utils.py 53.5 KB
Newer Older
1
import argparse
2
import asyncio
3
import contextlib
4
import datetime
Woosuk Kwon's avatar
Woosuk Kwon committed
5
import enum
6
import gc
7
import getpass
8
import importlib.util
9
import inspect
10
import ipaddress
11
import os
12
import socket
13
import subprocess
14
import sys
15
16
import tempfile
import threading
17
import time
Zhuohan Li's avatar
Zhuohan Li committed
18
import uuid
19
import warnings
20
import weakref
21
from asyncio import FIRST_COMPLETED, AbstractEventLoop, Future, Task
22
23
from collections import defaultdict
from collections.abc import Iterable, Mapping
24
from functools import lru_cache, partial, wraps
25
from platform import uname
26
from typing import (Any, AsyncGenerator, Awaitable, Callable, Dict, Generic,
27
28
                    Hashable, List, Literal, Optional, OrderedDict, Set, Tuple,
                    Type, TypeVar, Union, overload)
29
from uuid import uuid4
Zhuohan Li's avatar
Zhuohan Li committed
30

31
import numpy as np
32
import numpy.typing as npt
33
import psutil
Zhuohan Li's avatar
Zhuohan Li committed
34
import torch
35
import torch.types
36
import yaml
37
from packaging.version import Version
38
from torch.library import Library
39
from typing_extensions import ParamSpec, TypeIs, assert_never
40

41
import vllm.envs as envs
42
from vllm.logger import enable_trace_function_call, init_logger
43
from vllm.platforms import current_platform
44
45
46

logger = init_logger(__name__)

47
48
# Exception strings for non-implemented encoder/decoder scenarios

49
50
51
# Reminder: Please update docs/source/serving/compatibility_matrix.rst
# If the feature combo become valid

52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
STR_NOT_IMPL_ENC_DEC_SWA = \
    "Sliding window attention for encoder/decoder models " + \
                    "is not currently supported."

STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE = \
    "Prefix caching for encoder/decoder models " + \
                    "is not currently supported."

STR_NOT_IMPL_ENC_DEC_CHUNKED_PREFILL = \
    "Chunked prefill for encoder/decoder models " + \
                    "is not currently supported."

STR_NOT_IMPL_ENC_DEC_LOGIT_SOFTCAP = (
    "Models with logits_soft_cap "
    "require FlashInfer backend, which is "
    "currently not supported for encoder/decoder "
    "models.")

STR_NOT_IMPL_ENC_DEC_LORA = ("LoRA is currently not currently "
                             "supported with encoder/decoder "
                             "models.")

STR_NOT_IMPL_ENC_DEC_PP = ("Pipeline parallelism is not "
                           "currently supported with "
                           "encoder/decoder models.")

STR_NOT_IMPL_ENC_DEC_MM = ("Multimodal is not currently "
                           "supported with encoder/decoder "
                           "models.")

STR_NOT_IMPL_ENC_DEC_SPEC_DEC = ("Speculative decoding is not "
                                 "currently supported with encoder/"
                                 "decoder models.")

86
87
STR_NOT_IMPL_ENC_DEC_BACKEND = ("XFormers and Flash-Attention are the only "
                                "backends currently supported with encoder/"
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
                                "decoder models.")

STR_NOT_IMPL_ENC_DEC_PROMPT_ADAPTER = ("Prompt adapters are not "
                                       "currently supported with encoder/"
                                       "decoder models.")

# Efficiently import all enc/dec error strings
# rather than having to import all of the above
STR_NOT_IMPL_ENC_DEC_ERR_STRS = {
    "STR_NOT_IMPL_ENC_DEC_SWA": STR_NOT_IMPL_ENC_DEC_SWA,
    "STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE": STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE,
    "STR_NOT_IMPL_ENC_DEC_CHUNKED_PREFILL":
    STR_NOT_IMPL_ENC_DEC_CHUNKED_PREFILL,
    "STR_NOT_IMPL_ENC_DEC_LOGIT_SOFTCAP": STR_NOT_IMPL_ENC_DEC_LOGIT_SOFTCAP,
    "STR_NOT_IMPL_ENC_DEC_LORA": STR_NOT_IMPL_ENC_DEC_LORA,
    "STR_NOT_IMPL_ENC_DEC_PP": STR_NOT_IMPL_ENC_DEC_PP,
    "STR_NOT_IMPL_ENC_DEC_MM": STR_NOT_IMPL_ENC_DEC_MM,
    "STR_NOT_IMPL_ENC_DEC_SPEC_DEC": STR_NOT_IMPL_ENC_DEC_SPEC_DEC,
    "STR_NOT_IMPL_ENC_DEC_BACKEND": STR_NOT_IMPL_ENC_DEC_BACKEND,
    "STR_NOT_IMPL_ENC_DEC_PROMPT_ADAPTER": STR_NOT_IMPL_ENC_DEC_PROMPT_ADAPTER,
}

# Constants related to forcing the attention backend selection

# String name of register which may be set in order to
# force auto-selection of attention backend by Attention
# wrapper
STR_BACKEND_ENV_VAR: str = "VLLM_ATTENTION_BACKEND"

# Possible string values of STR_BACKEND_ENV_VAR
# register, corresponding to possible backends
STR_FLASHINFER_ATTN_VAL: str = "FLASHINFER"
STR_TORCH_SDPA_ATTN_VAL: str = "TORCH_SDPA"
STR_ROCM_FLASH_ATTN_VAL: str = "ROCM_FLASH"
STR_XFORMERS_ATTN_VAL: str = "XFORMERS"
STR_FLASH_ATTN_VAL: str = "FLASH_ATTN"
STR_INVALID_VAL: str = "INVALID"

126
127
128
GB_bytes = 1_000_000_000
"""The number of bytes in one gigabyte (GB)."""

129
130
131
GiB_bytes = 1 << 30
"""The number of bytes in one gibibyte (GiB)."""

132
133
134
135
STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.half,
    "bfloat16": torch.bfloat16,
    "float": torch.float,
136
    "fp8": torch.uint8,
137
138
    "fp8_e4m3": torch.uint8,
    "fp8_e5m2": torch.uint8,
139
}
Zhuohan Li's avatar
Zhuohan Li committed
140

141
142
143
144
145
146
147
148
149
TORCH_DTYPE_TO_NUMPY_DTYPE = {
    torch.float16: np.float16,
    torch.float32: np.float32,
    torch.float64: np.float64,
    torch.uint8: np.uint8,
    torch.int32: np.int32,
    torch.int64: np.int64,
}

150
151
152
P = ParamSpec('P')
K = TypeVar("K")
T = TypeVar("T")
153
U = TypeVar("U")
154

Woosuk Kwon's avatar
Woosuk Kwon committed
155

156
157
158
159
160
161
162
class _Sentinel:
    ...


