utils.py 12.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import asyncio
5
import dataclasses
6
import functools
7
import os
8
from argparse import Namespace
9
10
from logging import Logger
from string import Template
11
from typing import TYPE_CHECKING, Any
12

13
import regex as re
14
from fastapi import Request
15
16
from fastapi.responses import JSONResponse, StreamingResponse
from starlette.background import BackgroundTask, BackgroundTasks
17

18
from vllm import envs
19
from vllm.engine.arg_utils import EngineArgs
20
from vllm.logger import current_formatter_type, init_logger
21
from vllm.platforms import current_platform
22
from vllm.utils.argparse_utils import FlexibleArgumentParser
23

24
25
26
27
if TYPE_CHECKING:
    from vllm.entrypoints.openai.chat_completion.protocol import (
        ChatCompletionRequest,
    )
28
    from vllm.entrypoints.openai.completion.protocol import (
29
        CompletionRequest,
30
31
    )
    from vllm.entrypoints.openai.engine.protocol import (
32
33
        StreamOptions,
    )
34
    from vllm.entrypoints.openai.models.protocol import LoRAModulePath
35
36
37
38
39
40
41
else:
    ChatCompletionRequest = object
    CompletionRequest = object
    StreamOptions = object
    LoRAModulePath = object


42
43
logger = init_logger(__name__)

44
VLLM_SUBCMD_PARSER_EPILOG = (
45
46
47
    "For full list:            vllm {subcmd} --help=all\n"
    "For a section:            vllm {subcmd} --help=ModelConfig    (case-insensitive)\n"  # noqa: E501
    "For a flag:               vllm {subcmd} --help=max-model-len  (_ or - accepted)\n"  # noqa: E501
48
49
    "Documentation:            https://docs.vllm.ai\n"
)
50

51
52
53
54
55
56

async def listen_for_disconnect(request: Request) -> None:
    """Returns if a disconnect message is received"""
    while True:
        message = await request.receive()
        if message["type"] == "http.disconnect":
57
58
59
            # If load tracking is enabled *and* the counter exists, decrement
            # it. Combines the previous nested checks into a single condition
            # to satisfy the linter rule.
60
61
62
            if getattr(
                request.app.state, "enable_server_load_tracking", False
            ) and hasattr(request.app.state, "server_load_metrics"):
63
                request.app.state.server_load_metrics -= 1
64
65
66
67
68
69
            break


def with_cancellation(handler_func):
    """Decorator that allows a route handler to be cancelled by client
    disconnections.
70

71
    This does _not_ use request.is_disconnected, which does not work with
72
    middleware. Instead this follows the pattern from
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
    starlette.StreamingResponse, which simultaneously awaits on two tasks- one
    to wait for an http disconnect message, and the other to do the work that we
    want done. When the first task finishes, the other is cancelled.

    A core assumption of this method is that the body of the request has already
    been read. This is a safe assumption to make for fastapi handlers that have
    already parsed the body of the request into a pydantic model for us.
    This decorator is unsafe to use elsewhere, as it will consume and throw away
    all incoming messages for the request while it looks for a disconnect
    message.

    In the case where a `StreamingResponse` is returned by the handler, this
    wrapper will stop listening for disconnects and instead the response object
    will start listening for disconnects.
    """

    # Functools.wraps is required for this wrapper to appear to fastapi as a
    # normal route handler, with the correct request type hinting.
    @functools.wraps(handler_func)
    async def wrapper(*args, **kwargs):
        # The request is either the second positional arg or `raw_request`
        request = args[1] if len(args) > 1 else kwargs["raw_request"]

        handler_task = asyncio.create_task(handler_func(*args, **kwargs))
        cancellation_task = asyncio.create_task(listen_for_disconnect(request))

99
100
101
        done, pending = await asyncio.wait(
            [handler_task, cancellation_task], return_when=asyncio.FIRST_COMPLETED
        )
102
103
104
105
106
107
108
109
        for task in pending:
            task.cancel()

        if handler_task in done:
            return handler_task.result()
        return None

    return wrapper
110
111
112
113
114
115
116
117


def decrement_server_load(request: Request):
    request.app.state.server_load_metrics -= 1


def load_aware_call(func):
    @functools.wraps(func)
118
    async def wrapper(*args, **kwargs):
119
        raw_request = kwargs.get("raw_request", args[1] if len(args) > 1 else None)
120
121
122

        if raw_request is None:
            raise ValueError(
123
124
                "raw_request required when server load tracking is enabled"
            )
125

