utils.py 90 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
14
"""Common utilities."""
15

16
17
from __future__ import annotations

18
import builtins
19
import ctypes
20
import dataclasses
21
import functools
22
import importlib
23
import io
24
import ipaddress
25
import itertools
26
import json
27
import logging
Lianmin Zheng's avatar
Lianmin Zheng committed
28
import os
29
import pickle
JieXin Liang's avatar
JieXin Liang committed
30
import platform
Lianmin Zheng's avatar
Lianmin Zheng committed
31
import random
Lianmin Zheng's avatar
Lianmin Zheng committed
32
import re
33
import resource
34
35
import shutil
import signal
Lianmin Zheng's avatar
Lianmin Zheng committed
36
import socket
37
import subprocess
38
import sys
Lianmin Zheng's avatar
Lianmin Zheng committed
39
import tempfile
40
import threading
Lianmin Zheng's avatar
Lianmin Zheng committed
41
import time
42
import traceback
43
import warnings
44
from collections import OrderedDict, defaultdict
45
from contextlib import contextmanager
46
from enum import Enum
47
from functools import lru_cache
48
from importlib.metadata import PackageNotFoundError, version
49
from importlib.util import find_spec
Lianmin Zheng's avatar
Lianmin Zheng committed
50
from io import BytesIO
51
from json import JSONDecodeError
52
from multiprocessing.reduction import ForkingPickler
53
from pathlib import Path
54
55
56
57
58
59
60
61
62
63
64
65
66
from typing import (
    Any,
    Callable,
    Dict,
    Generic,
    List,
    Optional,
    Protocol,
    Set,
    Tuple,
    TypeVar,
    Union,
)
Lianmin Zheng's avatar
Lianmin Zheng committed
67
68

import numpy as np
69
import psutil
70
import pybase64
Lianmin Zheng's avatar
Lianmin Zheng committed
71
72
import requests
import torch
73
import torch.distributed
74
import torch.distributed as dist
75
import triton
76
import zmq
77
from fastapi.responses import ORJSONResponse
78
from packaging import version as pkg_version
Mick's avatar
Mick committed
79
from PIL import Image
Lianmin Zheng's avatar
Lianmin Zheng committed
80
from starlette.routing import Mount
81
from torch import nn
82
from torch.func import functional_call
83
from torch.library import Library
84
from torch.profiler import ProfilerActivity, profile, record_function
85
from torch.utils._contextlib import _DecoratorContextManager
86
from triton.runtime.cache import FileCacheManager
87

88
89
logger = logging.getLogger(__name__)

Liangsheng Yin's avatar
Liangsheng Yin committed
90
91
show_time_cost = False
time_infos = {}
Lianmin Zheng's avatar
Lianmin Zheng committed
92

93
94
HIP_FP8_E4M3_FNUZ_MAX = 224.0

95

96
# https://pytorch.org/docs/stable/notes/hip.html#checking-for-hip
97
98
99
100
def is_hip() -> bool:
    return torch.version.hip is not None


101
102
103
104
105
106
107
108
109
110
111
if is_hip():
    FP8_E4M3_MAX = HIP_FP8_E4M3_FNUZ_MAX
else:
    FP8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max

FP8_E4M3_MIN = -FP8_E4M3_MAX

builtins.FP8_E4M3_MAX = FP8_E4M3_MAX
builtins.FP8_E4M3_MIN = FP8_E4M3_MIN


112
def is_cuda():
113
    return torch.cuda.is_available() and torch.version.cuda
114
115
116
117
118
119
120
121
122
123
124
125
126
127


def is_cuda_alike():
    return is_cuda() or is_hip()


def is_hpu() -> bool:
    return hasattr(torch, "hpu") and torch.hpu.is_available()


def is_xpu() -> bool:
    return hasattr(torch, "xpu") and torch.xpu.is_available()


128
129
130
131
def is_npu() -> bool:
    return hasattr(torch, "npu") and torch.npu.is_available()


132
def is_host_cpu_x86() -> bool:
JieXin Liang's avatar
JieXin Liang committed
133
134
135
136
137
138
139
140
    machine = platform.machine().lower()
    return (
        machine in ("x86_64", "amd64", "i386", "i686")
        and hasattr(torch, "cpu")
        and torch.cpu.is_available()
    )


141
142
143
144
def is_cpu() -> bool:
    return os.getenv("SGLANG_USE_CPU_ENGINE", "0") == "1" and is_host_cpu_x86()


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
def get_cuda_version():
    if torch.version.cuda:
        return tuple(map(int, torch.version.cuda.split(".")))
    return (0, 0)


def _check(cc_major):
    if not is_cuda():
        return False
    return torch.cuda.get_device_capability()[0] == cc_major and tuple(
        map(int, torch.version.cuda.split(".")[:2])
    ) >= (12, 3)


is_ampere_with_cuda_12_3 = lambda: _check(8)
is_hopper_with_cuda_12_3 = lambda: _check(9)


def is_blackwell():
    if not is_cuda():
        return False
    return torch.cuda.get_device_capability()[0] == 10


_warned_bool_env_var_keys = set()


def get_bool_env_var(name: str, default: str = "false") -> bool:
    value = os.getenv(name, default)
    value = value.lower()

    truthy_values = ("true", "1")
    falsy_values = ("false", "0")

    if (value not in truthy_values) and (value not in falsy_values):
        if value not in _warned_bool_env_var_keys:
            logger.warning(
                f"get_bool_env_var({name}) see non-understandable value={value} and treat as false"
            )
        _warned_bool_env_var_keys.add(value)

    return value in truthy_values


def get_int_env_var(name: str, default: int = 0) -> int:
    value = os.getenv(name)
    if value is None or not value.strip():
        return default
    try:
        return int(value)
    except ValueError:
        return default


def support_triton(backend: str) -> bool:
200
    return backend not in ["torch_native", "intel_amx", "ascend"]
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220


try:
    import sgl_kernel

    is_intel_amx_backend_available = hasattr(
        torch.ops.sgl_kernel, "convert_weight_packed"
    )
except:
    is_intel_amx_backend_available = False


def cpu_has_amx_support():
    return torch._C._cpu._is_amx_tile_supported() and is_intel_amx_backend_available


def use_intel_amx_backend(layer):
    return getattr(layer, "use_intel_amx_backend", False)


221
222
223
224
225
def is_flashinfer_available():
    """
    Check whether flashinfer is available.
    As of Oct. 6, 2024, it is only available on NVIDIA GPUs.
    """
226
    if not get_bool_env_var("SGLANG_IS_FLASHINFER_AVAILABLE", default="true"):
227
        return False
228
    return importlib.util.find_spec("flashinfer") is not None and is_cuda()
229
230


231
_ENABLE_TORCH_INFERENCE_MODE = get_bool_env_var(
232
    "SGLANG_ENABLE_TORCH_INFERENCE_MODE", "false"
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


class DynamicGradMode(_DecoratorContextManager):
    """
    A combination of torch.no_grad and torch.inference_mode,
    with their behavior controlled by an environment variable. Just refer to them.
    """

    @staticmethod
    def set_inference_mode(mode: bool):
        if isinstance(mode, bool):
            global _ENABLE_TORCH_INFERENCE_MODE

            _ENABLE_TORCH_INFERENCE_MODE = mode
        else:
            logger.warning("mode is not a boolean object")

    def __init__(self, mode=True):
        if not torch._jit_internal.is_scripting():
            super().__init__()
        if _ENABLE_TORCH_INFERENCE_MODE:
            self.mode = mode
        else:
            self.prev = False

    def __new__(cls, mode_or_orig_func=True if _ENABLE_TORCH_INFERENCE_MODE else None):
        if mode_or_orig_func is None or isinstance(mode_or_orig_func, bool):
            return super().__new__(cls)
        return cls()(mode_or_orig_func)

    def __enter__(self) -> None:
        if _ENABLE_TORCH_INFERENCE_MODE:
            self._inference_mode_context = torch._C._InferenceMode(self.mode)
            self._inference_mode_context.__enter__()
        else:
            self.prev = torch.is_grad_enabled()
            torch.set_grad_enabled(False)

