Repository navigation
Expand file tree
/
Copy pathutils.py
More file actions
1998 lines (1597 loc) · 65.3 KB
/
Copy pathutils.py
File metadata and controls
1998 lines (1597 loc) · 65.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
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
85
86
87
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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
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
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
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
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
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
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import argparse
import asyncio
import concurrent
import contextlib
import datetime
import enum
import gc
import getpass
import importlib.metadata
import importlib.util
import inspect
import ipaddress
import multiprocessing
import os
import re
import resource
import signal
import socket
import subprocess
import sys
import tempfile
import threading
import time
import traceback
import uuid
import warnings
import weakref
from asyncio import FIRST_COMPLETED, AbstractEventLoop, Task
from collections import OrderedDict, UserDict, defaultdict
from collections.abc import Hashable, Iterable, Mapping
from dataclasses import dataclass, field
from functools import lru_cache, partial, wraps
from typing import (TYPE_CHECKING, Any, AsyncGenerator, Awaitable, Callable,
Dict, Generator, Generic, Iterator, List, Literal,
NamedTuple, Optional, Tuple, Type, TypeVar, Union,
overload)
from uuid import uuid4
import numpy as np
import numpy.typing as npt
import psutil
import torch
import torch.types
import yaml
import zmq
import zmq.asyncio
from packaging.version import Version
from torch.library import Library
from typing_extensions import ParamSpec, TypeIs, assert_never
import vllm.envs as envs
from vllm.logger import enable_trace_function_call, init_logger
if TYPE_CHECKING:
from vllm.config import VllmConfig
logger = init_logger(__name__)
# Exception strings for non-implemented encoder/decoder scenarios
# Reminder: Please update docs/source/usage/compatibility_matrix.md
# If the feature combo become valid
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 and Flash-Attention are the only "
"backends currently supported with encoder/"
"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"
GB_bytes = 1_000_000_000
"""The number of bytes in one gigabyte (GB)."""
GiB_bytes = 1 << 30
"""The number of bytes in one gibibyte (GiB)."""
STR_DTYPE_TO_TORCH_DTYPE = {
"half": torch.half,
"bfloat16": torch.bfloat16,
"float": torch.float,
"fp8": torch.uint8,
"fp8_e4m3": torch.uint8,
"fp8_e5m2": torch.uint8,
}
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,
}
P = ParamSpec('P')
T = TypeVar("T")
U = TypeVar("U")
_K = TypeVar("_K", bound=Hashable)
_V = TypeVar("_V")
class _Sentinel:
...
ALL_PINNED_SENTINEL = _Sentinel()
class Device(enum.Enum):
GPU = enum.auto()
CPU = enum.auto()
class LayerBlockType(enum.Enum):
attention = "attention"
mamba = "mamba"
class Counter:
def __init__(self, start: int = 0) -> None:
self.counter = start
def __next__(self) -> int:
i = self.counter
self.counter += 1
return i
def reset(self) -> None:
self.counter = 0
class CacheInfo(NamedTuple):
hits: int
total: int
@property
def hit_ratio(self) -> float:
if self.total == 0:
return 0
return self.hits / self.total
class LRUCache(Generic[_K, _V]):
"""Note: This class is not thread safe!"""
def __init__(self, capacity: int) -> None:
self.cache = OrderedDict[_K, _V]()
self.pinned_items = set[_K]()
self.capacity = capacity
self._hits = 0
self._total = 0
def __contains__(self, key: _K) -> bool:
return key in self.cache
def __len__(self) -> int:
return len(self.cache)
def __getitem__(self, key: _K) -> _V:
value = self.cache[key] # Raise KeyError if not exists
self.cache.move_to_end(key)
return value
def __setitem__(self, key: _K, value: _V) -> None:
self.put(key, value)
def __delitem__(self, key: _K) -> None:
self.pop(key)
def stat(self) -> CacheInfo:
return CacheInfo(hits=self._hits, total=self._total)
def touch(self, key: _K) -> None:
self.cache.move_to_end(key)
def get(self, key: _K, default: Optional[_V] = None) -> Optional[_V]:
value: Optional[_V]
if key in self.cache:
value = self.cache[key]
self.cache.move_to_end(key)
self._hits += 1
else:
value = default
self._total += 1
return value
def put(self, key: _K, value: _V) -> None:
self.cache[key] = value
self.cache.move_to_end(key)
self._remove_old_if_needed()
def pin(self, key: _K) -> 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: _K) -> None:
self.pinned_items.remove(key)
def _on_remove(self, key: _K, value: Optional[_V]) -> None:
pass
def remove_oldest(self, *, remove_pinned: bool = False) -> None:
if not self.cache:
return
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) # type: ignore
def _remove_old_if_needed(self) -> None:
while len(self.cache) > self.capacity:
self.remove_oldest()
def pop(self, key: _K, default: Optional[_V] = None) -> Optional[_V]:
run_on_remove = key in self.cache
value = self.cache.pop(key, default)
# remove from pinned items
if key in self.pinned_items:
self._unpin(key)
if run_on_remove:
self._on_remove(key, value)
return value
def clear(self) -> None:
while len(self.cache) > 0:
self.remove_oldest(remove_pinned=True)
self.cache.clear()
class PyObjectCache:
"""Used to cache python objects to avoid object allocations
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):
"""Returns a pre-allocated cached object. If there is not enough
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
@lru_cache(maxsize=None)
def get_max_shared_memory_bytes(gpu: int = 0) -> int:
"""Returns the maximum shared memory per thread block in bytes."""