ALL_PINNED_SENTINEL = _Sentinel()


Woosuk Kwon's avatar
Woosuk Kwon committed
163
164
165
166
167
168
169
170
171
172
class Device(enum.Enum):
    GPU = enum.auto()
    CPU = enum.auto()


class Counter:

    def __init__(self, start: int = 0) -> None:
        self.counter = start

Woosuk Kwon's avatar
Woosuk Kwon committed
173
    def __next__(self) -> int:
174
        i = self.counter
Woosuk Kwon's avatar
Woosuk Kwon committed
175
        self.counter += 1
176
        return i
Woosuk Kwon's avatar
Woosuk Kwon committed
177
178
179

    def reset(self) -> None:
        self.counter = 0
Zhuohan Li's avatar
Zhuohan Li committed
180

181

182
class LRUCache(Generic[T]):
183
184

    def __init__(self, capacity: int):
185
        self.cache: OrderedDict[Hashable, T] = OrderedDict()
186
        self.pinned_items: Set[Hashable] = set()
187
188
189
190
191
192
193
194
        self.capacity = capacity

    def __contains__(self, key: Hashable) -> bool:
        return key in self.cache

    def __len__(self) -> int:
        return len(self.cache)

195
196
197
198
    def __getitem__(self, key: Hashable) -> T:
        value = self.cache[key]  # Raise KeyError if not exists
        self.cache.move_to_end(key)
        return value
199

200
    def __setitem__(self, key: Hashable, value: T) -> None:
201
202
203
204
205
206
207
208
        self.put(key, value)

    def __delitem__(self, key: Hashable) -> None:
        self.pop(key)

    def touch(self, key: Hashable) -> None:
        self.cache.move_to_end(key)

209
210
211
    def get(self,
            key: Hashable,
            default_value: Optional[T] = None) -> Optional[T]:
212
        value: Optional[T]
213
        if key in self.cache:
214
            value = self.cache[key]
215
216
217
218
219
            self.cache.move_to_end(key)
        else:
            value = default_value
        return value

220
    def put(self, key: Hashable, value: T) -> None:
221
222
223
224
        self.cache[key] = value
        self.cache.move_to_end(key)
        self._remove_old_if_needed()

225
226
227
228
229
230
231
232
233
234
235
236
    def pin(self, key: Hashable) -> None:
        """
        Pins a key in the cache preventing it from being
        evicted in the LRU order.
        """
        if key not in self.cache:
            raise ValueError(f"Cannot pin key: {key} not in cache.")
        self.pinned_items.add(key)

    def _unpin(self, key: Hashable) -> None:
        self.pinned_items.remove(key)

237
    def _on_remove(self, key: Hashable, value: Optional[T]):
238
239
        pass

240
    def remove_oldest(self, remove_pinned=False):
241
242
        if not self.cache:
            return
243
244
245
246
247
248
249
250
251
252
253
254

        if not remove_pinned:
            # pop the oldest item in the cache that is not pinned
            lru_key = next(
                (key for key in self.cache if key not in self.pinned_items),
                ALL_PINNED_SENTINEL)
            if lru_key is ALL_PINNED_SENTINEL:
                raise RuntimeError("All items are pinned, "
                                   "cannot remove oldest from the cache.")
        else:
            lru_key = next(iter(self.cache))
        self.pop(lru_key)
255
256
257
258
259

    def _remove_old_if_needed(self) -> None:
        while len(self.cache) > self.capacity:
            self.remove_oldest()

260
261
262
    def pop(self,
            key: Hashable,
            default_value: Optional[T] = None) -> Optional[T]:
263
        run_on_remove = key in self.cache
264
        value: Optional[T] = self.cache.pop(key, default_value)
265
266
267
        # remove from pinned items
        if key in self.pinned_items:
            self._unpin(key)
268
269
270
271
272
273
        if run_on_remove:
            self._on_remove(key, value)
        return value

    def clear(self):
        while len(self.cache) > 0:
274
            self.remove_oldest(remove_pinned=True)
275
276
277
        self.cache.clear()


278
class PyObjectCache:
279
    """Used to cache python objects to avoid object allocations
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
    across scheduler iterations.
    """

    def __init__(self, obj_builder):
        self._obj_builder = obj_builder
        self._index = 0

        self._obj_cache = []
        for _ in range(128):
            self._obj_cache.append(self._obj_builder())

    def _grow_cache(self):
        # Double the size of the cache
        num_objs = len(self._obj_cache)
        for _ in range(num_objs):
            self._obj_cache.append(self._obj_builder())

    def get_object(self):
298
        """Returns a pre-allocated cached object. If there is not enough
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
        objects, then the cache size will double.
        """
        if self._index >= len(self._obj_cache):
            self._grow_cache()
            assert self._index < len(self._obj_cache)

        obj = self._obj_cache[self._index]
        self._index += 1

        return obj

    def reset(self):
        """Makes all cached-objects available for the next scheduler iteration.
        """
        self._index = 0


316
@lru_cache(maxsize=None)
317
318
def get_max_shared_memory_bytes(gpu: int = 0) -> int:
    """Returns the maximum shared memory per thread block in bytes."""
319
    from vllm import _custom_ops as ops
320
    max_shared_mem = (
321
        ops.get_max_shared_memory_per_block_device_attribute(gpu))
322
323
    # value 0 will cause MAX_SEQ_LEN become negative and test_attention.py
    # will fail
324
    assert max_shared_mem > 0, "max_shared_mem can not be zero"
325
326
327
    return int(max_shared_mem)


328
def get_cpu_memory() -> int:
329
    """Returns the total CPU memory of the node in bytes."""
330
    return psutil.virtual_memory().total
Zhuohan Li's avatar
Zhuohan Li committed
331
332
333
334


def random_uuid() -> str:
    return str(uuid.uuid4().hex)
335

336

337
@lru_cache(maxsize=None)
338
def get_vllm_instance_id() -> str:
339
340
341
342
343
344
    """
    If the environment variable VLLM_INSTANCE_ID is set, return it.
    Otherwise, return a random UUID.
    Instance id represents an instance of the VLLM. All processes in the same
    instance should have the same instance id.
    """
345
    return envs.VLLM_INSTANCE_ID or f"vllm-instance-{random_uuid()}"
346
347


348
@lru_cache(maxsize=None)
349
350
351
def in_wsl() -> bool:
    # Reference: https://github.com/microsoft/WSL/issues/4071
    return "microsoft" in " ".join(uname()).lower()
352
353


354
def make_async(func: Callable[P, T]) -> Callable[P, Awaitable[T]]:
355
356
357
358
359
360
361
    """Take a blocking function, and run it on in an executor thread.

    This function prevents the blocking function from blocking the
    asyncio event loop.
    The code in this function needs to be thread safe.
    """

362
    def _async_wrapper(*args: P.args, **kwargs: P.kwargs) -> asyncio.Future:
363
364
365
366
367
368
369
        loop = asyncio.get_event_loop()
        p_func = partial(func, *args, **kwargs)
        return loop.run_in_executor(executor=None, func=p_func)

    return _async_wrapper


370
371
372
373
374
375
def _next_task(iterator: AsyncGenerator[T, None],
               loop: AbstractEventLoop) -> Task:
    # Can use anext() in python >= 3.10
    return loop.create_task(iterator.__anext__())  # type: ignore[arg-type]


376
377
378
379
380
381
382
383
async def iterate_with_cancellation(
    iterator: AsyncGenerator[T, None],
    is_cancelled: Callable[[], Awaitable[bool]],
) -> AsyncGenerator[T, None]:
    """Convert async iterator into one that polls the provided function
    at least once per second to check for client cancellation.
    """

384
385
386
387
    loop = asyncio.get_running_loop()

    awaits: List[Future[T]] = [_next_task(iterator, loop)]
    next_cancel_check: float = 0
388
    while True:
389
390
391
392
393
394
395
396
397
398
399
400
        done, pending = await asyncio.wait(awaits, timeout=1.5)

        # Check for cancellation at most once per second
        time_now = time.time()
        if time_now >= next_cancel_check:
            if await is_cancelled():
                with contextlib.suppress(BaseException):
                    awaits[0].cancel()
                    await iterator.aclose()
                raise asyncio.CancelledError("client cancelled")
            next_cancel_check = time_now + 1

401
402
403
        if done:
            try:
                item = await awaits[0]
404
                awaits[0] = _next_task(iterator, loop)
405
406
407
408
                yield item
            except StopAsyncIteration:
                # we are done
                return
409
410


411
412
async def merge_async_iterators(
    *iterators: AsyncGenerator[T, None],
413
    is_cancelled: Optional[Callable[[], Awaitable[bool]]] = None,
414
) -> AsyncGenerator[Tuple[int, T], None]:
415
416
417
418
419
    """Merge multiple asynchronous iterators into a single iterator.