126
        if not getattr(raw_request.app.state, "enable_server_load_tracking", False):
127
            return await func(*args, **kwargs)
128

129
130
131
132
        # ensure the counter exists
        if not hasattr(raw_request.app.state, "server_load_metrics"):
            raw_request.app.state.server_load_metrics = 0

133
134
        raw_request.app.state.server_load_metrics += 1
        try:
135
            response = await func(*args, **kwargs)
136
137
138
139
140
141
        except Exception:
            raw_request.app.state.server_load_metrics -= 1
            raise

        if isinstance(response, (JSONResponse, StreamingResponse)):
            if response.background is None:
142
                response.background = BackgroundTask(decrement_server_load, raw_request)
143
            elif isinstance(response.background, BackgroundTasks):
144
                response.background.add_task(decrement_server_load, raw_request)
145
146
147
148
            elif isinstance(response.background, BackgroundTask):
                # Convert the single BackgroundTask to BackgroundTasks
                # and chain the decrement_server_load task to it
                tasks = BackgroundTasks()
149
150
151
152
153
                tasks.add_task(
                    response.background.func,
                    *response.background.args,
                    **response.background.kwargs,
                )
154
155
156
157
158
159
160
161
                tasks.add_task(decrement_server_load, raw_request)
                response.background = tasks
        else:
            raw_request.app.state.server_load_metrics -= 1

        return response

    return wrapper
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182


def cli_env_setup():
    # The safest multiprocessing method is `spawn`, as the default `fork` method
    # is not compatible with some accelerators. The default method will be
    # changing in future versions of Python, so we should use it explicitly when
    # possible.
    #
    # We only set it here in the CLI entrypoint, because changing to `spawn`
    # could break some existing code using vLLM as a library. `spawn` will cause
    # unexpected behavior if the code is not protected by
    # `if __name__ == "__main__":`.
    #
    # References:
    # - https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods
    # - https://pytorch.org/docs/stable/notes/multiprocessing.html#cuda-in-multiprocessing
    # - https://pytorch.org/docs/stable/multiprocessing.html#sharing-cuda-tensors
    # - https://docs.habana.ai/en/latest/PyTorch/Getting_Started_with_PyTorch_and_Gaudi/Getting_Started_with_PyTorch.html?highlight=multiprocessing#torch-multiprocessing-for-dataloaders
    if "VLLM_WORKER_MULTIPROC_METHOD" not in os.environ:
        logger.debug("Setting VLLM_WORKER_MULTIPROC_METHOD to 'spawn'")
        os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
183
184
185
186


def _validate_truncation_size(
    max_model_len: int,
187
188
189
    truncate_prompt_tokens: int | None,
    tokenization_kwargs: dict[str, Any] | None = None,
) -> int | None:
190
191
192
193
194
195
196
197
    if truncate_prompt_tokens is not None:
        if truncate_prompt_tokens <= -1:
            truncate_prompt_tokens = max_model_len

        if truncate_prompt_tokens > max_model_len:
            raise ValueError(
                f"truncate_prompt_tokens value ({truncate_prompt_tokens}) "
                f"is greater than max_model_len ({max_model_len})."
198
199
                f" Please, select a smaller truncation size."
            )
200
201
202
203
204

        if tokenization_kwargs is not None:
            tokenization_kwargs["truncation"] = True
            tokenization_kwargs["max_length"] = truncate_prompt_tokens

205
206
207
208
    else:
        if tokenization_kwargs is not None:
            tokenization_kwargs["truncation"] = False

209
    return truncate_prompt_tokens
210
211


212
213
def get_max_tokens(
    max_model_len: int,
zhuwenwen's avatar
zhuwenwen committed
214
215
    request: "ChatCompletionRequest | CompletionRequest",
    input_length: int,
216
217
    default_sampling_params: dict,
) -> int:
zhuwenwen's avatar
zhuwenwen committed
218
    max_tokens = getattr(request, "max_completion_tokens", None) or request.max_tokens
219
220
221
    default_max_tokens = max_model_len - input_length
    max_output_tokens = current_platform.get_max_output_tokens(input_length)