    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
        if _ENABLE_TORCH_INFERENCE_MODE:
            self._inference_mode_context.__exit__(exc_type, exc_value, traceback)
        else:
            torch.set_grad_enabled(self.prev)

    def clone(self) -> "DynamicGradMode":
        r"""
        Create a copy of this class
        """
        if _ENABLE_TORCH_INFERENCE_MODE:
            return self.__class__(self.mode)
        else:
            return self.__class__()


Liangsheng Yin's avatar
Liangsheng Yin committed
288
289
290
291
def enable_show_time_cost():
    global show_time_cost
    show_time_cost = True

Lianmin Zheng's avatar
Lianmin Zheng committed
292

Liangsheng Yin's avatar
Liangsheng Yin committed
293
294
295
296
297
298
class TimeInfo:
    def __init__(self, name, interval=0.1, color=0, indent=0):
        self.name = name
        self.interval = interval
        self.color = color
        self.indent = indent
Lianmin Zheng's avatar
Lianmin Zheng committed
299

Liangsheng Yin's avatar
Liangsheng Yin committed
300
301
        self.acc_time = 0
        self.last_acc_time = 0
Lianmin Zheng's avatar
Lianmin Zheng committed
302

Liangsheng Yin's avatar
Liangsheng Yin committed
303
304
305
306
307
    def check(self):
        if self.acc_time - self.last_acc_time > self.interval:
            self.last_acc_time = self.acc_time
            return True
        return False
Lianmin Zheng's avatar
Lianmin Zheng committed
308

Liangsheng Yin's avatar
Liangsheng Yin committed
309
310
311
312
    def pretty_print(self):
        print(f"\x1b[{self.color}m", end="")
        print("-" * self.indent * 2, end="")
        print(f"{self.name}: {self.acc_time:.3f}s\x1b[0m")
Lianmin Zheng's avatar
Lianmin Zheng committed
313
314


Liangsheng Yin's avatar
Liangsheng Yin committed
315
316
317
318
def mark_start(name, interval=0.1, color=0, indent=0):
    global time_infos, show_time_cost
    if not show_time_cost:
        return
Lianmin Zheng's avatar
Lianmin Zheng committed
319
    torch.cuda.synchronize()
Liangsheng Yin's avatar
Liangsheng Yin committed
320
321
    if time_infos.get(name, None) is None:
        time_infos[name] = TimeInfo(name, interval, color, indent)
322
    time_infos[name].acc_time -= time.perf_counter()
Lianmin Zheng's avatar
Lianmin Zheng committed
323
324


Liangsheng Yin's avatar
Liangsheng Yin committed
325
326
327
328
def mark_end(name):
    global time_infos, show_time_cost
    if not show_time_cost:
        return
Lianmin Zheng's avatar
Lianmin Zheng committed
329
    torch.cuda.synchronize()
330
    time_infos[name].acc_time += time.perf_counter()
Liangsheng Yin's avatar
Liangsheng Yin committed
331
332
    if time_infos[name].check():
        time_infos[name].pretty_print()
Lianmin Zheng's avatar
Lianmin Zheng committed
333
334
335
336
337
338
339


def calculate_time(show=False, min_cost_ms=0.0):
    def wrapper(func):
        def inner_func(*args, **kwargs):
            torch.cuda.synchronize()
            if show:
340
                start_time = time.perf_counter()
Lianmin Zheng's avatar
Lianmin Zheng committed
341
342
343
            result = func(*args, **kwargs)
            torch.cuda.synchronize()
            if show:
344
                cost_time = (time.perf_counter() - start_time) * 1000
Lianmin Zheng's avatar
Lianmin Zheng committed
345
346
347
348
349
350
351
352
353
                if cost_time > min_cost_ms:
                    print(f"Function {func.__name__} took {cost_time} ms to run.")
            return result

        return inner_func

    return wrapper


354
355
356
def get_available_gpu_memory(
    device, gpu_id, distributed=False, empty_cache=True, cpu_group=None
):
Lianmin Zheng's avatar
Lianmin Zheng committed
357
358
359
360
    """
    Get available memory for cuda:gpu_id device.
    When distributed is True, the available memory is the minimum available memory of all GPUs.
    """
Zhang, Liangang's avatar
Zhang, Liangang committed
361
    if device == "cuda":
362
        num_gpus = torch.cuda.device_count()
Zhang, Liangang's avatar
Zhang, Liangang committed
363
364
365
366
367
368
369
370
        assert gpu_id < num_gpus

        if torch.cuda.current_device() != gpu_id:
            print(
                f"WARNING: current device is not {gpu_id}, but {torch.cuda.current_device()}, ",
                "which may cause useless memory allocation for torch CUDA context.",
            )

371
372
        if empty_cache:
            torch.cuda.empty_cache()
Zhang, Liangang's avatar
Zhang, Liangang committed
373
374
375
376
377
378
379
380
381
382
383
        free_gpu_memory, _ = torch.cuda.mem_get_info(gpu_id)

    elif device == "xpu":
        num_gpus = torch.xpu.device_count()
        assert gpu_id < num_gpus

        if torch.xpu.current_device() != gpu_id:
            print(
                f"WARNING: current device is not {gpu_id}, but {torch.xpu.current_device()}, ",
                "which may cause useless memory allocation for torch XPU context.",
            )
384
385
386

        if empty_cache:
            torch.xpu.empty_cache()
Zhang, Liangang's avatar
Zhang, Liangang committed
387
388
389
        used_memory = torch.xpu.memory_allocated()
        total_gpu_memory = torch.xpu.get_device_properties(gpu_id).total_memory
        free_gpu_memory = total_gpu_memory - used_memory
Lianmin Zheng's avatar
Lianmin Zheng committed
390

391
392
393
394
395
396
397
398
399
400
401
402
    elif device == "hpu":
        num_gpus = torch.hpu.device_count()
        assert gpu_id < num_gpus

        if torch.hpu.current_device() != gpu_id:
            print(
                f"WARNING: current device is not {gpu_id}, but {torch.hpu.current_device()}, ",
                "which may cause useless memory allocation for torch HPU context.",
            )

        free_gpu_memory, total_gpu_memory = torch.hpu.mem_get_info()

403
404
405
    elif device == "cpu":
        # TODO: rename the variables in the current function to be not GPU specific
        free_gpu_memory = psutil.virtual_memory().available
406
407
408
409
410
411
412
413
414
415
    elif device == "npu":
        num_gpus = torch.npu.device_count()
        assert gpu_id < num_gpus

        if torch.npu.current_device() != gpu_id:
            print(
                f"WARNING: current device is not {gpu_id}, but {torch.npu.current_device()}, ",
                "which may cause useless memory allocation for torch NPU context.",
            )
        free_gpu_memory, total_gpu_memory = torch.npu.mem_get_info()
416

Lianmin Zheng's avatar
Lianmin Zheng committed
417
    if distributed:
418
419
420
        tensor = torch.tensor(free_gpu_memory, dtype=torch.float32)
        torch.distributed.all_reduce(
            tensor, op=torch.distributed.ReduceOp.MIN, group=cpu_group
Lianmin Zheng's avatar
Lianmin Zheng committed
421
422
423
424
425
426
        )
        free_gpu_memory = tensor.item()

    return free_gpu_memory / (1 << 30)


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
def is_pin_memory_available() -> bool:
    return torch.cuda.is_available()