    This method handle the case where some iterators finish before others.
    When it yields, it yields a tuple (i, item) where i is the index of the
    iterator that yields the item.
420

421
422
    It also optionally polls a provided function at least once per second
    to check for client cancellation.
423
    """
424

425
426
427
428
429
    loop = asyncio.get_running_loop()

    awaits = {_next_task(pair[1], loop): pair for pair in enumerate(iterators)}
    timeout = None if is_cancelled is None else 1.5
    next_cancel_check: float = 0
430
431
432
433
    try:
        while awaits:
            done, pending = await asyncio.wait(awaits.keys(),
                                               return_when=FIRST_COMPLETED,
434
                                               timeout=timeout)
435
436
437
438
439
440
441
            if is_cancelled is not None:
                # Check for cancellation at most once per second
                time_now = time.time()
                if time_now >= next_cancel_check:
                    if await is_cancelled():
                        raise asyncio.CancelledError("client cancelled")
                    next_cancel_check = time_now + 1
442
443
444
445
446
            for d in done:
                pair = awaits.pop(d)
                try:
                    item = await d
                    i, it = pair
447
                    awaits[_next_task(it, loop)] = pair
448
449
450
451
452
453
454
455
456
                    yield i, item
                except StopAsyncIteration:
                    pass
    finally:
        # Cancel any remaining iterators
        for f, (_, it) in awaits.items():
            with contextlib.suppress(BaseException):
                f.cancel()
                await it.aclose()
457
458


459
460
461
462
463
464
465
466
467
async def collect_from_async_generator(
        iterator: AsyncGenerator[T, None]) -> List[T]:
    """Collect all items from an async generator into a list."""
    items = []
    async for item in iterator:
        items.append(item)
    return items


468
def get_ip() -> str:
469
    host_ip = envs.VLLM_HOST_IP
470
471
472
473
474
475
476
    if "HOST_IP" in os.environ and "VLLM_HOST_IP" not in os.environ:
        logger.warning(
            "The environment variable HOST_IP is deprecated and ignored, as"
            " it is often used by Docker and other software to"
            "interact with the container's network stack. Please"
            "use VLLM_HOST_IP instead to set the IP address for vLLM processes"
            " to communicate with each other.")
477
478
479
480
481
    if host_ip:
        return host_ip

    # IP is not set, try to get it from the network interface

482
    # try ipv4
483
    s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
484
    try:
485
        s.connect(("8.8.8.8", 80))  # Doesn't need to be reachable
486
        return s.getsockname()[0]
487
488
489
490
491
    except Exception:
        pass

    # try ipv6
    try:
492
        s = socket.socket(socket.AF_INET6, socket.SOCK_DGRAM)
493
494
495
        # Google's public DNS server, see
        # https://developers.google.com/speed/public-dns/docs/using#addresses
        s.connect(("2001:4860:4860::8888", 80))  # Doesn't need to be reachable
496
        return s.getsockname()[0]
497
498
499
500
501
    except Exception:
        pass

    warnings.warn(
        "Failed to get the IP address, using 0.0.0.0 by default."
502
503
        "The value can be set by the environment variable"
        " VLLM_HOST_IP or HOST_IP.",
504
505
        stacklevel=2)
    return "0.0.0.0"
506
507


508
509
510
511
512
513
514
515
def is_valid_ipv6_address(address: str) -> bool:
    try:
        ipaddress.IPv6Address(address)
        return True
    except ValueError:
        return False


516
def get_distributed_init_method(ip: str, port: int) -> str:
517
518
519
    # Brackets are not permitted in ipv4 addresses,
    # see https://github.com/python/cpython/issues/103848
    return f"tcp://[{ip}]:{port}" if ":" in ip else f"tcp://{ip}:{port}"
520
521


522
523
524
525
526
527
528
def get_open_zmq_ipc_path() -> str:
    base_rpc_path = envs.VLLM_RPC_BASE_PATH
    return f"ipc://{base_rpc_path}/{uuid4()}"


def get_open_port() -> int:
    port = envs.VLLM_PORT
529
    if port is not None:
530
531
532
533
534
535
536
537
538
        while True:
            try:
                with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
                    s.bind(("", port))
                    return port
            except OSError:
                port += 1  # Increment port number if already in use
                logger.info("Port %d is already in use, trying port %d",
                            port - 1, port)
539
540
541
542
543
544
545
546
547
548
    # try ipv4
    try:
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            s.bind(("", 0))
            return s.getsockname()[1]
    except OSError:
        # try ipv6
        with socket.socket(socket.AF_INET6, socket.SOCK_STREAM) as s:
            s.bind(("", 0))
            return s.getsockname()[1]
549
550


551
552
553
554
555
556
557
558
559
560
def find_process_using_port(port: int) -> Optional[psutil.Process]:
    for conn in psutil.net_connections():
        if conn.laddr.port == port:
            try:
                return psutil.Process(conn.pid)
            except psutil.NoSuchProcess:
                return None
    return None


561
562
def update_environment_variables(envs: Dict[str, str]):
    for k, v in envs.items():
563
        if k in os.environ and os.environ[k] != v:
564
565
566
            logger.warning(
                "Overwriting environment variable %s "
                "from '%s' to '%s'", k, os.environ[k], v)
567
        os.environ[k] = v
568
569


570
def chunk_list(lst: List[T], chunk_size: int):
571
    """Yield successive chunk_size chunks from lst."""
572
573
    for i in range(0, len(lst), chunk_size):
        yield lst[i:i + chunk_size]
574
575
576
577
578
579
580


def cdiv(a: int, b: int) -> int:
    """Ceiling division."""
    return -(a // -b)


581
def _generate_random_fp8(
582
    tensor: torch.Tensor,
583
584
585
586
587
588
    low: float,
    high: float,
) -> None:
    # NOTE(zhaoyang): Due to NaN and Inf representation for fp8 data type,
    # it may occur Inf or NaN if we directly use torch.randint
    # to generate random data for fp8 data.
589
    # For example, s.11111.00 in fp8e5m2 format represents Inf.
590
591
592
593
    #     | E4M3        | E5M2
    #-----|-------------|-------------------
    # Inf | N/A         | s.11111.00
    # NaN | s.1111.111  | s.11111.{01,10,11}
594
    from vllm import _custom_ops as ops
595
596
    tensor_tmp = torch.empty_like(tensor, dtype=torch.float16)
    tensor_tmp.uniform_(low, high)
597
    ops.convert_fp8(tensor, tensor_tmp)
598
599
600
    del tensor_tmp