from vllm import _custom_ops as ops
max_shared_mem = (
ops.get_max_shared_memory_per_block_device_attribute(gpu))
# value 0 will cause MAX_SEQ_LEN become negative and test_attention.py
# will fail
assert max_shared_mem > 0, "max_shared_mem can not be zero"
return int(max_shared_mem)
def get_cpu_memory() -> int:
"""Returns the total CPU memory of the node in bytes."""
return psutil.virtual_memory().total
def random_uuid() -> str:
return str(uuid.uuid4().hex)
def make_async(
func: Callable[P, T],
executor: Optional[concurrent.futures.Executor] = None
) -> Callable[P, Awaitable[T]]:
"""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.
"""
def _async_wrapper(*args: P.args, **kwargs: P.kwargs) -> asyncio.Future:
loop = asyncio.get_event_loop()
p_func = partial(func, *args, **kwargs)
return loop.run_in_executor(executor=executor, func=p_func)
return _async_wrapper
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]
async def merge_async_iterators(
*iterators: AsyncGenerator[T,
None], ) -> AsyncGenerator[Tuple[int, T], None]:
"""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.
"""
loop = asyncio.get_running_loop()
awaits = {_next_task(pair[1], loop): pair for pair in enumerate(iterators)}
try:
while awaits:
done, _ = await asyncio.wait(awaits.keys(),
return_when=FIRST_COMPLETED)
for d in done:
pair = awaits.pop(d)
try:
item = await d
i, it = pair
awaits[_next_task(it, loop)] = 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()
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
def get_ip() -> str:
host_ip = envs.VLLM_HOST_IP
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.")
if host_ip:
return host_ip
# IP is not set, try to get it from the network interface
# try ipv4
s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
try:
s.connect(("8.8.8.8", 80)) # Doesn't need to be reachable
return s.getsockname()[0]
except Exception:
pass
# try ipv6
try:
s = socket.socket(socket.AF_INET6, socket.SOCK_DGRAM)
# 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
return s.getsockname()[0]
except Exception:
pass
warnings.warn(
"Failed to get the IP address, using 0.0.0.0 by default."
"The value can be set by the environment variable"
" VLLM_HOST_IP or HOST_IP.",
stacklevel=2)
return "0.0.0.0"
def is_valid_ipv6_address(address: str) -> bool:
try:
ipaddress.IPv6Address(address)
return True
except ValueError:
return False
def get_distributed_init_method(ip: str, port: int) -> str:
# 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}"
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
if port is not None:
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)
# 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]
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
def update_environment_variables(envs: Dict[str, str]):
for k, v in envs.items():
if k in os.environ and os.environ[k] != v:
logger.warning(
"Overwriting environment variable %s "
"from '%s' to '%s'", k, os.environ[k], v)
os.environ[k] = v
def chunk_list(lst: List[T], chunk_size: int):
"""Yield successive chunk_size chunks from lst."""
for i in range(0, len(lst), chunk_size):
yield lst[i:i + chunk_size]
def cdiv(a: int, b: int) -> int:
"""Ceiling division."""
return -(a // -b)
def _generate_random_fp8(
tensor: torch.Tensor,
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.