_CPU_OFFLOAD_BYTES = 0
_CPU_OFFLOAD_MAX_BYTES = 0


def set_cpu_offload_max_bytes(max_bytes: int) -> None:
    global _CPU_OFFLOAD_MAX_BYTES, _CPU_OFFLOAD_BYTES
    _CPU_OFFLOAD_BYTES = 0
    _CPU_OFFLOAD_MAX_BYTES = max_bytes


def maybe_offload_to_cpu(module: torch.nn.Module) -> torch.nn.Module:
    device = next(module.parameters()).device

    if device == torch.device("cpu"):
        return module

    global _CPU_OFFLOAD_MAX_BYTES, _CPU_OFFLOAD_BYTES
    if _CPU_OFFLOAD_BYTES >= _CPU_OFFLOAD_MAX_BYTES:
        return module

    pin_memory = is_pin_memory_available()
    # offload parameters to CPU
    # use pin_memory if possible, which helps cudagraph capture speed
    offloaded_parameters = False
    for p in module.parameters():
        if _CPU_OFFLOAD_BYTES >= _CPU_OFFLOAD_MAX_BYTES:
            # we use per-parameter offloading
            # one module might have some parameters offloaded and some not
            break

        # `torch.empty_like` does not support `pin_memory` argument
        cpu_data = torch.empty_strided(
            size=p.data.size(),
            stride=p.data.stride(),
            dtype=p.data.dtype,
            layout=p.data.layout,
            device="cpu",
            pin_memory=pin_memory,
        )
        cpu_data.copy_(p.data)
        p.data = cpu_data
        _CPU_OFFLOAD_BYTES += p.data.numel() * p.data.element_size()
        offloaded_parameters = True

    if offloaded_parameters:
        original_forward = module.forward

        def forward(*args, **kwargs):
            module.forward = original_forward
            device_state = {
                # here we blindly call `to(device)`
                # if the parameter is already on the device, it will be a no-op
                k: v.to(device, non_blocking=True)
                for k, v in module.state_dict().items()
            }
            output = functional_call(module, device_state, args=args, kwargs=kwargs)
            module.forward = forward
            return output

        module.forward = forward

    return module


class LayerFn(Protocol):

    def __call__(self, layer_id: int, prefix: str) -> torch.nn.Module: ...


def make_layers(
    num_hidden_layers: int,
    layer_fn: LayerFn,
503
504
    pp_rank: Optional[int] = None,
    pp_size: Optional[int] = None,
505
    prefix: str = "",
506
    return_tuple: bool = False,
507
508
) -> Tuple[int, int, torch.nn.ModuleList]:
    """Make a list of layers with the given layer function"""
509
510
511
512
513
514
515
516
517
518
519
520
521
522
    # circula imports
    from sglang.srt.distributed import get_pp_indices
    from sglang.srt.layers.utils import PPMissingLayer

    assert not pp_size or num_hidden_layers >= pp_size
    start_layer, end_layer = (
        get_pp_indices(
            num_hidden_layers,
            pp_rank,
            pp_size,
        )
        if pp_rank is not None and pp_size is not None
        else (0, num_hidden_layers)
    )
523
    modules = torch.nn.ModuleList(
524
525
        [PPMissingLayer(return_tuple=return_tuple) for _ in range(start_layer)]
        + [
526
            maybe_offload_to_cpu(layer_fn(idx=idx, prefix=add_prefix(idx, prefix)))
527
528
529
530
531
            for idx in range(start_layer, end_layer)
        ]
        + [
            PPMissingLayer(return_tuple=return_tuple)
            for _ in range(end_layer, num_hidden_layers)
532
533
        ]
    )
534
535
536
    if pp_rank is None or pp_size is None:
        return modules
    return modules, start_layer, end_layer
537
538


Lianmin Zheng's avatar
Lianmin Zheng committed
539
def set_random_seed(seed: int) -> None:
540
    """Set the random seed for all libraries."""
Lianmin Zheng's avatar
Lianmin Zheng committed
541
    random.seed(seed)
542
    np.random.seed(seed)
Lianmin Zheng's avatar
Lianmin Zheng committed
543
544
545
546
547
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)


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
def find_process_using_port(port: int) -> Optional[psutil.Process]:
    for conn in psutil.net_connections(kind="inet"):
        if conn.laddr.port == port:
            try:
                return psutil.Process(conn.pid)
            except psutil.NoSuchProcess:
                # It could happen by race condition (the proc dies when psutil.Process is called).
                pass

    return None


def wait_port_available(
    port: int, port_name: str, timeout_s: int = 30, raise_exception: bool = True
) -> bool:
    for i in range(timeout_s):
        if is_port_available(port):
            return True

        if i > 10 and i % 5 == 0:
            process = find_process_using_port(port)
            if process is None:
                logger.warning(
                    f"The port {port} is in use, but we could not find the process that uses it."
                )

            pid = process.pid
            error_message = f"{port_name} is used by a process already. {process.name()=}' {process.cmdline()=} {process.status()=} {pid=}"
            logger.info(
                f"port {port} is in use. Waiting for {i} seconds for {port_name} to be available. {error_message}"
            )
        time.sleep(0.1)

    if raise_exception:
        raise ValueError(
            f"{port_name} at {port} is not available in {timeout_s} seconds. {error_message}"
        )
    return False


588
def is_port_available(port):
589
    """Return whether a port is available."""
590
591
    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
        try:
592
            s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
593
            s.bind(("", port))
594
            s.listen(1)
595
596
597
            return True
        except socket.error:
            return False
TianYu GUO's avatar
TianYu GUO committed
598
599
        except OverflowError:
            return False
600
601


602
603
604
605
606
607
608
609
610
611
612
613
614
def get_free_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]


Yuanhan Zhang's avatar
Yuanhan Zhang committed
615
616
617
618
def decode_video_base64(video_base64):
    from PIL import Image

    # Decode the base64 string
619
    video_bytes = pybase64.b64decode(video_base64, validate=True)
Yuanhan Zhang's avatar
Yuanhan Zhang committed
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

    # Placeholder for the start indices of each PNG image
    img_starts = []

    frame_format = "PNG"  # str(os.getenv('FRAME_FORMAT', "JPEG"))

    assert frame_format in [
        "PNG",
        "JPEG",
    ], "FRAME_FORMAT must be either 'PNG' or 'JPEG'"

    if frame_format == "PNG":
        # Find each PNG start signature to isolate images
        i = 0
        while i < len(video_bytes) - 7:  # Adjusted for the length of the PNG signature
            # Check if we found the start of a PNG file
            if (
                video_bytes[i] == 0x89
                and video_bytes[i + 1] == 0x50
                and video_bytes[i + 2] == 0x4E
                and video_bytes[i + 3] == 0x47
                and video_bytes[i + 4] == 0x0D
                and video_bytes[i + 5] == 0x0A
                and video_bytes[i + 6] == 0x1A
                and video_bytes[i + 7] == 0x0A
            ):
                img_starts.append(i)
                i += 8  # Skip the PNG signature
            else:
                i += 1
    else:
        # Find each JPEG start (0xFFD8) to isolate images
        i = 0
        while (
            i < len(video_bytes) - 1
        ):  # Adjusted for the length of the JPEG SOI signature
            # Check if we found the start of a JPEG file
            if video_bytes[i] == 0xFF and video_bytes[i + 1] == 0xD8:
                img_starts.append(i)
                # Move to the next byte to continue searching for the next image start
                i += 2
            else:
                i += 1

    frames = []
    for start_idx in img_starts:
        # Assuming each image is back-to-back, the end of one image is the start of another
        # The last image goes until the end of the byte string
        end_idx = (
            img_starts[img_starts.index(start_idx) + 1]
            if img_starts.index(start_idx) + 1 < len(img_starts)
            else len(video_bytes)
        )
        img_bytes = video_bytes[start_idx:end_idx]