601
602
603
def get_kv_cache_torch_dtype(
        cache_dtype: Optional[Union[str, torch.dtype]],
        model_dtype: Optional[Union[str, torch.dtype]] = None) -> torch.dtype:
604
605
606
607
608
609
610
611
612
613
    if isinstance(cache_dtype, str):
        if cache_dtype == "auto":
            if isinstance(model_dtype, str):
                torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[model_dtype]
            elif isinstance(model_dtype, torch.dtype):
                torch_dtype = model_dtype
            else:
                raise ValueError(f"Invalid model dtype: {model_dtype}")
        elif cache_dtype in ["half", "bfloat16", "float"]:
            torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_dtype]
614
        elif cache_dtype == "fp8":
615
616
617
618
619
620
621
            torch_dtype = torch.uint8
        else:
            raise ValueError(f"Invalid kv cache dtype: {cache_dtype}")
    elif isinstance(cache_dtype, torch.dtype):
        torch_dtype = cache_dtype
    else:
        raise ValueError(f"Invalid kv cache dtype: {cache_dtype}")
622
623
624
625
626
627
628
629
630
631
632
633
634
635
    return torch_dtype


def create_kv_caches_with_random_flash(
    num_blocks: int,
    block_size: int,
    num_layers: int,
    num_heads: int,
    head_size: int,
    cache_dtype: Optional[Union[str, torch.dtype]],
    model_dtype: Optional[Union[str, torch.dtype]] = None,
    seed: int = 0,
    device: Optional[str] = "cuda",
) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
636
    current_platform.seed_everything(seed)
637
638
639
640

    torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
    key_value_cache_shape = (num_blocks, 2, block_size, num_heads, head_size)
    scale = head_size**-0.5
641
642
643
644

    key_caches: List[torch.Tensor] = []
    value_caches: List[torch.Tensor] = []

645
646
647
648
    for _ in range(num_layers):
        key_value_cache = torch.empty(size=key_value_cache_shape,
                                      dtype=torch_dtype,
                                      device=device)
649
650
651
652
653
654
655
        if cache_dtype in ["auto", "half", "bfloat16", "float"]:
            key_value_cache.uniform_(-scale, scale)
        elif cache_dtype == 'fp8':
            _generate_random_fp8(key_value_cache, -scale, scale)
        else:
            raise ValueError(
                f"Does not support key cache of type {cache_dtype}")
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
        key_caches.append(key_value_cache[:, 0])
        value_caches.append(key_value_cache[:, 1])
    return key_caches, value_caches


def create_kv_caches_with_random(
    num_blocks: int,
    block_size: int,
    num_layers: int,
    num_heads: int,
    head_size: int,
    cache_dtype: Optional[Union[str, torch.dtype]],
    model_dtype: Optional[Union[str, torch.dtype]] = None,
    seed: int = 0,
    device: Optional[str] = "cuda",
) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
Joe's avatar
Joe committed
672
673
674
675
676
677

    if cache_dtype == "fp8" and head_size % 16:
        raise ValueError(
            f"Does not support key cache of type fp8 with head_size {head_size}"
        )

678
    current_platform.seed_everything(seed)
679
680

    torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
681
682
683
684

    scale = head_size**-0.5
    x = 16 // torch.tensor([], dtype=torch_dtype).element_size()
    key_cache_shape = (num_blocks, num_heads, head_size // x, block_size, x)
685
    key_caches: List[torch.Tensor] = []
686
687
688
689
    for _ in range(num_layers):
        key_cache = torch.empty(size=key_cache_shape,
                                dtype=torch_dtype,
                                device=device)
690
        if cache_dtype in ["auto", "half", "bfloat16", "float"]:
691
            key_cache.uniform_(-scale, scale)
692
693
        elif cache_dtype == 'fp8':
            _generate_random_fp8(key_cache, -scale, scale)
694
695
696
        else:
            raise ValueError(
                f"Does not support key cache of type {cache_dtype}")
697
698
699
        key_caches.append(key_cache)

    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
700
    value_caches: List[torch.Tensor] = []
701
702
703
704
    for _ in range(num_layers):
        value_cache = torch.empty(size=value_cache_shape,
                                  dtype=torch_dtype,
                                  device=device)
705
        if cache_dtype in ["auto", "half", "bfloat16", "float"]:
706
            value_cache.uniform_(-scale, scale)
707
708
        elif cache_dtype == 'fp8':
            _generate_random_fp8(value_cache, -scale, scale)
709
710
711
        else:
            raise ValueError(
                f"Does not support value cache of type {cache_dtype}")
712
713
        value_caches.append(value_cache)
    return key_caches, value_caches
714
715


716
717
718
719
720
721
@lru_cache
def print_info_once(msg: str) -> None:
    # Set the stacklevel to 2 to print the caller's line info
    logger.info(msg, stacklevel=2)


722
723
@lru_cache
def print_warning_once(msg: str) -> None:
724
725
    # Set the stacklevel to 2 to print the caller's line info
    logger.warning(msg, stacklevel=2)
726
727
728
729
730
731
732
733
734
735
736


@lru_cache(maxsize=None)
def is_pin_memory_available() -> bool:

    if in_wsl():
        # Pinning memory in WSL is not supported.
        # https://docs.nvidia.com/cuda/wsl-user-guide/index.html#known-limitations-for-linux-cuda-applications
        print_warning_once("Using 'pin_memory=False' as WSL is detected. "
                           "This may slow down the performance.")
        return False
737
    elif current_platform.is_xpu():
738
739
        print_warning_once("Pin memory is not supported on XPU.")
        return False
740
    elif current_platform.is_neuron():
741
742
        print_warning_once("Pin memory is not supported on Neuron.")
        return False
743
744
745
    elif current_platform.is_hpu():
        print_warning_once("Pin memory is not supported on HPU.")
        return False
746
    elif current_platform.is_cpu() or current_platform.is_openvino():
747
        return False
748
749
750
    return True


751
class DeviceMemoryProfiler:
752

753
    def __init__(self, device: Optional[torch.types.Device] = None):
754
755
756
757
        self.device = device

    def current_memory_usage(self) -> float:
        # Return the memory usage in bytes.
758
        if current_platform.is_cuda_alike():
759
760
            torch.cuda.reset_peak_memory_stats(self.device)
            mem = torch.cuda.max_memory_allocated(self.device)
761
        elif current_platform.is_xpu():
762
763
            torch.xpu.reset_peak_memory_stats(self.device)  # type: ignore
            mem = torch.xpu.max_memory_allocated(self.device)  # type: ignore
764
765
766
767
768
769
770
771
772
773
774
775
776
        return mem

    def __enter__(self):
        self.initial_memory = self.current_memory_usage()
        # This allows us to call methods of the context manager if needed
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.final_memory = self.current_memory_usage()
        self.consumed_memory = self.final_memory - self.initial_memory

        # Force garbage collection
        gc.collect()
777
778


779
780
781
782
783
784
785
786
787
def make_ndarray_with_pad(
    x: List[List[T]],
    pad: T,
    dtype: npt.DTypeLike,
    *,
    max_len: Optional[int] = None,
) -> npt.NDArray:
    """
    Make a padded array from 2D inputs.
788
789
790
791

    The padding is applied to the end of each inner list until it reaches
    `max_len`.
    """
792
793
794
795
796
    if max_len is None:
        # Unlike for most functions, map is faster than a genexpr over `len`
        max_len = max(map(len, x), default=0)

    padded_x = np.full((len(x), max_len), pad, dtype=dtype)
797
798
799
    for ind, blocktb in enumerate(x):
        assert len(blocktb) <= max_len
        padded_x[ind, :len(blocktb)] = blocktb
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826

    return padded_x


def make_tensor_with_pad(
    x: List[List[T]],
    pad: T,
    dtype: torch.dtype,
    *,
    max_len: Optional[int] = None,
    device: Optional[Union[str, torch.device]] = None,
    pin_memory: bool = False,
) -> torch.Tensor:
    """
    Make a padded tensor from 2D inputs.

