utils.py 41.6 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 inspect
8
import os
9
import random
10
import socket
11
import subprocess
12
import sys
13
14
import tempfile
import threading
Zhuohan Li's avatar
Zhuohan Li committed
15
import uuid
16
import warnings
17
import weakref
18
from asyncio import FIRST_COMPLETED, ensure_future
19
from functools import lru_cache, partial, wraps
20
from platform import uname
21
from typing import (Any, AsyncGenerator, Awaitable, Callable, Dict, Generic,
22
23
                    Hashable, List, Literal, Optional, OrderedDict, Set, Tuple,
                    Type, TypeVar, Union, overload)
24
from uuid import uuid4
Zhuohan Li's avatar
Zhuohan Li committed
25

26
import numpy as np
27
import numpy.typing as npt
28
import psutil
Zhuohan Li's avatar
Zhuohan Li committed
29
import torch
30
import torch.types
31
import yaml
32
from packaging.version import Version
33
from typing_extensions import ParamSpec, TypeIs, assert_never
34

35
import vllm.envs as envs
36
from vllm.logger import enable_trace_function_call, init_logger
37
from vllm.platforms import current_platform
38
39
40

logger = init_logger(__name__)

41
42
43
44
45
46
47
48
49
50
51
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
# Exception strings for non-implemented encoder/decoder scenarios

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.")

STR_NOT_IMPL_ENC_DEC_BACKEND = ("XFormers is the only backend "
                                "currently supported with encoder/"
                                "decoder models.")

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

85
86
87
STR_NOT_IMPL_ENC_DEC_CPU = ("CPU is not currently supported with "
                            "encoder/decoder models.")

88
89
90
91
92
93
94
95
96
97
98
99
100
101
# 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,
102
    "STR_NOT_IMPL_ENC_DEC_CPU": STR_NOT_IMPL_ENC_DEC_CPU
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
}

# 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"

121
122
123
GiB_bytes = 1 << 30
"""The number of bytes in one gibibyte (GiB)."""

124
125
126
127
STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.half,
    "bfloat16": torch.bfloat16,
    "float": torch.float,
128
    "fp8": torch.uint8,
129
130
    "fp8_e4m3": torch.uint8,
    "fp8_e5m2": torch.uint8,
131
}
Zhuohan Li's avatar
Zhuohan Li committed
132

133
134
135
136
137
138
139
140
141
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,
}

142
143
144
P = ParamSpec('P')
K = TypeVar("K")
T = TypeVar("T")
145
U = TypeVar("U")
146

Woosuk Kwon's avatar
Woosuk Kwon committed
147

148
149
150
151
152
153
154
class _Sentinel:
    ...


ALL_PINNED_SENTINEL = _Sentinel()


Woosuk Kwon's avatar
Woosuk Kwon committed
155
156
157
158
159
160
161
162
163
164
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
165
    def __next__(self) -> int:
166
        i = self.counter
Woosuk Kwon's avatar
Woosuk Kwon committed
167
        self.counter += 1
168
        return i
Woosuk Kwon's avatar
Woosuk Kwon committed
169
170
171

    def reset(self) -> None:
        self.counter = 0
Zhuohan Li's avatar
Zhuohan Li committed
172

173

174
class LRUCache(Generic[T]):
175
176

    def __init__(self, capacity: int):
177
        self.cache: OrderedDict[Hashable, T] = OrderedDict()
178
        self.pinned_items: Set[Hashable] = set()
179
180
181
182
183
184
185
186
        self.capacity = capacity

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

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

187
188
189
190
    def __getitem__(self, key: Hashable) -> T:
        value = self.cache[key]  # Raise KeyError if not exists
        self.cache.move_to_end(key)
        return value
191

192
    def __setitem__(self, key: Hashable, value: T) -> None:
193
194
195
196
197
198
199
200
        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)

201
202
203
    def get(self,
            key: Hashable,
            default_value: Optional[T] = None) -> Optional[T]:
204
        value: Optional[T]
205
        if key in self.cache:
206
            value = self.cache[key]
207
208
209
210
211
            self.cache.move_to_end(key)
        else:
            value = default_value
        return value

212
    def put(self, key: Hashable, value: T) -> None:
213
214
215
216
        self.cache[key] = value
        self.cache.move_to_end(key)
        self._remove_old_if_needed()

217
218
219
220
221
222
223
224
225
226
227
228
    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)

229
    def _on_remove(self, key: Hashable, value: Optional[T]):
230
231
        pass

232
    def remove_oldest(self, remove_pinned=False):
233
234
        if not self.cache:
            return
235
236
237
238
239
240
241
242
243
244
245
246