        # Convert bytes to a PIL Image
        img = Image.open(BytesIO(img_bytes))

        # Convert PIL Image to a NumPy array
        frame = np.array(img)

        # Append the frame to the list of frames
        frames.append(frame)

    # Ensure there's at least one frame to avoid errors with np.stack
    if frames:
        return np.stack(frames, axis=0), img.size
    else:
        return np.array([]), (
            0,
            0,
        )  # Return an empty array and size tuple if no frames were found
Lianmin Zheng's avatar
Lianmin Zheng committed
692
693


694
695
696
def load_audio(
    audio_file: str, sr: Optional[int] = None, mono: bool = True
) -> np.ndarray:
Mick's avatar
Mick committed
697
698
699
700
701
    # Use soundfile here, since librosa use it under the hood,
    # and librosa will not support audio loading in the future
    import soundfile as sf
    from scipy.signal import resample

702
703
704
    if sr is None:
        sr = 16000

Mick's avatar
Mick committed
705
706
707
708
709
    # Load audio data
    if isinstance(audio_file, bytes):
        audio, original_sr = sf.read(BytesIO(audio_file))
    elif audio_file.startswith("data:"):
        audio_file = audio_file.split(",")[1]
710
711
712
        audio, original_sr = sf.read(
            BytesIO(pybase64.b64decode(audio_file, validate=True))
        )
Mick's avatar
Mick committed
713
714
715
716
717
718
    elif audio_file.startswith("http://") or audio_file.startswith("https://"):
        timeout = int(os.getenv("REQUEST_TIMEOUT", "5"))
        response = requests.get(audio_file, stream=True, timeout=timeout)
        audio_file = BytesIO(response.content)
        response.close()
        audio, original_sr = sf.read(audio_file)
Mick's avatar
Mick committed
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
    elif isinstance(audio_file, str):
        audio, original_sr = sf.read(audio_file)
    else:
        raise ValueError(f"Invalid audio format: {audio_file}")

    # Resample audio if the original sample rate is different from the desired sample rate
    if original_sr != sr:
        num_samples = int(len(audio) * float(sr) / original_sr)
        audio = resample(audio, num_samples)

    # Convert to mono if requested and audio is stereo
    if mono and len(audio.shape) > 1:
        audio = np.mean(audio, axis=1)

    return audio

Lianmin Zheng's avatar
Lianmin Zheng committed
735

736
def load_image(
737
    image_file: Union[Image.Image, str, bytes],
738
) -> tuple[Image.Image, tuple[int, int]]:
Yuanhan Zhang's avatar
Yuanhan Zhang committed
739
    image = image_size = None
740
741
742
743
    if isinstance(image_file, Image.Image):
        image = image_file
        image_size = (image.width, image.height)
    elif isinstance(image_file, bytes):
744
745
        image = Image.open(BytesIO(image_file))
    elif image_file.startswith("http://") or image_file.startswith("https://"):
Lianmin Zheng's avatar
Lianmin Zheng committed
746
        timeout = int(os.getenv("REQUEST_TIMEOUT", "3"))
747
748
749
        response = requests.get(image_file, stream=True, timeout=timeout).raw
        image = Image.open(response)
        response.close()
Lianmin Zheng's avatar
Lianmin Zheng committed
750
751
752
    elif image_file.lower().endswith(("png", "jpg", "jpeg", "webp", "gif")):
        image = Image.open(image_file)
    elif image_file.startswith("data:"):
753
        image_file = image_file.split(",")[1]
754
        image = Image.open(BytesIO(pybase64.b64decode(image_file, validate=True)))
755
    elif isinstance(image_file, str):
756
        image = Image.open(BytesIO(pybase64.b64decode(image_file, validate=True)))
757
758
    else:
        raise ValueError(f"Invalid image: {image}")
Lianmin Zheng's avatar
Lianmin Zheng committed
759

Yuanhan Zhang's avatar
Yuanhan Zhang committed
760
    return image, image_size
761
762


763
764
def load_video(video_file: Union[str, bytes], use_gpu: bool = True):
    # We import decord here to avoid a strange Segmentation fault (core dumped) issue.
765
766
767
768
769
770
771
772
773
    from decord import VideoReader, cpu, gpu

    try:
        from decord.bridge import decord_bridge

        ctx = gpu(0)
        _ = decord_bridge.get_ctx_device(ctx)
    except Exception:
        ctx = cpu(0)
774
775
776
777
778
779
780
781

    tmp_file = None
    vr = None
    try:
        if isinstance(video_file, bytes):
            tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
            tmp_file.write(video_file)
            tmp_file.close()
782
            vr = VideoReader(tmp_file.name, ctx=ctx)
783
784
785
786
787
788
789
790
791
        elif isinstance(video_file, str):
            if video_file.startswith(("http://", "https://")):
                timeout = int(os.getenv("REQUEST_TIMEOUT", "10"))
                response = requests.get(video_file, stream=True, timeout=timeout)
                response.raise_for_status()
                tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
                for chunk in response.iter_content(chunk_size=8192):
                    tmp_file.write(chunk)
                tmp_file.close()
792
                vr = VideoReader(tmp_file.name, ctx=ctx)
793
794
795
796
797
798
            elif video_file.startswith("data:"):
                _, encoded = video_file.split(",", 1)
                video_bytes = base64.b64decode(encoded)
                tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
                tmp_file.write(video_bytes)
                tmp_file.close()
799
                vr = VideoReader(tmp_file.name, ctx=ctx)
800
            elif os.path.isfile(video_file):
801
                vr = VideoReader(video_file, ctx=ctx)
802
803
804
805
806
            else:
                video_bytes = base64.b64decode(video_file)
                tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
                tmp_file.write(video_bytes)
                tmp_file.close()
807
                vr = VideoReader(tmp_file.name, ctx=ctx)
808
809
810
811
812
813
814
815
816
817
        else:
            raise ValueError(f"Unsupported video input type: {type(video_file)}")

        return vr

    finally:
        if tmp_file and os.path.exists(tmp_file.name):
            os.unlink(tmp_file.name)


818
def suppress_other_loggers():
819
820
821
822
    warnings.filterwarnings(
        "ignore", category=UserWarning, message="The given NumPy array is not writable"
    )

Yineng Zhang's avatar
Yineng Zhang committed
823
824
825
826
    try:
        from vllm.logger import logger as vllm_default_logger
    except ImportError:
        return
827
828

    vllm_default_logger.setLevel(logging.WARN)
829
830
831
    logging.getLogger("vllm.distributed.device_communicators.pynccl").setLevel(
        logging.WARN
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
832
833
834
    logging.getLogger("vllm.distributed.device_communicators.shm_broadcast").setLevel(
        logging.WARN
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
835
    logging.getLogger("vllm.config").setLevel(logging.ERROR)
836
837


838
def assert_pkg_version(pkg: str, min_version: str, message: str):
839
840
841
842
    try:
        installed_version = version(pkg)
        if pkg_version.parse(installed_version) < pkg_version.parse(min_version):
            raise Exception(
843
                f"{pkg} is installed with version {installed_version}, which "
Ying Sheng's avatar
Ying Sheng committed
844
                f"is less than the minimum required version {min_version}. " + message
845
846
            )
    except PackageNotFoundError:
Yuanhan Zhang's avatar
Yuanhan Zhang committed
847
        raise Exception(
Ying Sheng's avatar
Ying Sheng committed
848
849
            f"{pkg} with minimum required version {min_version} is not installed. "
            + message
Yuanhan Zhang's avatar
Yuanhan Zhang committed
850
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
851
852