    The padding is applied to the end of each inner list until it reaches
    `max_len`.
    """
    np_dtype = TORCH_DTYPE_TO_NUMPY_DTYPE[dtype]
    padded_x = make_ndarray_with_pad(x, pad, np_dtype, max_len=max_len)

    tensor = torch.from_numpy(padded_x).to(device)
    if pin_memory:
        tensor = tensor.pin_memory()

    return tensor
827
828
829
830
831
832
833
834
835
836
837
838
839


def async_tensor_h2d(
    data: list,
    dtype: torch.dtype,
    target_device: Union[str, torch.device],
    pin_memory: bool,
) -> torch.Tensor:
    """Asynchronously create a tensor and copy it from host to device."""
    t = torch.tensor(data, dtype=dtype, pin_memory=pin_memory, device="cpu")
    return t.to(device=target_device, non_blocking=True)


840
841
842
843
844
def get_dtype_size(dtype: torch.dtype) -> int:
    """Get the size of the data type in bytes."""
    return torch.tensor([], dtype=dtype).element_size()


845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
# `collections` helpers
def is_list_of(
    value: object,
    typ: Type[T],
    *,
    check: Literal["first", "all"] = "first",
) -> TypeIs[List[T]]:
    if not isinstance(value, list):
        return False

    if check == "first":
        return len(value) == 0 or isinstance(value[0], typ)
    elif check == "all":
        return all(isinstance(v, typ) for v in value)

    assert_never(check)


863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
JSONTree = Union[Dict[str, "JSONTree[T]"], List["JSONTree[T]"],
                 Tuple["JSONTree[T]", ...], T]
"""A nested JSON structure where the leaves need not be JSON-serializable."""


@overload
def json_map_leaves(
    func: Callable[[T], U],
    value: Dict[str, JSONTree[T]],
) -> Dict[str, JSONTree[U]]:
    ...


@overload
def json_map_leaves(
    func: Callable[[T], U],
    value: List[JSONTree[T]],
) -> List[JSONTree[U]]:
    ...


@overload
def json_map_leaves(
    func: Callable[[T], U],
    value: Tuple[JSONTree[T], ...],
) -> Tuple[JSONTree[U], ...]:
    ...


@overload
def json_map_leaves(
    func: Callable[[T], U],
    value: JSONTree[T],
) -> JSONTree[U]:
    ...


def json_map_leaves(func: Callable[[T], U], value: JSONTree[T]) -> JSONTree[U]:
    if isinstance(value, dict):
        return {k: json_map_leaves(func, v) for k, v in value.items()}
    elif isinstance(value, list):
        return [json_map_leaves(func, v) for v in value]
    elif isinstance(value, tuple):
        return tuple(json_map_leaves(func, v) for v in value)
    else:
        return func(value)


911
912
913
914
915
def flatten_2d_lists(lists: List[List[T]]) -> List[T]:
    """Flatten a list of lists to a single list."""
    return [item for sublist in lists for item in sublist]


916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
_K = TypeVar("_K", bound=Hashable)
_V = TypeVar("_V")


def full_groupby(values: Iterable[_V], *, key: Callable[[_V], _K]):
    """
    Unlike :class:`itertools.groupby`, groups are not broken by
    non-contiguous data.
    """
    groups = defaultdict[_K, list[_V]](list)

    for value in values:
        groups[key(value)].append(value)

    return groups.items()


933
934
# TODO: This function can be removed if transformer_modules classes are
# serialized by value when communicating between processes
935
def init_cached_hf_modules() -> None:
936
937
938
939
940
    """
    Lazy initialization of the Hugging Face modules.
    """
    from transformers.dynamic_module_utils import init_hf_modules
    init_hf_modules()
941
942


943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
@lru_cache(maxsize=None)
def find_library(lib_name: str) -> str:
    """
    Find the library file in the system.
    `lib_name` is full filename, with both prefix and suffix.
    This function resolves `lib_name` to the full path of the library.
    """
    # Adapted from https://github.com/openai/triton/blob/main/third_party/nvidia/backend/driver.py#L19 # noqa
    # According to https://en.wikipedia.org/wiki/Filesystem_Hierarchy_Standard
    # `/sbin/ldconfig` should exist in all Linux systems.
    # `/sbin/ldconfig` searches the library in the system
    libs = subprocess.check_output(["/sbin/ldconfig", "-p"]).decode()
    # each line looks like the following:
    # libcuda.so.1 (libc6,x86-64) => /lib/x86_64-linux-gnu/libcuda.so.1
    locs = [line.split()[-1] for line in libs.splitlines() if lib_name in line]
    # `LD_LIBRARY_PATH` searches the library in the user-defined paths
959
    env_ld_library_path = envs.LD_LIBRARY_PATH
960
961
962
963
964
965
966
967
968
969
970
    if not locs and env_ld_library_path:
        locs = [
            os.path.join(dir, lib_name)
            for dir in env_ld_library_path.split(":")
            if os.path.exists(os.path.join(dir, lib_name))
        ]
    if not locs:
        raise ValueError(f"Cannot find {lib_name} in the system.")
    return locs[0]


971
def find_nccl_library() -> str:
972
973
974
975
976
977
    """
    We either use the library file specified by the `VLLM_NCCL_SO_PATH`
    environment variable, or we find the library file brought by PyTorch.
    After importing `torch`, `libnccl.so.2` or `librccl.so.1` can be
    found by `ctypes` automatically.
    """
978
    so_file = envs.VLLM_NCCL_SO_PATH
979
980
981
982

    # manually load the nccl library
    if so_file:
        logger.info(
983
984
            "Found nccl from environment variable VLLM_NCCL_SO_PATH=%s",
            so_file)
985
986
    else:
        if torch.version.cuda is not None:
987
            so_file = "libnccl.so.2"
988
        elif torch.version.hip is not None:
989
            so_file = "librccl.so.1"
990
991
        else:
            raise ValueError("NCCL only supports CUDA and ROCm backends.")
992
        logger.info("Found nccl from library %s", so_file)
993
    return so_file
994
995
996
997
998
999
1000


def enable_trace_function_call_for_thread() -> None:
    """Set up function tracing for the current thread,
    if enabled via the VLLM_TRACE_FUNCTION environment variable
    """

1001
    if envs.VLLM_TRACE_FUNCTION:
1002
        tmp_dir = tempfile.gettempdir()
1003
1004
        # add username to tmp_dir to avoid permission issues
        tmp_dir = os.path.join(tmp_dir, getpass.getuser())
1005
1006
1007
1008
1009
1010
1011
        filename = (f"VLLM_TRACE_FUNCTION_for_process_{os.getpid()}"
                    f"_thread_{threading.get_ident()}_"
                    f"at_{datetime.datetime.now()}.log").replace(" ", "_")
        log_path = os.path.join(tmp_dir, "vllm", get_vllm_instance_id(),
                                filename)
        os.makedirs(os.path.dirname(log_path), exist_ok=True)
        enable_trace_function_call(log_path)
1012
1013