        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)
247
248
249
250
251

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

252
253
254
    def pop(self,
            key: Hashable,
            default_value: Optional[T] = None) -> Optional[T]:
255
        run_on_remove = key in self.cache
256
        value: Optional[T] = self.cache.pop(key, default_value)
257
258
259
        # remove from pinned items
        if key in self.pinned_items:
            self._unpin(key)
260
261
262
263
264
265
        if run_on_remove:
            self._on_remove(key, value)
        return value

    def clear(self):
        while len(self.cache) > 0:
266
            self.remove_oldest(remove_pinned=True)
267
268
269
        self.cache.clear()


270
class PyObjectCache:
271
    """Used to cache python objects to avoid object allocations
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
    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):
290
        """Returns a pre-allocated cached object. If there is not enough
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
        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


308
309
310
311
def is_hip() -> bool:
    return torch.version.hip is not None


312
313
@lru_cache(maxsize=None)
def is_cpu() -> bool:
314
315
316
317
318
    from importlib.metadata import PackageNotFoundError, version
    try:
        return "cpu" in version("vllm")
    except PackageNotFoundError:
        return False
319
320


321
322
323
324
325
326
327
328
329
@lru_cache(maxsize=None)
def is_openvino() -> bool:
    from importlib.metadata import PackageNotFoundError, version
    try:
        return "openvino" in version("vllm")
    except PackageNotFoundError:
        return False


330
@lru_cache(maxsize=None)
331
332
333
334
335
336
337
338
def is_neuron() -> bool:
    try:
        import transformers_neuronx
    except ImportError:
        transformers_neuronx = None
    return transformers_neuronx is not None


339
340
@lru_cache(maxsize=None)
def is_xpu() -> bool:
341
342
343
344
345
    from importlib.metadata import PackageNotFoundError, version
    try:
        is_xpu_flag = "xpu" in version("vllm")
    except PackageNotFoundError:
        return False
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
    # vllm is not build with xpu
    if not is_xpu_flag:
        return False
    try:
        import intel_extension_for_pytorch as ipex  # noqa: F401
        _import_ipex = True
    except ImportError as e:
        logger.warning("Import Error for IPEX: %s", e.msg)
        _import_ipex = False
    # ipex dependency is not ready
    if not _import_ipex:
        logger.warning("not found ipex lib")
        return False
    return hasattr(torch, "xpu") and torch.xpu.is_available()


362
@lru_cache(maxsize=None)
363
364
def get_max_shared_memory_bytes(gpu: int = 0) -> int:
    """Returns the maximum shared memory per thread block in bytes."""
365
    from vllm import _custom_ops as ops
366
    max_shared_mem = (
367
        ops.get_max_shared_memory_per_block_device_attribute(gpu))
368
369
    # value 0 will cause MAX_SEQ_LEN become negative and test_attention.py
    # will fail
370
    assert max_shared_mem > 0, "max_shared_mem can not be zero"
371
372
373
    return int(max_shared_mem)


374
def get_cpu_memory() -> int:
375
    """Returns the total CPU memory of the node in bytes."""
376
    return psutil.virtual_memory().total
Zhuohan Li's avatar
Zhuohan Li committed
377
378


379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
def seed_everything(seed: int) -> None:
    """
    Set the seed of each random module.

    Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
    """
    random.seed(seed)
    np.random.seed(seed)

    if current_platform.is_cuda_alike():
        torch.cuda.manual_seed_all(seed)

    if is_xpu():
        torch.xpu.manual_seed_all(seed)


Zhuohan Li's avatar
Zhuohan Li committed
395
396
def random_uuid() -> str:
    return str(uuid.uuid4().hex)
397

398

399
@lru_cache(maxsize=None)
400
def get_vllm_instance_id() -> str:
401
402
403
404
405
406
    """
    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.
    """
407
    return envs.VLLM_INSTANCE_ID or f"vllm-instance-{random_uuid()}"
408
409


410
@lru_cache(maxsize=None)
411
412
413
def in_wsl() -> bool:
    # Reference: https://github.com/microsoft/WSL/issues/4071
    return "microsoft" in " ".join(uname()).lower()
414
415


416
def make_async(func: Callable[P, T]) -> Callable[P, Awaitable[T]]:
417
418
419
420
421
422
423
    """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.
    """

424
    def _async_wrapper(*args: P.args, **kwargs: P.kwargs) -> asyncio.Future:
425
426
427
428
429
430
431
        loop = asyncio.get_event_loop()
        p_func = partial(func, *args, **kwargs)
        return loop.run_in_executor(executor=None, func=p_func)

    return _async_wrapper


432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
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.
    """