853
854
def kill_process_tree(parent_pid, include_parent: bool = True, skip_pid: int = None):
    """Kill the process and all its child processes."""
855
856
857
858
    # Remove sigchld handler to avoid spammy logs.
    if threading.current_thread() is threading.main_thread():
        signal.signal(signal.SIGCHLD, signal.SIG_DFL)

859
860
861
    if parent_pid is None:
        parent_pid = os.getpid()
        include_parent = False
Lianmin Zheng's avatar
Lianmin Zheng committed
862

863
    try:
864
        itself = psutil.Process(parent_pid)
865
866
867
    except psutil.NoSuchProcess:
        return

Lianmin Zheng's avatar
Lianmin Zheng committed
868
    children = itself.children(recursive=True)
869
    for child in children:
870
871
        if child.pid == skip_pid:
            continue
872
873
874
875
876
        try:
            child.kill()
        except psutil.NoSuchProcess:
            pass

877
    if include_parent:
878
        try:
Lianmin Zheng's avatar
Lianmin Zheng committed
879
880
881
882
            if parent_pid == os.getpid():
                itself.kill()
                sys.exit(0)

883
            itself.kill()
884

885
886
887
888
889
            # Sometime processes cannot be killed with SIGKILL (e.g, PID=1 launched by kubernetes),
            # so we send an additional signal to kill them.
            itself.send_signal(signal.SIGQUIT)
        except psutil.NoSuchProcess:
            pass
890
891


892
def monkey_patch_p2p_access_check():
893
    """
894
    Monkey patch the slow p2p access check.
895
896
897
    NOTE: We assume the p2p access is always allowed, which can be wrong for some setups.
    """

898
    import sglang.srt.distributed.device_communicators.custom_all_reduce_utils as tgt
Liangsheng Yin's avatar
Liangsheng Yin committed
899

900
    setattr(tgt, "gpu_p2p_access_check", lambda *arg, **kwargs: True)
901

Lianmin Zheng's avatar
Lianmin Zheng committed
902
    # Suppress the warnings from this delete function when using sglang.bench_one_batch
903
904
905
    from sglang.srt.distributed.device_communicators.custom_all_reduce import (
        CustomAllreduce,
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
906
907
908

    setattr(CustomAllreduce, "__del__", lambda *args, **kwargs: None)

909

910
def monkey_patch_vllm_gguf_config():
Yineng Zhang's avatar
Yineng Zhang committed
911
912
913
914
915
916
917
918
    try:
        from vllm.model_executor.layers.quantization.gguf import (
            GGUFConfig,
            GGUFEmbeddingMethod,
            GGUFLinearMethod,
        )
    except ImportError:
        return
919

Yineng Zhang's avatar
Yineng Zhang committed
920
    from sglang.srt.layers.linear import LinearBase
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
    from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding

    def get_quant_method_with_embedding_replaced(
        self, layer: torch.nn.Module, prefix: str
    ) -> Optional["QuantizeMethodBase"]:
        if isinstance(layer, LinearBase):
            return GGUFLinearMethod(self)
        elif isinstance(layer, VocabParallelEmbedding):
            # patch to own VocabParallelEmbedding
            return GGUFEmbeddingMethod(self)
        return None

    setattr(GGUFConfig, "get_quant_method", get_quant_method_with_embedding_replaced)


936
937
938
939
940
941
def maybe_set_triton_cache_manager() -> None:
    """Set environment variable to tell Triton to use a
    custom cache manager"""
    cache_manger = os.environ.get("TRITON_CACHE_MANAGER", None)
    if cache_manger is None:
        manager = "sglang.srt.utils:CustomCacheManager"
942
        logger.debug("Setting Triton cache manager to: %s", manager)
943
944
945
946
947
948
        os.environ["TRITON_CACHE_MANAGER"] = manager


class CustomCacheManager(FileCacheManager):
    # Adapted from: https://github.com/tdoublep/vllm/blob/3307522289fdfefe323b6c00d0db696651989a2f/vllm/triton_utils/custom_cache_manager.py
    def __init__(self, key, override=False, dump=False):
949
        from sglang.srt.distributed.parallel_state import get_tp_group
950
951
952

        self.key = key
        self.lock_path = None
953
954
955
956
957
958
959
960
961
962
963
964
965

        try:
            module_path = "triton.runtime.cache"
            cache_module = importlib.import_module(module_path)

            default_cache_dir = getattr(cache_module, "default_cache_dir", None)
            default_dump_dir = getattr(cache_module, "default_dump_dir", None)
            default_override_dir = getattr(cache_module, "default_override_dir", None)
        except (ModuleNotFoundError, AttributeError) as e:
            default_cache_dir = None
            default_dump_dir = None
            default_override_dir = None

966
        if dump:
967
968
969
970
971
            self.cache_dir = (
                default_dump_dir()
                if default_dump_dir is not None
                else os.path.join(Path.home(), ".triton", "dump")
            )
972
973
974
975
            self.cache_dir = os.path.join(self.cache_dir, self.key)
            self.lock_path = os.path.join(self.cache_dir, "lock")
            os.makedirs(self.cache_dir, exist_ok=True)
        elif override:
976
977
978
979
980
            self.cache_dir = (
                default_override_dir()
                if default_override_dir is not None
                else os.path.join(Path.home(), ".triton", "override")
            )
981
982
983
            self.cache_dir = os.path.join(self.cache_dir, self.key)
        else:
            # create cache directory if it doesn't exist
984
985
986
987
            self.cache_dir = os.getenv("TRITON_CACHE_DIR", "").strip() or (
                default_cache_dir()
                if default_cache_dir is not None
                else os.path.join(Path.home(), ".triton", "cache")
988
989
            )
            if self.cache_dir:
990
991
992
993
                try:
                    self.cache_dir = f"{self.cache_dir}_{get_tp_group().local_rank}"
                except:
                    self.cache_dir = f"{self.cache_dir}_{os.getpid()}"
994
995
996
997
998
999
1000
                self.cache_dir = os.path.join(self.cache_dir, self.key)
                self.lock_path = os.path.join(self.cache_dir, "lock")
                os.makedirs(self.cache_dir, exist_ok=True)
            else:
                raise RuntimeError("Could not create or locate cache dir")


1001
def set_ulimit(target_soft_limit=65535):
1002
    # number of open files
1003
1004
1005
1006
1007
1008
1009
    resource_type = resource.RLIMIT_NOFILE
    current_soft, current_hard = resource.getrlimit(resource_type)

    if current_soft < target_soft_limit:
        try:
            resource.setrlimit(resource_type, (target_soft_limit, current_hard))
        except ValueError as e:
Lianmin Zheng's avatar
Lianmin Zheng committed
1010
            logger.warning(f"Fail to set RLIMIT_NOFILE: {e}")
1011

1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
    # stack size
    resource_type = resource.RLIMIT_STACK
    current_soft, current_hard = resource.getrlimit(resource_type)
    target_soft_limit_stack_size = 1024 * target_soft_limit
    if current_soft < target_soft_limit_stack_size:
        try:
            resource.setrlimit(
                resource_type, (target_soft_limit_stack_size, current_hard)
            )
        except ValueError as e:
            logger.warning(f"Fail to set RLIMIT_STACK: {e}")

1024

1025
def add_api_key_middleware(app, api_key: str):
1026
1027
1028
1029
1030
1031
    @app.middleware("http")
    async def authentication(request, call_next):
        if request.method == "OPTIONS":
            return await call_next(request)
        if request.url.path.startswith("/health"):
            return await call_next(request)
1032
1033
        if request.url.path.startswith("/metrics"):
            return await call_next(request)
1034
        if request.headers.get("Authorization") != "Bearer " + api_key:
1035
            return ORJSONResponse(content={"error": "Unauthorized"}, status_code=401)
1036
        return await call_next(request)
1037
1038