1014
# `functools` helpers
1015
1016
def identity(value: T, **kwargs) -> T:
    """Returns the first provided value."""
1017
1018
1019
1020
1021
1022
    return value


F = TypeVar('F', bound=Callable[..., Any])


1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
def deprecate_args(
    start_index: int,
    is_deprecated: Union[bool, Callable[[], bool]] = True,
    additional_message: Optional[str] = None,
) -> Callable[[F], F]:

    if not callable(is_deprecated):
        is_deprecated = partial(identity, is_deprecated)

    def wrapper(fn: F) -> F:

        params = inspect.signature(fn).parameters
        pos_types = (
            inspect.Parameter.POSITIONAL_ONLY,
            inspect.Parameter.POSITIONAL_OR_KEYWORD,
        )
        pos_kws = [
            kw for kw, param in params.items() if param.kind in pos_types
        ]

        @wraps(fn)
        def inner(*args, **kwargs):
            if is_deprecated():
                deprecated_args = pos_kws[start_index:len(args)]
                if deprecated_args:
                    msg = (
                        f"The positional arguments {deprecated_args} are "
                        "deprecated and will be removed in a future update.")
                    if additional_message is not None:
                        msg += f" {additional_message}"

                    warnings.warn(
                        DeprecationWarning(msg),
                        stacklevel=3,  # The inner function takes up one level
                    )

            return fn(*args, **kwargs)

        return inner  # type: ignore

    return wrapper


1066
def deprecate_kwargs(
1067
1068
1069
1070
    *kws: str,
    is_deprecated: Union[bool, Callable[[], bool]] = True,
    additional_message: Optional[str] = None,
) -> Callable[[F], F]:
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
    deprecated_kws = set(kws)

    if not callable(is_deprecated):
        is_deprecated = partial(identity, is_deprecated)

    def wrapper(fn: F) -> F:

        @wraps(fn)
        def inner(*args, **kwargs):
            if is_deprecated():
                deprecated_kwargs = kwargs.keys() & deprecated_kws
                if deprecated_kwargs:
                    msg = (
                        f"The keyword arguments {deprecated_kwargs} are "
                        "deprecated and will be removed in a future update.")
                    if additional_message is not None:
                        msg += f" {additional_message}"

                    warnings.warn(
                        DeprecationWarning(msg),
                        stacklevel=3,  # The inner function takes up one level
                    )

            return fn(*args, **kwargs)

        return inner  # type: ignore

    return wrapper
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115


@lru_cache(maxsize=8)
def _cuda_device_count_stateless(
        cuda_visible_devices: Optional[str] = None) -> int:
    # Note: cuda_visible_devices is not used, but we keep it as an argument for
    # LRU Cache purposes.

    # Code below is based on
    # https://github.com/pytorch/pytorch/blob/
    # c1cd946818442aca8c7f812b16d187ce1586c3bc/
    # torch/cuda/__init__.py#L831C1-L831C17
    import torch.cuda
    import torch.version

    if not torch.cuda._is_compiled():
        return 0
1116
    if current_platform.is_rocm():
1117
1118
1119
1120
1121
1122
1123
        # ROCm uses amdsmi instead of nvml for stateless device count
        # This requires a sufficiently modern version of Torch 2.4.0
        raw_count = torch.cuda._device_count_amdsmi() if (hasattr(
            torch.cuda, "_device_count_amdsmi")) else -1
    else:
        raw_count = torch.cuda._device_count_nvml()
    r = torch._C._cuda_getDeviceCount() if raw_count < 0 else raw_count
1124
1125
1126
1127
1128
1129
    return r


def cuda_device_count_stateless() -> int:
    """Get number of CUDA devices, caching based on the value of
    CUDA_VISIBLE_DEVICES at the time of call.
1130

1131
1132
1133
1134
1135
1136
1137
    This should be used instead of torch.cuda.device_count()
    unless CUDA_VISIBLE_DEVICES has already been set to the desired
    value."""

    # This can be removed and simply replaced with torch.cuda.get_device_count
    # after https://github.com/pytorch/pytorch/pull/122815 is released.
    return _cuda_device_count_stateless(envs.CUDA_VISIBLE_DEVICES)
1138
1139


1140
1141
1142
1143
1144
1145
1146
def cuda_is_initialized() -> bool:
    """Check if CUDA is initialized."""
    if not torch.cuda._is_compiled():
        return False
    return torch.cuda.is_initialized()


1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
def weak_bind(bound_method: Callable[..., Any], ) -> Callable[..., None]:
    """Make an instance method that weakly references
    its associated instance and no-ops once that
    instance is collected."""
    ref = weakref.ref(bound_method.__self__)  # type: ignore[attr-defined]
    unbound = bound_method.__func__  # type: ignore[attr-defined]

    def weak_bound(*args, **kwargs) -> None:
        if inst := ref():
            unbound(inst, *args, **kwargs)

    return weak_bound


1161
#From: https://stackoverflow.com/a/4104188/2749989
1162
def run_once(f: Callable[P, None]) -> Callable[P, None]:
1163

1164
    def wrapper(*args: P.args, **kwargs: P.kwargs) -> None:
1165
1166
1167
1168
1169
1170
        if not wrapper.has_run:  # type: ignore[attr-defined]
            wrapper.has_run = True  # type: ignore[attr-defined]
            return f(*args, **kwargs)

    wrapper.has_run = False  # type: ignore[attr-defined]
    return wrapper
1171
1172


1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
class StoreBoolean(argparse.Action):

    def __call__(self, parser, namespace, values, option_string=None):
        if values.lower() == "true":
            setattr(namespace, self.dest, True)
        elif values.lower() == "false":
            setattr(namespace, self.dest, False)
        else:
            raise ValueError(f"Invalid boolean value: {values}. "
                             "Expected 'true' or 'false'.")


1185
1186
1187
1188
1189
class SortedHelpFormatter(argparse.HelpFormatter):
    """SortedHelpFormatter that sorts arguments by their option strings."""

    def add_arguments(self, actions):
        actions = sorted(actions, key=lambda x: x.option_strings)
1190
        super().add_arguments(actions)
1191
1192


1193
1194
1195
class FlexibleArgumentParser(argparse.ArgumentParser):
    """ArgumentParser that allows both underscore and dash in names."""

