_tool_call( |
| 182 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 183 | ): |
| 184 | """Test tool call in streaming mode.""" |
| 185 | stream = await client.chat.completions.create( |
| 186 | model=server_config["model_arg"], |
| 187 | messages=MESSAGES, |
| 188 | tools=TOOLS, |
| 189 | tool_choice="auto", |
| 190 | temperature=0.0, |
| 191 | stream=True, |
| 192 | ) |
| 193 | |
| 194 | tool_call_chunks = {} |
| 195 | async for chunk in stream: |
| 196 | if not chunk.choices: |
| 197 | continue |
| 198 | |
| 199 | delta = chunk.choices[0].delta |
| 200 | if not delta or not delta.tool_calls: |
| 201 | continue |
| 202 | |
| 203 | for tool_chunk in delta.tool_calls: |
| 204 | index = tool_chunk.index |
| 205 | if index not in tool_call_chunks: |
| 206 | tool_call_chunks[index] = {"name": "", "arguments": ""} |
| 207 | |
| 208 | if tool_chunk.function.name: |
| 209 | tool_call_chunks[index]["name"] += tool_chunk.function.name |
| 210 | if tool_chunk.function.arguments: |
| 211 | tool_call_chunks[index]["arguments"] += tool_chunk.function.arguments |
| 212 | |
| 213 | assert len(tool_call_chunks) == 1 |
| 214 | reconstructed_tool_call = tool_call_chunks[0] |
| 215 | |
| 216 | assert reconstructed_tool_call["name"] == "get_current_weather" |
| 217 | |
| 218 | arguments = json.loads(reconstructed_tool_call["arguments"]) |
| 219 | assert "location" in arguments |
| 220 | assert "Boston" in arguments["location"] |
| 221 | print("\n[Streaming Test Passed]") |
| 222 | print(f"Reconstructed Tool Call: {reconstructed_tool_call['name']}") |
| 223 | print(f"Reconstructed Arguments: {arguments}") |
| 224 | |
| 225 | |
| 226 | @pytest.mark.asyncio |
| 227 | async def test_non_streaming_product_tool_call( |
| 228 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 229 | ): |
| 230 | """Test tool call integer and boolean parameters in non-streaming mode.""" |
| 231 | response = await client.chat.completions.create( |
| 232 | model=server_config["model_arg"], |
| 233 | messages=PRODUCT_MESSAGES, |
| 234 | tools=PRODUCT_TOOLS, |
| 235 | tool_choice="auto", |
| 236 | temperature=0.66, |
| 237 | ) |
| 238 | |
| 239 | assert response.choices |
| 240 | choice = response.choices[0] |
| 241 | message = choice.message |
| 242 | |
| 243 | assert choice.finish_reason == "tool_calls" |
| 244 | assert message.tool_calls is not None |
| 245 | |
| 246 | tool_call = message.tool_calls[0] |
| 247 | assert tool_call.type == "function" |
| 248 | assert tool_call.function.name == "get_product_info" |
| 249 | |
| 250 | arguments = json.loads(tool_call.function.arguments) |
| 251 | assert "product_id" in arguments |
| 252 | assert "inserted" in arguments |
| 253 | |
| 254 | product_id = arguments.get("product_id") |
| 255 | inserted = arguments.get("inserted") |
| 256 | |
| 257 | assert isinstance(product_id, int) |
| 258 | assert product_id == 7355608 |
| 259 | assert isinstance(inserted, bool) |
| 260 | assert inserted is True |
| 261 | |
| 262 | print("\n[Non-Streaming Product Test Passed]") |
| 263 | print(f"Tool Call: {tool_call.function.name}") |
| 264 | print(f"Arguments: {arguments}") |
| 265 | |
| 266 | |
| 267 | @pytest.mark.asyncio |
| 268 | async def test_streaming_product_tool_call( |
| 269 | client: openai.AsyncOpenAI, server_config: ServerConfig |
| 270 | ): |
| 271 | """Test tool call integer and boolean parameters in streaming mode.""" |
| 272 | stream = await client.chat.completions.create( |
| 273 | model=server_config["model_arg"], |
| 274 | messages=PRODUCT_MESSAGES, |
| 275 | tools=PRODUCT_TOOLS, |
| 276 | tool_choice="auto", |
| 277 | temperature=0.66, |
| 278 | stream=True, |
| 279 | ) |
| 280 | |
| 281 | tool_call_chunks = {} |
| 282 | async for chunk in stream: |
| 283 | if not chunk.choices: |
| 284 | continue |
| 285 | |
| 286 | delta = chunk.choices[0].delta |
| 287 | if not delta or not delta.tool_calls: |
| 288 | continue |
| 289 | |
| 290 | for tool_chunk in delta.tool_calls: |
| 291 | index = tool_chunk.index |
| 292 | if index not in tool_call_chunks: |
| 293 | tool_call_chunks[index] = {"name": "", "arguments": ""} |
| 294 | |
| 295 | if tool_chunk.function.name: |
| 296 | tool_call_chunks[index]["name"] += tool_chunk.function.name |
| 297 | if tool_chunk.function.arguments: |
| 298 | tool_call_chunks[index]["arguments"] += tool_chunk.function.arguments |
| 299 | |
| 300 | assert len(tool_call_chunks) == 1 |
| 301 | reconstructed_tool_call = tool_call_chunks[0] |
| 302 | |
| 303 | assert reconstructed_tool_call["name"] == "get_product_info" |
| 304 | |
| 305 | arguments = json.loads(reconstructed_tool_call["arguments"]) |
| 306 | assert "product_id" in arguments |
| 307 | assert "inserted" in arguments |
| 308 | |
| 309 | # Handle type coercion for streaming test as well |
| 310 | product_id = arguments.get("product_id") |
| 311 | inserted = arguments.get("inserted") |
| 312 | |
| 313 | assert isinstance(product_id, int) |
| 314 | assert product_id == 7355608 |
| 315 | assert isinstance(inserted, bool) |
| 316 | assert inserted is True |
| 317 | |
| 318 | print("\n[Streaming Product Test Passed]") |
| 319 | print(f"Reconstructed Tool Call: {reconstructed_tool_call['name']}") |
| 320 | print(f"Reconstructed Arguments: {arguments}") |