1039
def prepare_model_and_tokenizer(model_path: str, tokenizer_path: str):
1040
    if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
1041
1042
1043
        if not os.path.exists(model_path):
            from modelscope import snapshot_download

1044
1045
            model_path = snapshot_download(model_path)
            tokenizer_path = snapshot_download(
1046
1047
                tokenizer_path, ignore_patterns=["*.bin", "*.safetensors"]
            )
1048
    return model_path, tokenizer_path
1049
1050
1051


def configure_logger(server_args, prefix: str = ""):
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
    if SGLANG_LOGGING_CONFIG_PATH := os.getenv("SGLANG_LOGGING_CONFIG_PATH"):
        if not os.path.exists(SGLANG_LOGGING_CONFIG_PATH):
            raise Exception(
                "Setting SGLANG_LOGGING_CONFIG_PATH from env with "
                f"{SGLANG_LOGGING_CONFIG_PATH} but it does not exist!"
            )
        with open(SGLANG_LOGGING_CONFIG_PATH, encoding="utf-8") as file:
            custom_config = json.loads(file.read())
        logging.config.dictConfig(custom_config)
        return
1062
    format = f"[%(asctime)s{prefix}] %(message)s"
Lianmin Zheng's avatar
Lianmin Zheng committed
1063
    # format = f"[%(asctime)s.%(msecs)03d{prefix}] %(message)s"
1064
1065
1066
    logging.basicConfig(
        level=getattr(logging, server_args.log_level.upper()),
        format=format,
1067
        datefmt="%Y-%m-%d %H:%M:%S",
1068
1069
        force=True,
    )
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080


# source: https://github.com/vllm-project/vllm/blob/93b38bea5dd03e1b140ca997dfaadef86f8f1855/vllm/lora/utils.py#L9
def replace_submodule(
    model: nn.Module, module_name: str, new_module: nn.Module
) -> nn.Module:
    """Replace a submodule in a model with a new module."""
    parent = model.get_submodule(".".join(module_name.split(".")[:-1]))
    target_name = module_name.split(".")[-1]
    setattr(parent, target_name, new_module)
    return new_module
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100


def set_weight_attrs(
    weight: torch.Tensor,
    weight_attrs: Optional[Dict[str, Any]],
):
    """Set attributes on a weight tensor.

    This method is used to set attributes on a weight tensor. This method
    will not overwrite existing attributes.

    Args:
        weight: The weight tensor.
        weight_attrs: A dictionary of attributes to set on the weight tensor.
    """
    if weight_attrs is None:
        return
    for key, value in weight_attrs.items():
        assert not hasattr(weight, key), f"Overwriting existing tensor attribute: {key}"
        setattr(weight, key, value)
1101
1102
1103


def broadcast_pyobj(
1104
1105
1106
    data: List[Any],
    rank: int,
    dist_group: Optional[torch.distributed.ProcessGroup] = None,
1107
    src: int = 0,
1108
    force_cpu_device: bool = True,
1109
):
1110
1111
1112
1113
    """Broadcast inputs from src rank to all other ranks with torch.dist backend.
    The `rank` here refer to the source rank on global process group (regardless
    of dist_group argument).
    """
1114
1115
1116
    device = torch.device(
        "cuda" if torch.cuda.is_available() and not force_cpu_device else "cpu"
    )
1117

1118
    if rank == src:
1119
        if len(data) == 0:
1120
            tensor_size = torch.tensor([0], dtype=torch.long, device=device)
1121
            dist.broadcast(tensor_size, src=src, group=dist_group)
1122
1123
1124
        else:
            serialized_data = pickle.dumps(data)
            size = len(serialized_data)
1125

1126
1127
            tensor_data = torch.ByteTensor(
                np.frombuffer(serialized_data, dtype=np.uint8)
1128
1129
            ).to(device)
            tensor_size = torch.tensor([size], dtype=torch.long, device=device)
1130

1131
1132
            dist.broadcast(tensor_size, src=src, group=dist_group)
            dist.broadcast(tensor_data, src=src, group=dist_group)
1133
1134
        return data
    else:
1135
        tensor_size = torch.tensor([0], dtype=torch.long, device=device)
1136
        dist.broadcast(tensor_size, src=src, group=dist_group)
1137
1138
1139
1140
1141
        size = tensor_size.item()

        if size == 0:
            return []

1142
        tensor_data = torch.empty(size, dtype=torch.uint8, device=device)
1143
        dist.broadcast(tensor_data, src=src, group=dist_group)
1144

1145
        serialized_data = bytes(tensor_data.cpu().numpy())
1146
1147
        data = pickle.loads(serialized_data)
        return data
1148
1149


1150
1151
1152
1153
1154
1155
1156
def point_to_point_pyobj(
    data: List[Any],
    rank: int,
    group: Optional[torch.distributed.ProcessGroup] = None,
    src: int = 0,
    dst: int = 1,
):
1157
    """Send data from src to dst in group using DeviceToDevice communication."""
1158

1159
1160
    if rank == src:
        if len(data) == 0:
1161
1162
1163
            tensor_size = torch.tensor(
                [0], dtype=torch.long, device=torch.cuda.current_device()
            )
1164
1165
1166
1167
1168
1169
            dist.send(tensor_size, dst=dst, group=group)
        else:
            serialized_data = pickle.dumps(data)
            size = len(serialized_data)
            tensor_data = torch.ByteTensor(
                np.frombuffer(serialized_data, dtype=np.uint8)
1170
1171
1172
1173
1174
1175
            ).cuda(
                device=torch.cuda.current_device()
            )  # Move to GPU
            tensor_size = torch.tensor(
                [size], dtype=torch.long, device=torch.cuda.current_device()
            )
1176
1177
1178
1179
1180
1181

            dist.send(tensor_size, dst=dst, group=group)
            dist.send(tensor_data, dst=dst, group=group)
        return data

    elif rank == dst:
1182
1183
1184
        tensor_size = torch.tensor(
            [0], dtype=torch.long, device=torch.cuda.current_device()
        )
1185
1186
1187
1188
1189
1190
        dist.recv(tensor_size, src=src, group=group)
        size = tensor_size.item()

        if size == 0:
            return []

1191
1192
1193
        tensor_data = torch.empty(
            size, dtype=torch.uint8, device=torch.cuda.current_device()
        )
1194
1195
        dist.recv(tensor_data, src=src, group=group)

1196
1197
1198
        serialized_data = bytes(
            tensor_data.cpu().numpy()
        )  # Move back to host for deserialization
1199
1200
1201
1202
1203
1204
1205
        data = pickle.loads(serialized_data)
        return data

    # Other ranks in pp_group do nothing
    return []


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
1233
1234
step_counter = 0


def pytorch_profile(name, func, *args, data_size=-1):
    """
    Args:
        name (string): the name of recorded function.
        func: the function to be profiled.
        args: the arguments of the profiled function.
        data_size (int): some measurement of the computation complexity.
            Usually, it could be the batch size.
    """
    global step_counter
    os.makedirs("trace", exist_ok=True)
    with profile(
        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
        # schedule=torch.profiler.schedule(wait=1, warmup=1, active=3, repeat=2),
        # on_trace_ready=tensorboard_trace_handler('./log_dir'),
        record_shapes=True,
        profile_memory=True,
        with_stack=True,
    ) as prof:
        with record_function(name):
            with open(f"trace/size_{step_counter}.json", "w") as f:
                json.dump({"size": data_size}, f)
            result = func(*args)
    prof.export_chrome_trace(f"trace/{name}_{step_counter}.json")
    step_counter += 1
    return result
1235
1236