1196
1197
1198
1199
1200
1201
    def __init__(self, *args, **kwargs):
        # Set the default 'formatter_class' to SortedHelpFormatter
        if 'formatter_class' not in kwargs:
            kwargs['formatter_class'] = SortedHelpFormatter
        super().__init__(*args, **kwargs)

1202
1203
1204
1205
    def parse_args(self, args=None, namespace=None):
        if args is None:
            args = sys.argv[1:]

1206
        if '--config' in args:
1207
            args = self._pull_args_from_config(args)
1208

1209
1210
1211
1212
        # Convert underscores to dashes and vice versa in argument names
        processed_args = []
        for arg in args:
            if arg.startswith('--'):
1213
1214
1215
1216
1217
1218
1219
                if '=' in arg:
                    key, value = arg.split('=', 1)
                    key = '--' + key[len('--'):].replace('_', '-')
                    processed_args.append(f'{key}={value}')
                else:
                    processed_args.append('--' +
                                          arg[len('--'):].replace('_', '-'))
1220
1221
1222
1223
            elif arg.startswith('-O') and arg != '-O' and len(arg) == 2:
                # allow -O flag to be used without space, e.g. -O3
                processed_args.append('-O')
                processed_args.append(arg[2:])
1224
1225
1226
1227
            else:
                processed_args.append(arg)

        return super().parse_args(processed_args, namespace)
1228

1229
    def _pull_args_from_config(self, args: List[str]) -> List[str]:
1230
1231
        """Method to pull arguments specified in the config file
        into the command-line args variable.
1232
1233

        The arguments in config file will be inserted between
1234
        the argument list.
1235

1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
        example:
        ```yaml
            port: 12323
            tensor-parallel-size: 4
        ```
        ```python
        $: vllm {serve,chat,complete} "facebook/opt-12B" \
            --config config.yaml -tp 2
        $: args = [
            "serve,chat,complete",
1246
1247
            "facebook/opt-12B",
            '--config', 'config.yaml',
1248
1249
1250
1251
            '-tp', '2'
        ]
        $: args = [
            "serve,chat,complete",
1252
1253
1254
            "facebook/opt-12B",
            '--port', '12323',
            '--tensor-parallel-size', '4',
1255
1256
1257
1258
1259
            '-tp', '2'
            ]
        ```

        Please note how the config args are inserted after the sub command.
1260
        this way the order of priorities is maintained when these are args
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
        parsed by super().
        """
        assert args.count(
            '--config') <= 1, "More than one config file specified!"

        index = args.index('--config')
        if index == len(args) - 1:
            raise ValueError("No config file specified! \
                             Please check your command-line arguments.")

        file_path = args[index + 1]

1273
        config_args = self._load_config_file(file_path)
1274
1275

        # 0th index is for {serve,chat,complete}
1276
        # followed by model_tag (only for serve)
1277
1278
1279
1280
        # followed by config args
        # followed by rest of cli args.
        # maintaining this order will enforce the precedence
        # of cli > config > defaults
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
        if args[0] == "serve":
            if index == 1:
                raise ValueError(
                    "No model_tag specified! Please check your command-line"
                    " arguments.")
            args = [args[0]] + [
                args[1]
            ] + config_args + args[2:index] + args[index + 2:]
        else:
            args = [args[0]] + config_args + args[1:index] + args[index + 2:]
1291
1292
1293

        return args

1294
    def _load_config_file(self, file_path: str) -> List[str]:
1295
        """Loads a yaml file and returns the key value pairs as a
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
        flattened list with argparse like pattern
        ```yaml
            port: 12323
            tensor-parallel-size: 4
        ```
        returns:
            processed_args: list[str] = [
                '--port': '12323',
                '--tensor-parallel-size': '4'
            ]
1306

1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
        """

        extension: str = file_path.split('.')[-1]
        if extension not in ('yaml', 'yml'):
            raise ValueError(
                "Config file must be of a yaml/yml type.\
                              %s supplied", extension)

        # only expecting a flat dictionary of atomic types
        processed_args: List[str] = []

        config: Dict[str, Union[int, str]] = {}
        try:
1320
            with open(file_path) as config_file:
1321
1322
1323
1324
1325
1326
1327
                config = yaml.safe_load(config_file)
        except Exception as ex:
            logger.error(
                "Unable to read the config file at %s. \
                Make sure path is correct", file_path)
            raise ex

1328
1329
1330
1331
1332
        store_boolean_arguments = [
            action.dest for action in self._actions
            if isinstance(action, StoreBoolean)
        ]

1333
        for key, value in config.items():
1334
1335
1336
1337
1338
1339
            if isinstance(value, bool) and key not in store_boolean_arguments:
                if value:
                    processed_args.append('--' + key)
            else:
                processed_args.append('--' + key)
                processed_args.append(str(value))
1340
1341
1342

        return processed_args

1343
1344
1345
1346
1347
1348

async def _run_task_with_lock(task: Callable, lock: asyncio.Lock, *args,
                              **kwargs):
    """Utility function to run async task in a lock"""
    async with lock:
        return await task(*args, **kwargs)
1349
1350


1351
1352
1353
1354
1355
1356
1357
1358
1359
def supports_kw(
    callable: Callable[..., object],
    kw_name: str,
    requires_kw_only: bool = False,
    allow_var_kwargs: bool = True,
) -> bool:
    """Check if a keyword is a valid kwarg for a callable; if requires_kw_only
    disallows kwargs names that can also be positional arguments.
    """
1360
    params = inspect.signature(callable).parameters
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
    if not params:
        return False

    param_val = params.get(kw_name)

    # Types where the it may be valid, i.e., explicitly defined & nonvariadic
    passable_kw_types = set((inspect.Parameter.POSITIONAL_ONLY,
                             inspect.Parameter.POSITIONAL_OR_KEYWORD,
                             inspect.Parameter.KEYWORD_ONLY))

    if param_val:
        is_sig_param = param_val.kind in passable_kw_types
        # We want kwargs only, but this is passable as a positional arg
        if (requires_kw_only and is_sig_param
                and param_val.kind != inspect.Parameter.KEYWORD_ONLY):
            return False
        if ((requires_kw_only
             and param_val.kind == inspect.Parameter.KEYWORD_ONLY)
                or (not requires_kw_only and is_sig_param)):
            return True

    # If we're okay with var-kwargs, it's supported as long as
    # the kw_name isn't something like *args, **kwargs
    if allow_var_kwargs:
        # Get the last param; type is ignored here because params is a proxy
        # mapping, but it wraps an ordered dict, and they appear in order.
        # Ref: https://docs.python.org/3/library/inspect.html#inspect.Signature.parameters
        last_param = params[next(reversed(params))]  # type: ignore
        return (last_param.kind == inspect.Parameter.VAR_KEYWORD
                and last_param.name != kw_name)
    return False


def resolve_mm_processor_kwargs(
    init_kwargs: Optional[Dict[str, Any]],
    inference_kwargs: Optional[Dict[str, Any]],
    callable: Callable[..., object],
    allow_var_kwargs: bool = False,
) -> Dict[str, Any]:
    """Applies filtering to eliminate invalid mm_processor_kwargs, i.e.,
    those who are not explicit keywords to the given callable (of one is
    given; otherwise no filtering is done), then merges the kwarg dicts,
    giving priority to inference_kwargs if there are any collisions.

    In the case that no kwarg overrides are provided, returns an empty
    dict so that it can still be kwarg expanded into the callable later on.

    If allow_var_kwargs=True, allows for things that can be expanded into
    kwargs as long as they aren't naming collision for var_kwargs or potential
    positional arguments.
    """
    # Filter inference time multimodal processor kwargs provided
    runtime_mm_kwargs = get_allowed_kwarg_only_overrides(
        callable,
        overrides=inference_kwargs,
        allow_var_kwargs=allow_var_kwargs)

    # Filter init time multimodal processor kwargs provided
    init_mm_kwargs = get_allowed_kwarg_only_overrides(
        callable, overrides=init_kwargs, allow_var_kwargs=allow_var_kwargs)
1421

1422
1423
1424
1425
    # Merge the final processor kwargs, prioritizing inference
    # time values over the initialization time values.
    mm_processor_kwargs = {**init_mm_kwargs, **runtime_mm_kwargs}
    return mm_processor_kwargs
1426
1427


1428
1429
1430
def get_allowed_kwarg_only_overrides(
    callable: Callable[..., object],
    overrides: Optional[Dict[str, Any]],
1431
    allow_var_kwargs: bool = False,
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
) -> Dict[str, Any]:
    """
    Given a callable which has one or more keyword only params and a dict
    mapping param names to values, drop values that can be not be kwarg
    expanded to overwrite one or more keyword-only args. This is used in a
    few places to handle custom processor overrides for multimodal models,
    e.g., for profiling when processor options provided by the user
    may affect the number of mm tokens per instance.