# For example, s.11111.00 in fp8e5m2 format represents Inf.
# | E4M3 | E5M2
#-----|-------------|-------------------
# Inf | N/A | s.11111.00
# NaN | s.1111.111 | s.11111.{01,10,11}
from vllm import _custom_ops as ops
tensor_tmp = torch.empty_like(tensor, dtype=torch.float16)
tensor_tmp.uniform_(low, high)
ops.convert_fp8(tensor, tensor_tmp)
del tensor_tmp
def get_kv_cache_torch_dtype(
cache_dtype: Optional[Union[str, torch.dtype]],
model_dtype: Optional[Union[str, torch.dtype]] = None) -> torch.dtype:
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]
elif cache_dtype == "fp8":
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}")
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]]:
from vllm.platforms import current_platform
current_platform.seed_everything(seed)
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
key_caches: List[torch.Tensor] = []
value_caches: List[torch.Tensor] = []
for _ in range(num_layers):
key_value_cache = torch.empty(size=key_value_cache_shape,
dtype=torch_dtype,
device=device)
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}")
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]]:
if cache_dtype == "fp8" and head_size % 16:
raise ValueError(
f"Does not support key cache of type fp8 with head_size {head_size}"
)
from vllm.platforms import current_platform
current_platform.seed_everything(seed)
torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
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)
key_caches: List[torch.Tensor] = []
for _ in range(num_layers):
key_cache = torch.empty(size=key_cache_shape,
dtype=torch_dtype,
device=device)
if cache_dtype in ["auto", "half", "bfloat16", "float"]:
key_cache.uniform_(-scale, scale)
elif cache_dtype == 'fp8':
_generate_random_fp8(key_cache, -scale, scale)
else:
raise ValueError(
f"Does not support key cache of type {cache_dtype}")
key_caches.append(key_cache)
value_cache_shape = (num_blocks, num_heads, head_size, block_size)
value_caches: List[torch.Tensor] = []
for _ in range(num_layers):
value_cache = torch.empty(size=value_cache_shape,
dtype=torch_dtype,
device=device)
if cache_dtype in ["auto", "half", "bfloat16", "float"]:
value_cache.uniform_(-scale, scale)
elif cache_dtype == 'fp8':
_generate_random_fp8(value_cache, -scale, scale)
else:
raise ValueError(
f"Does not support value cache of type {cache_dtype}")
value_caches.append(value_cache)
return key_caches, value_caches
@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)
@lru_cache
def print_warning_once(msg: str) -> None:
# Set the stacklevel to 2 to print the caller's line info
logger.warning(msg, stacklevel=2)
@lru_cache(maxsize=None)
def is_pin_memory_available() -> bool:
from vllm.platforms import current_platform
return current_platform.is_pin_memory_available()
class DeviceMemoryProfiler:
def __init__(self, device: Optional[torch.types.Device] = None):
self.device = device
def current_memory_usage(self) -> float:
# Return the memory usage in bytes.
from vllm.platforms import current_platform
if current_platform.is_cuda_alike():
torch.cuda.reset_peak_memory_stats(self.device)
mem = torch.cuda.max_memory_allocated(self.device)
elif current_platform.is_xpu():
torch.xpu.reset_peak_memory_stats(self.device) # type: ignore
mem = torch.xpu.max_memory_allocated(self.device) # type: ignore
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()
def make_ndarray_with_pad_3d(
x: List[List[List[T]]],
pad: T,
dtype: np.dtype,
*,
num_layers: int, # Fixed second dimension (required)
max_len: int,
) -> np.ndarray:
"""Convert 3D block tables to padded array with:
- First dim: sequences
- Second dim: fixed layers (pad if needed)
- Third dim: padded blocks per layer
"""
num_seqs = len(x)
# Create padded array: (num_sequences, num_layers, max_blocks)
padded_x = np.full((num_seqs, num_layers, max_len), pad, dtype=dtype)
# Inside make_ndarray_with_pad_3d
for i, seq in enumerate(x):
assert len(seq) <= num_layers # Checks middle dimension
for j, layer in enumerate(seq):
assert len(layer) <= max_len # Checks inner-most dimension
padded_x[i, j, :len(layer)] = layer
return padded_x
def make_tensor_with_pad_3d(
x: List[List[List[T]]],
pad: T,
dtype: npt.DTypeLike,
*,
max_len: Optional[int] = None,
device: Optional[Union[str, torch.device]] = None,
pin_memory: bool = False,
num_layers: Optional[int] = None,
) -> torch.Tensor:
np_dtype = TORCH_DTYPE_TO_NUMPY_DTYPE[dtype]
if num_layers is None:
num_layers = max(map(len, x), default=0)
if max_len is None:
max_len = max(map(len, (layer for seq in x for layer in seq)), default=0) if x else 0
padded_x = make_ndarray_with_pad_3d(x, pad, dtype=np_dtype, num_layers=num_layers, max_len=max_len)
tensor = torch.from_numpy(padded_x).to(device)
if pin_memory:
tensor = tensor.pin_memory()
return tensor
return padded_x
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.
The padding is applied to the end of each inner list until it reaches
`max_len`.
"""
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)
for ind, blocktb in enumerate(x):
assert len(blocktb) <= max_len
padded_x[ind, :len(blocktb)] = blocktb
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
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)
def get_dtype_size(dtype: torch.dtype) -> int:
"""Get the size of the data type in bytes."""
return torch.tensor([], dtype=dtype).element_size()
# `collections` helpers
def is_list_of(
value: object,
typ: Union[type[T], tuple[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)
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)
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]
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()
# TODO: This function can be removed if transformer_modules classes are
# serialized by value when communicating between processes
def init_cached_hf_modules() -> None:
"""
Lazy initialization of the Hugging Face modules.
"""
from transformers.dynamic_module_utils import init_hf_modules
init_hf_modules()
@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
env_ld_library_path = envs.LD_LIBRARY_PATH
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]
def find_nccl_library() -> str:
"""
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.
"""
so_file = envs.VLLM_NCCL_SO_PATH
# manually load the nccl library
if so_file:
logger.info(
"Found nccl from environment variable VLLM_NCCL_SO_PATH=%s",
so_file)
else:
if torch.version.cuda is not None:
so_file = "libnccl.so.2"
elif torch.version.hip is not None:
so_file = "librccl.so.1"
else:
raise ValueError("NCCL only supports CUDA and ROCm backends.")
logger.info("Found nccl from library %s", so_file)
return so_file