Lianmin Zheng's avatar
Lianmin Zheng committed
1237
1238
1239
def get_zmq_socket(
    context: zmq.Context, socket_type: zmq.SocketType, endpoint: str, bind: bool
):
1240
1241
1242
1243
1244
1245
1246
1247
    mem = psutil.virtual_memory()
    total_mem = mem.total / 1024**3
    available_mem = mem.available / 1024**3
    if total_mem > 32 and available_mem > 16:
        buf_size = int(0.5 * 1024**3)
    else:
        buf_size = -1

1248
    socket = context.socket(socket_type)
1249
1250
    if endpoint.find("[") != -1:
        socket.setsockopt(zmq.IPV6, 1)
1251
1252

    def set_send_opt():
1253
        socket.setsockopt(zmq.SNDHWM, 0)
1254
        socket.setsockopt(zmq.SNDBUF, buf_size)
1255
1256

    def set_recv_opt():
1257
        socket.setsockopt(zmq.RCVHWM, 0)
1258
        socket.setsockopt(zmq.RCVBUF, buf_size)
1259
1260
1261
1262
1263
1264
1265
1266

    if socket_type == zmq.PUSH:
        set_send_opt()
    elif socket_type == zmq.PULL:
        set_recv_opt()
    elif socket_type == zmq.DEALER:
        set_send_opt()
        set_recv_opt()
1267
1268
1269
    else:
        raise ValueError(f"Unsupported socket type: {socket_type}")

Lianmin Zheng's avatar
Lianmin Zheng committed
1270
1271
1272
1273
1274
    if bind:
        socket.bind(endpoint)
    else:
        socket.connect(endpoint)

1275
    return socket
1276
1277
1278


def dump_to_file(dirpath, name, value):
1279
    from sglang.srt.distributed import get_tensor_model_parallel_rank
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316

    if get_tensor_model_parallel_rank() != 0:
        return

    os.makedirs(dirpath, exist_ok=True)
    if value.dtype is torch.bfloat16:
        value = value.float()
    value = value.cpu().numpy()
    output_filename = os.path.join(dirpath, f"pytorch_dump_{name}.npy")
    logger.info(f"Dump a tensor to {output_filename}. Shape = {value.shape}")
    np.save(output_filename, value)


def is_triton_3():
    return triton.__version__.startswith("3.")


def maybe_torch_compile(*args, **kwargs):
    """
    torch.compile does not work for triton 2.2.0, which is needed in xlm1's jax.
    Therefore, we disable it here.
    """

    def decorator(func):
        if is_triton_3():
            return torch.compile(*args, **kwargs)(func)
        return func

    return decorator


def delete_directory(dirpath):
    try:
        # This will remove the directory and all its contents
        shutil.rmtree(dirpath)
    except OSError as e:
        print(f"Warning: {dirpath} : {e.strerror}")
Lianmin Zheng's avatar
Lianmin Zheng committed
1317
1318
1319
1320
1321
1322
1323
1324
1325


# Temporary directory for prometheus multiprocess mode
# Cleaned up automatically when this object is garbage collected
prometheus_multiproc_dir: tempfile.TemporaryDirectory


def set_prometheus_multiproc_dir():
    # Set prometheus multiprocess directory
1326
    # sglang uses prometheus multiprocess mode
Lianmin Zheng's avatar
Lianmin Zheng committed
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
    # we need to set this before importing prometheus_client
    # https://prometheus.github.io/client_python/multiprocess/
    global prometheus_multiproc_dir

    if "PROMETHEUS_MULTIPROC_DIR" in os.environ:
        logger.debug("User set PROMETHEUS_MULTIPROC_DIR detected.")
        prometheus_multiproc_dir = tempfile.TemporaryDirectory(
            dir=os.environ["PROMETHEUS_MULTIPROC_DIR"]
        )
    else:
        prometheus_multiproc_dir = tempfile.TemporaryDirectory()
        os.environ["PROMETHEUS_MULTIPROC_DIR"] = prometheus_multiproc_dir.name
    logger.debug(f"PROMETHEUS_MULTIPROC_DIR: {os.environ['PROMETHEUS_MULTIPROC_DIR']}")


def add_prometheus_middleware(app):
1343
    # We need to import prometheus_client after setting the env variable `PROMETHEUS_MULTIPROC_DIR`
Lianmin Zheng's avatar
Lianmin Zheng committed
1344
1345
1346
1347
1348
1349
1350
1351
1352
    from prometheus_client import CollectorRegistry, make_asgi_app, multiprocess

    registry = CollectorRegistry()
    multiprocess.MultiProcessCollector(registry)
    metrics_route = Mount("/metrics", make_asgi_app(registry=registry))

    # Workaround for 307 Redirect for /metrics
    metrics_route.path_regex = re.compile("^/metrics(?P<path>.*)$")
    app.routes.append(metrics_route)
1353
1354


1355
1356
1357
1358
1359
1360
1361
1362
1363
def bind_port(port):
    """Bind to a specific port, assuming it's available."""
    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
    sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)  # Allows address reuse
    sock.bind(("", port))
    sock.listen(1)
    return sock


HAI's avatar
HAI committed
1364
1365
1366
1367
def get_amdgpu_memory_capacity():
    try:
        # Run rocm-smi and capture the output
        result = subprocess.run(
1368
            [
HAI's avatar
HAI committed
1369
                "rocminfo | grep 'gfx' -A 100 | grep 'Pool 1' -A 5 | grep 'Size:' | awk '{print $2}'"
1370
            ],
HAI's avatar
HAI committed
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            shell=True,
            text=True,
        )
        if result.returncode != 0:
            raise RuntimeError(f"rocm-smi error: {result.stderr.strip()}")

        # Parse the output to extract memory values in MiB
        memory_values = [
1381
            float(mem.split("(")[0].strip()) / 1024
HAI's avatar
HAI committed
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
            for mem in result.stdout.strip().split("\n")
        ]

        if not memory_values:
            raise ValueError("No GPU memory values found.")

        # Return the minimum memory value
        return min(memory_values)

    except FileNotFoundError:
        raise RuntimeError(
            "rocm-smi not found. Ensure AMD ROCm drivers are installed and accessible."
        )


1397
1398
1399
1400
1401
1402
1403
def get_device_sm():
    if torch.cuda.is_available():
        major, minor = torch.cuda.get_device_capability()
        return major * 10 + minor
    return 0


HAI's avatar
HAI committed
1404
def get_nvgpu_memory_capacity():
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
    try:
        # Run nvidia-smi and capture the output
        result = subprocess.run(
            ["nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits"],
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True,
        )

        if result.returncode != 0:
            raise RuntimeError(f"nvidia-smi error: {result.stderr.strip()}")

        # Parse the output to extract memory values
        memory_values = [
            float(mem)
            for mem in result.stdout.strip().split("\n")
            if re.match(r"^\d+(\.\d+)?$", mem.strip())
        ]

        if not memory_values:
1425
1426
1427
1428
1429
1430
1431
            # Fallback to torch.cuda.mem_get_info() when failed to get memory capacity from nvidia-smi,
            # typically in NVIDIA MIG mode.
            if torch.cuda.is_available():
                logger.warning(
                    "Failed to get GPU memory capacity from nvidia-smi, falling back to torch.cuda.mem_get_info()."
                )
                return torch.cuda.mem_get_info()[1] // 1024 // 1024  # unit: MB
1432
1433
1434
1435
1436
1437
1438
1439
1440
            raise ValueError("No GPU memory values found.")