    Args:
        callable: Callable which takes 0 or more keyword only arguments.
1443
                  If None is provided, all overrides names are allowed.
1444
        overrides: Potential overrides to be used when invoking the callable.
1445
        allow_var_kwargs: Allows overrides that are expandable for var kwargs.
1446
1447
1448
1449
1450
1451
1452
1453
1454

    Returns:
        Dictionary containing the kwargs to be leveraged which may be used
        to overwrite one or more keyword only arguments when invoking the
        callable.
    """
    if not overrides:
        return {}

1455
1456
    # Drop any mm_processor_kwargs provided by the user that
    # are not kwargs, unless it can fit it var_kwargs param
1457
1458
1459
    filtered_overrides = {
        kwarg_name: val
        for kwarg_name, val in overrides.items()
1460
1461
1462
1463
        if supports_kw(callable,
                       kwarg_name,
                       requires_kw_only=True,
                       allow_var_kwargs=allow_var_kwargs)
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
    }

    # If anything is dropped, log a warning
    dropped_keys = overrides.keys() - filtered_overrides.keys()
    if dropped_keys:
        logger.warning(
            "The following intended overrides are not keyword-only args "
            "and and will be dropped: %s", dropped_keys)

    return filtered_overrides


1476
1477
1478
1479
1480
1481
# Using dynamo with vLLM doesn't really work well with PyTorch versions < 2.4.0.
# In particular, the FakeScalarType is not supported for earlier versions of
# PyTorch which breaks dynamo for any ops registered using ScalarType.
def supports_dynamo() -> bool:
    base_torch_version = Version(Version(torch.__version__).base_version)
    return base_torch_version >= Version("2.4.0")
1482
1483


1484
1485
1486
1487
1488
1489
# Some backends use pytorch version < 2.4.0 which doesn't
# support `torch.library.custom_op`.
def supports_custom_op() -> bool:
    return hasattr(torch.library, "custom_op")


1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
class AtomicCounter:
    """An atomic, thread-safe counter"""

    def __init__(self, initial=0):
        """Initialize a new atomic counter to given initial value"""
        self._value = initial
        self._lock = threading.Lock()

    def inc(self, num=1):
        """Atomically increment the counter by num and return the new value"""
        with self._lock:
            self._value += num
            return self._value

    def dec(self, num=1):
        """Atomically decrement the counter by num and return the new value"""
        with self._lock:
            self._value -= num
            return self._value

    @property
    def value(self):
        return self._value
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528


# Adapted from: https://stackoverflow.com/a/47212782/5082708
class LazyDict(Mapping, Generic[T]):

    def __init__(self, factory: Dict[str, Callable[[], T]]):
        self._factory = factory
        self._dict: Dict[str, T] = {}

    def __getitem__(self, key) -> T:
        if key not in self._dict:
            if key not in self._factory:
                raise KeyError(key)
            self._dict[key] = self._factory[key]()
        return self._dict[key]

1529
1530
1531
    def __setitem__(self, key: str, value: Callable[[], T]):
        self._factory[key] = value

1532
1533
1534
1535
1536
    def __iter__(self):
        return iter(self._factory)

    def __len__(self):
        return len(self._factory)
1537
1538
1539
1540
1541
1542
1543
1544
1545


def weak_ref_tensor(tensor: torch.Tensor) -> torch.Tensor:
    """
    Create a weak reference to a tensor.
    The new tensor will share the same data as the original tensor,
    but will not keep the original tensor alive.
    """
    return torch.ops._C.weak_ref_tensor(tensor)
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561


def weak_ref_tensors(
    tensors: Union[torch.Tensor, List[torch.Tensor], Tuple[torch.Tensor]]
) -> Union[torch.Tensor, List[torch.Tensor], Tuple[torch.Tensor]]:
    """
    Convenience function to create weak references to tensors,
    for single tensor, list of tensors or tuple of tensors.
    """
    if isinstance(tensors, torch.Tensor):
        return weak_ref_tensor(tensors)
    if isinstance(tensors, list):
        return [weak_ref_tensor(t) for t in tensors]
    if isinstance(tensors, tuple):
        return tuple(weak_ref_tensor(t) for t in tensors)
    raise ValueError("Invalid type for tensors")
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571


def is_in_doc_build() -> bool:
    try:
        from sphinx.ext.autodoc.mock import _MockModule
        return isinstance(torch, _MockModule)
    except ModuleNotFoundError:
        return False


1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
def import_from_path(module_name: str, file_path: Union[str, os.PathLike]):
    """
    Import a Python file according to its file path.

    Based on the official recipe:
    https://docs.python.org/3/library/importlib.html#importing-a-source-file-directly
    """
    spec = importlib.util.spec_from_file_location(module_name, file_path)
    if spec is None:
        raise ModuleNotFoundError(f"No module named '{module_name}'")

    assert spec.loader is not None

    module = importlib.util.module_from_spec(spec)
    sys.modules[module_name] = module
    spec.loader.exec_module(module)
    return module


1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
# create a library to hold the custom op
vllm_lib = Library("vllm", "FRAGMENT")  # noqa


def direct_register_custom_op(
    op_name: str,
    op_func: Callable,
    mutates_args: List[str],
    fake_impl: Optional[Callable] = None,
    target_lib: Optional[Library] = None,
1601
    dispatch_key: str = "CUDA",
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
):
    """
    `torch.library.custom_op` can have significant overhead because it
    needs to consider complicated dispatching logic. This function
    directly registers a custom op and dispatches it to the CUDA backend.
    See https://gist.github.com/youkaichao/ecbea9ec9fc79a45d2adce1784d7a9a5
    for more details.

    By default, the custom op is registered to the vLLM library. If you
    want to register it to a different library, you can pass the library
    object to the `target_lib` argument.

    IMPORTANT: the lifetime of the operator is tied to the lifetime of the
    library object. If you want to bind the operator to a different library,
    make sure the library object is alive when the operator is used.
    """
    if is_in_doc_build():
        return
1620
1621
1622
1623
1624
1625
1626
1627
    import torch.library
    if hasattr(torch.library, "infer_schema"):
        schema_str = torch.library.infer_schema(op_func,
                                                mutates_args=mutates_args)
    else:
        # for pytorch 2.4
        import torch._custom_op.impl
        schema_str = torch._custom_op.impl.infer_schema(op_func, mutates_args)
1628
1629
    my_lib = target_lib or vllm_lib
    my_lib.define(op_name + schema_str)
1630
    my_lib.impl(op_name, op_func, dispatch_key=dispatch_key)
1631
1632
    if fake_impl is not None:
        my_lib._register_fake(op_name, fake_impl)
1633
1634
1635
1636
1637
1638
1639
1640
1641


def resolve_obj_by_qualname(qualname: str) -> Any:
    """
    Resolve an object by its fully qualified name.
    """
    module_name, obj_name = qualname.rsplit(".", 1)
    module = importlib.import_module(module_name)
    return getattr(module, obj_name)