    # Can use anext() in python >= 3.10
    awaits = [ensure_future(iterator.__anext__())]
    while True:
        done, pending = await asyncio.wait(awaits, timeout=1)
        if await is_cancelled():
            with contextlib.suppress(BaseException):
                awaits[0].cancel()
                await iterator.aclose()
            raise asyncio.CancelledError("client cancelled")
        if done:
            try:
                item = await awaits[0]
                awaits[0] = ensure_future(iterator.__anext__())
                yield item
            except StopAsyncIteration:
                # we are done
                return
457
458


459
460
async def merge_async_iterators(
    *iterators: AsyncGenerator[T, None],
461
    is_cancelled: Optional[Callable[[], Awaitable[bool]]] = None,
462
) -> AsyncGenerator[Tuple[int, T], None]:
463
464
465
466
467
    """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.
468

469
470
    It also optionally polls a provided function at least once per second
    to check for client cancellation.
471
    """
472
473
474
475
476
477

    # Can use anext() in python >= 3.10
    awaits = {
        ensure_future(pair[1].__anext__()): pair
        for pair in enumerate(iterators)
    }
478
    timeout = None if is_cancelled is None else 1
479
480
481
482
    try:
        while awaits:
            done, pending = await asyncio.wait(awaits.keys(),
                                               return_when=FIRST_COMPLETED,
483
484
                                               timeout=timeout)
            if is_cancelled is not None and await is_cancelled():
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
                raise asyncio.CancelledError("client cancelled")
            for d in done:
                pair = awaits.pop(d)
                try:
                    item = await d
                    i, it = pair
                    awaits[ensure_future(it.__anext__())] = pair
                    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()
501
502


503
def get_ip() -> str:
504
    host_ip = envs.VLLM_HOST_IP
505
506
507
508
509
    if host_ip:
        return host_ip

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

510
    # try ipv4
511
    s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
512
    try:
513
        s.connect(("8.8.8.8", 80))  # Doesn't need to be reachable
514
        return s.getsockname()[0]
515
516
517
518
519
    except Exception:
        pass

    # try ipv6
    try:
520
        s = socket.socket(socket.AF_INET6, socket.SOCK_DGRAM)
521
522
523
        # 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
524
        return s.getsockname()[0]
525
526
527
528
529
    except Exception:
        pass

    warnings.warn(
        "Failed to get the IP address, using 0.0.0.0 by default."
530
531
        "The value can be set by the environment variable"
        " VLLM_HOST_IP or HOST_IP.",
532
533
        stacklevel=2)
    return "0.0.0.0"
534
535


536
def get_distributed_init_method(ip: str, port: int) -> str:
537
538
539
    # 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}"
540
541


542
543
544
545
546
547
548
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
549
    if port is not None:
550
551
552
553
554
555
556
557
558
        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)
559
560
561
562
563
564
565
566
567
568
    # 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]
569
570


571
572
573
574
575
576
577
578
579
580
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


581
582
def update_environment_variables(envs: Dict[str, str]):
    for k, v in envs.items():
583
        if k in os.environ and os.environ[k] != v:
584
585
586
            logger.warning(
                "Overwriting environment variable %s "
                "from '%s' to '%s'", k, os.environ[k], v)
587
        os.environ[k] = v
588
589


590
def chunk_list(lst: List[T], chunk_size: int):
591
    """Yield successive chunk_size chunks from lst."""
592
593
    for i in range(0, len(lst), chunk_size):
        yield lst[i:i + chunk_size]
594
595
596
597
598
599
600


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


601
def _generate_random_fp8(
602
    tensor: torch.Tensor,
603
604
605
606
607
608
    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.
609
    # For example, s.11111.00 in fp8e5m2 format represents Inf.
610
611
612
613
    #     | E4M3        | E5M2
    #-----|-------------|-------------------
    # Inf | N/A         | s.11111.00
    # NaN | s.1111.111  | s.11111.{01,10,11}
614
    from vllm import _custom_ops as ops
615
616
    tensor_tmp = torch.empty_like(tensor, dtype=torch.float16)
    tensor_tmp.uniform_(low, high)
617
    ops.convert_fp8(tensor, tensor_tmp)
618
619
620
    del tensor_tmp


621
622
623
def get_kv_cache_torch_dtype(
        cache_dtype: Optional[Union[str, torch.dtype]],
        model_dtype: Optional[Union[str, torch.dtype]] = None) -> torch.dtype:
624
625
626
627
628
629
630
631
632
633
    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]
634
        elif cache_dtype == "fp8":
635
636
637
638
639
640
641
            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}")
642
643
644
645
646
647
648
649
650
651
652
653
654
655
    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]]:
656
    seed_everything(seed)
657
658
659
660