222
223
224
225
226
227
228
229
230
231
    return min(
        val
        for val in (
            default_max_tokens,
            max_tokens,
            max_output_tokens,
            default_sampling_params.get("max_tokens"),
        )
        if val is not None
    )
232
233


234
def log_non_default_args(args: Namespace | EngineArgs):
235
236
    from vllm.entrypoints.openai.cli_args import make_arg_parser

237
238
    non_default_args = {}

239
240
    # Handle Namespace
    if isinstance(args, Namespace):
241
242
243
244
245
246
247
        parser = make_arg_parser(FlexibleArgumentParser())
        for arg, default in vars(parser.parse_args([])).items():
            if default != getattr(args, arg):
                non_default_args[arg] = getattr(args, arg)

    # Handle EngineArgs instance
    elif isinstance(args, EngineArgs):
248
        default_args = EngineArgs(model=args.model)  # Create default instance
249
250
251
252
253
        for field in dataclasses.fields(args):
            current_val = getattr(args, field.name)
            default_val = getattr(default_args, field.name)
            if current_val != default_val:
                non_default_args[field.name] = current_val
254
255
        if default_args.model != EngineArgs.model:
            non_default_args["model"] = default_args.model
256
    else:
257
258
259
        raise TypeError(
            "Unsupported argument type. Must be Namespace or EngineArgs instance."
        )
260
261

    logger.info("non-default args: %s", non_default_args)
262
263
264


def should_include_usage(
265
    stream_options: "StreamOptions | None", enable_force_include_usage: bool
266
267
268
269
270
271
272
273
274
) -> tuple[bool, bool]:
    if stream_options:
        include_usage = stream_options.include_usage or enable_force_include_usage
        include_continuous_usage = include_usage and bool(
            stream_options.continuous_usage_stats
        )
    else:
        include_usage, include_continuous_usage = enable_force_include_usage, False
    return include_usage, include_continuous_usage
275
276
277
278
279


def process_lora_modules(
    args_lora_modules: list[LoRAModulePath], default_mm_loras: dict[str, str] | None
) -> list[LoRAModulePath]:
280
    from vllm.entrypoints.openai.models.serving import LoRAModulePath
281

282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
    lora_modules = args_lora_modules
    if default_mm_loras:
        default_mm_lora_paths = [
            LoRAModulePath(
                name=modality,
                path=lora_path,
            )
            for modality, lora_path in default_mm_loras.items()
        ]
        if args_lora_modules is None:
            lora_modules = default_mm_lora_paths
        else:
            lora_modules += default_mm_lora_paths
    return lora_modules


298
299
300
def sanitize_message(message: str) -> str:
    # Avoid leaking memory address from object reprs
    return re.sub(r" at 0x[0-9a-f]+>", ">", message)
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325


def log_version_and_model(lgr: Logger, version: str, model_name: str) -> None:
    if envs.VLLM_DISABLE_LOG_LOGO or (formatter := current_formatter_type(lgr)) is None:
        message = "vLLM server version %s, serving model %s"
    else:
        logo_template = Template(
            "\n       ${w}█     █     █▄   ▄█${r}\n"
            " ${o}▄▄${r} ${b}▄█${r} ${w}█     █     █ ▀▄▀ █${r}  version ${w}%s${r}\n"
            "  ${o}█${r}${b}▄█▀${r} ${w}█     █     █     █${r}  model   ${w}%s${r}\n"
            "   ${b}▀▀${r}  ${w}▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀${r}\n"
        )
        colors = {
            "w": "\033[97;1m",  # white
            "o": "\033[93m",  # orange
            "b": "\033[94m",  # blue
            "r": "\033[0m",  # reset
        }
        if formatter != "color":
            # monochrome logo (no ansi escape codes)
            colors = dict.fromkeys(colors, "")

        message = logo_template.substitute(colors)

    lgr.info(message, version, model_name)