    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
661
662
663
664

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

665
666
667
668
    for _ in range(num_layers):
        key_value_cache = torch.empty(size=key_value_cache_shape,
                                      dtype=torch_dtype,
                                      device=device)
669
670
671
672
673
674
675
        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}")
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
        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
692
693
694
695
696
697

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

698
    seed_everything(seed)
699
700

    torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
701
702
703
704

    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)
705
    key_caches: List[torch.Tensor] = []
706
707
708
709
    for _ in range(num_layers):
        key_cache = torch.empty(size=key_cache_shape,
                                dtype=torch_dtype,
                                device=device)
710
        if cache_dtype in ["auto", "half", "bfloat16", "float"]:
711
            key_cache.uniform_(-scale, scale)
712
713
        elif cache_dtype == 'fp8':
            _generate_random_fp8(key_cache, -scale, scale)
714
715
716
        else:
            raise ValueError(
                f"Does not support key cache of type {cache_dtype}")
717
718
719
        key_caches.append(key_cache)

    value_cache_shape = (num_blocks, num_heads, head_size, block_size)
720
    value_caches: List[torch.Tensor] = []
721
722
723
724
    for _ in range(num_layers):
        value_cache = torch.empty(size=value_cache_shape,
                                  dtype=torch_dtype,
                                  device=device)
725
        if cache_dtype in ["auto", "half", "bfloat16", "float"]:
726
            value_cache.uniform_(-scale, scale)
727
728
        elif cache_dtype == 'fp8':
            _generate_random_fp8(value_cache, -scale, scale)
729
730
731
        else:
            raise ValueError(
                f"Does not support value cache of type {cache_dtype}")
732
733
        value_caches.append(value_cache)
    return key_caches, value_caches
734
735


736
737
738
739
740
741
742
743
744
745
746
747
748
749
@lru_cache
def print_warning_once(msg: str) -> None:
    logger.warning(msg)


@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
750
751
752
    elif is_xpu():
        print_warning_once("Pin memory is not supported on XPU.")
        return False
753
754
755
    elif is_neuron():
        print_warning_once("Pin memory is not supported on Neuron.")
        return False
756
    elif is_cpu() or is_openvino():
757
        return False
758
759
760
    return True


761
class DeviceMemoryProfiler:
762

763
    def __init__(self, device: Optional[torch.types.Device] = None):
764
765
766
767
        self.device = device

    def current_memory_usage(self) -> float:
        # Return the memory usage in bytes.
768
        if current_platform.is_cuda_alike():
769
770
771
            torch.cuda.reset_peak_memory_stats(self.device)
            mem = torch.cuda.max_memory_allocated(self.device)
        elif is_xpu():
772
773
            torch.xpu.reset_peak_memory_stats(self.device)  # type: ignore
            mem = torch.xpu.max_memory_allocated(self.device)  # type: ignore
774
775
776
777
778
779
780
781
782
783
784
785
786
        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()
787
788


789
790
791
792
793
794
795
796
797
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.
798
799
800
801

    The padding is applied to the end of each inner list until it reaches
    `max_len`.
    """
802
803
804
805
806
    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)
807
808
809
    for ind, blocktb in enumerate(x):
        assert len(blocktb) <= max_len
        padded_x[ind, :len(blocktb)] = blocktb
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836

    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
837
838
839
840
841
842
843
844
845
846
847
848
849


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)


850
851
852
853
854
def get_dtype_size(dtype: torch.dtype) -> int:
    """Get the size of the data type in bytes."""
    return torch.tensor([], dtype=dtype).element_size()


855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
# `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)


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
911
912
913
914
915
916
917
918
919
920
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)


921
922
923
924
925
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]


926
def init_cached_hf_modules() -> None:
927
928
929
930
931
    """
    Lazy initialization of the Hugging Face modules.
    """
    from transformers.dynamic_module_utils import init_hf_modules
    init_hf_modules()
932
933


934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
@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
950
    env_ld_library_path = envs.LD_LIBRARY_PATH
951
952
953
954
955
956
957
958
959
960
961
    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]


962
def find_nccl_library() -> str:
963
964
965
966
967
968
    """
    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.
    """
969
    so_file = envs.VLLM_NCCL_SO_PATH
970
971
972
973

    # manually load the nccl library
    if so_file:
        logger.info(
974
975
            "Found nccl from environment variable VLLM_NCCL_SO_PATH=%s",
            so_file)
976
977
    else:
        if torch.version.cuda is not None:
978
            so_file = "libnccl.so.2"
979
        elif torch.version.hip is not None:
980
            so_file = "librccl.so.1"
981
982
        else:
            raise ValueError("NCCL only supports CUDA and ROCm backends.")
983
        logger.info("Found nccl from library %s", so_file)
984
    return so_file
985
986
987
988
989
990
991


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
    """

992
    if envs.VLLM_TRACE_FUNCTION:
993
994
995
996
997
998
999
1000
        tmp_dir = tempfile.gettempdir()
        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)
1001
1002


1003
# `functools` helpers
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
def identity(value: T) -> T:
    return value


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


def deprecate_kwargs(
        *kws: str,
        is_deprecated: Union[bool, Callable[[], bool]] = True,
        additional_message: Optional[str] = None) -> Callable[[F], F]:
    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
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059


@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
1060
1061
1062
1063
1064
1065
1066
1067
    if is_hip():
        # 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
1068
1069
1070
1071
1072
1073
    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.
1074

1075
1076
1077
1078
1079
1080
1081
    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)
1082
1083


1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
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


1098
#From: https://stackoverflow.com/a/4104188/2749989
1099
def run_once(f: Callable[P, None]) -> Callable[P, None]:
1100

1101
    def wrapper(*args: P.args, **kwargs: P.kwargs) -> None:
1102
1103
1104
1105
1106
1107
        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
1108
1109
1110
1111
1112
1113
1114
1115
1116


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

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

1117
1118
1119
        if '--config' in args:
            args = FlexibleArgumentParser._pull_args_from_config(args)

1120
1121
1122
1123
        # Convert underscores to dashes and vice versa in argument names
        processed_args = []
        for arg in args:
            if arg.startswith('--'):
1124
1125
1126
1127
1128
1129
1130
                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('_', '-'))
1131
1132
1133
1134
            else:
                processed_args.append(arg)

        return super().parse_args(processed_args, namespace)
1135

1136
1137
1138
1139
    @staticmethod
    def _pull_args_from_config(args: List[str]) -> List[str]:
        """Method to pull arguments specified in the config file
        into the command-line args variable.
1140
1141

        The arguments in config file will be inserted between
1142
        the argument list.
1143

1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
        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",
1154
1155
            "facebook/opt-12B",
            '--config', 'config.yaml',
1156
1157
1158
1159
            '-tp', '2'
        ]
        $: args = [
            "serve,chat,complete",
1160
1161
1162
            "facebook/opt-12B",
            '--port', '12323',
            '--tensor-parallel-size', '4',
1163
1164
1165
1166
1167
            '-tp', '2'
            ]
        ```

        Please note how the config args are inserted after the sub command.
1168
        this way the order of priorities is maintained when these are args
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
        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]

        config_args = FlexibleArgumentParser._load_config_file(file_path)

        # 0th index is for {serve,chat,complete}
        # followed by config args
        # followed by rest of cli args.
        # maintaining this order will enforce the precedence
        # of cli > config > defaults
        args = [args[0]] + config_args + args[1:index] + args[index + 2:]

        return args

    @staticmethod
    def _load_config_file(file_path: str) -> List[str]:
1194
        """Loads a yaml file and returns the key value pairs as a
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
        flattened list with argparse like pattern
        ```yaml
            port: 12323
            tensor-parallel-size: 4
        ```
        returns:
            processed_args: list[str] = [
                '--port': '12323',
                '--tensor-parallel-size': '4'
            ]
1205

1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        """

        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:
            with open(file_path, 'r') as config_file:
                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

        for key, value in config.items():
            processed_args.append('--' + key)
            processed_args.append(str(value))

        return processed_args

1233
1234
1235
1236
1237
1238

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)
1239
1240


1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
def get_allowed_kwarg_only_overrides(
    callable: Callable[..., object],
    overrides: Optional[Dict[str, Any]],
) -> 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.
        overrides: Potential overrides to be used when invoking the callable.

    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 {}

    allowed_override_names = [
        name for name, param in inspect.signature(callable).parameters.items()
        if param.kind == inspect.Parameter.KEYWORD_ONLY
    ]

    # Drop any mm_processor_kwargs provided by the user that are
    # not kwarg names accepted by the provided input processor.
    filtered_overrides = {
        kwarg_name: val
        for kwarg_name, val in overrides.items()
        if kwarg_name in allowed_override_names
    }

    # 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


1288
1289
1290
1291
1292
1293
# 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")
1294
1295


1296
1297
1298
1299
1300
1301
# 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")


1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
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