        # Return the minimum memory value
        return min(memory_values)

    except FileNotFoundError:
        raise RuntimeError(
            "nvidia-smi not found. Ensure NVIDIA drivers are installed and accessible."
        )
1441
1442


1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
def get_hpu_memory_capacity():
    try:
        # Run hl-smi and capture the output
        result = subprocess.run(
            ["hl-smi --query | grep 'Total'"],
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            shell=True,
            text=True,
        )

        if result.returncode != 0:
            raise RuntimeError(f"hl-smi error: {result.stderr.strip()}")

        # Parse the output to extract memory values in MiB
        memory_values = [
            float(mem.split(" ")[-2]) for mem in result.stdout.strip().split("\n")
        ]

        if not memory_values:
            raise ValueError("No GPU memory values found.")

        # Return the minimum memory value
        return min(memory_values)

    except FileNotFoundError:
        raise RuntimeError(
            "hl-smi not found. Ensure Habana drivers are installed and accessible."
        )


1474
1475
1476
1477
1478
1479
1480
1481
1482
def get_npu_memory_capacity():
    try:
        import torch_npu

        return torch.npu.mem_get_info()[1] // 1024 // 1024  # unit: MB
    except ImportError as e:
        raise ImportError("torch_npu is required when run on npu device.")


Lianmin Zheng's avatar
Lianmin Zheng committed
1483
def get_device_memory_capacity(device: str = None):
1484
1485
1486
1487
1488
1489
    if is_cuda():
        gpu_mem = get_nvgpu_memory_capacity()
    elif is_hip():
        gpu_mem = get_amdgpu_memory_capacity()
    elif device == "hpu":
        gpu_mem = get_hpu_memory_capacity()
1490
1491
    elif device == "npu":
        gpu_mem = get_npu_memory_capacity()
1492
1493
1494
1495
1496
1497
1498
    else:
        # GPU memory is not known yet or no GPU is available.
        gpu_mem = None

    return gpu_mem


1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
# Copy from pytorch and OpenRLHF to allow creating multiple main groups.
# https://github.com/pytorch/pytorch/blob/main/torch/distributed/distributed_c10d.py
# https://github.com/OpenRLHF/OpenRLHF/blob/main/openrlhf/utils/distributed_util.py
def init_custom_process_group(
    backend=None,
    init_method=None,
    timeout=None,
    world_size=-1,
    rank=-1,
    store=None,
    group_name=None,
    pg_options=None,
):
    from torch.distributed.distributed_c10d import (
        Backend,
        PrefixStore,
        _new_process_group_helper,
        _world,
        default_pg_timeout,
        rendezvous,
    )

    assert (store is None) or (
        init_method is None
    ), "Cannot specify both init_method and store."

    if store is not None:
        assert world_size > 0, "world_size must be positive if using store"
        assert rank >= 0, "rank must be non-negative if using store"
    elif init_method is None:
        init_method = "env://"

    if backend:
        backend = Backend(backend)
    else:
        backend = Backend("undefined")

    if timeout is None:
        timeout = default_pg_timeout

    # backward compatible API
    if store is None:
        rendezvous_iterator = rendezvous(init_method, rank, world_size, timeout=timeout)
        store, rank, world_size = next(rendezvous_iterator)
        store.set_timeout(timeout)

        # Use a PrefixStore to avoid accidental overrides of keys used by
        # different systems (e.g. RPC) in case the store is multi-tenant.
        store = PrefixStore(group_name, store)

    # NOTE: The pg_options parameter was renamed into backend_options in PyTorch 2.6.0
    # https://github.com/pytorch/pytorch/commit/a0c7029a75628cd5fa8df83c0de0ea98ee7fd844
    # We need to determine the appropriate parameter name based on PyTorch version
    pg_options_param_name = (
        "backend_options" if str(torch.__version__) >= "2.6" else "pg_options"
    )
    pg, _ = _new_process_group_helper(
        world_size,
        rank,
        [],
        backend,
        store,
        group_name=group_name,
        **{pg_options_param_name: pg_options},
        timeout=timeout,
    )

    _world.pg_group_ranks[pg] = {i: i for i in range(world_size)}

    return pg


1571
1572
def crash_on_warnings():
    # Crash on warning if we are running CI tests
1573
    return get_bool_env_var("SGLANG_IS_IN_CI")
1574
1575


1576
1577
1578
1579
1580
def print_warning_once(msg: str) -> None:
    # Set the stacklevel to 2 to print the caller's line info
    logger.warning(msg, stacklevel=2)


1581
1582
1583
1584
1585
@functools.lru_cache(None)
def print_info_once(msg: str) -> None:
    logger.info(msg)


1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
def get_device_name(device_id: int = 0) -> str:
    if hasattr(torch, "cuda") and torch.cuda.is_available():
        return torch.cuda.get_device_name(device_id)

    if hasattr(torch, "xpu") and torch.xpu.is_available():
        return torch.xpu.get_device_name(device_id)

    if hasattr(torch, "hpu") and torch.hpu.is_available():
        return torch.hpu.get_device_name(device_id)

1596
1597
1598
    if hasattr(torch, "npu") and torch.npu.is_available():
        return torch.npu.get_device_name(device_id)

1599

1600
1601
1602
1603
1604
1605
1606
@lru_cache(maxsize=1)
def is_habana_available() -> bool:
    return find_spec("habana_frameworks") is not None


@lru_cache(maxsize=8)
def get_device(device_id: Optional[int] = None) -> str:
1607
1608
1609
1610
1611
1612
1613
1614
1615
    if is_cpu():
        if cpu_has_amx_support():
            logger.info("Intel AMX is detected, using CPU with Intel AMX support.")
        else:
            logger.warning(
                "CPU device enabled, using torch native backend, low performance expected."
            )
        return "cpu"

1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
    if hasattr(torch, "cuda") and torch.cuda.is_available():
        if device_id is None:
            return "cuda"
        return "cuda:{}".format(device_id)

    if hasattr(torch, "xpu") and torch.xpu.is_available():
        if device_id == None:
            return "xpu"
        return "xpu:{}".format(device_id)

1626
1627
1628
1629
1630
    if hasattr(torch, "npu") and torch.npu.is_available():
        if device_id == None:
            return "npu"
        return "npu:{}".format(device_id)

1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
    if is_habana_available():
        try:
            import habana_frameworks.torch.hpu

            if torch.hpu.is_available():
                if device_id == None:
                    return "hpu"
                return "hpu:{}".format(device_id)
        except ImportError as e:
            raise ImportError(
                "Habana frameworks detected, but failed to import 'habana_frameworks.torch.hpu'."
            )

    raise RuntimeError("No accelerator (CUDA, XPU, HPU) is available.")


@lru_cache(maxsize=1)
def get_device_count() -> int:
    if hasattr(torch, "cuda") and torch.cuda.is_available():
        try:
            return torch.cuda.device_count()
        except RuntimeError:
            return 0

    if hasattr(torch, "xpu") and torch.xpu.is_available():
        try:
            return torch.xpu.device_count()
        except RuntimeError:
            return 0

    if is_habana_available():
        try:
            import habana_frameworks.torch.hpu

            if torch.hpu.is_available():
                return torch.hpu.device_count()
        except (ImportError, RuntimeError):
            return 0

    return 0  # No accelerators available


1673
1674
1675
1676
1677
1678
1679
def get_device_core_count(device_id: int = 0) -> int:
    if hasattr(torch, "cuda") and torch.cuda.is_available():
        return torch.cuda.get_device_properties(device_id).multi_processor_count

    return 0


1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
def get_device_capability(device_id: int = 0) -> Tuple[int, int]:
    major, minor = None, None
    if hasattr(torch, "cuda") and torch.cuda.is_available():
        major, minor = torch.cuda.get_device_capability(device_id)

    if hasattr(torch, "xpu") and torch.xpu.is_available():
        major, minor, *_ = torch.xpu.get_device_capability(device_id)["version"].split(
            "."
        )
        major, minor = int(major), int(minor)

    if hasattr(torch, "hpu") and torch.hpu.is_available():
        try:
1693
1694
1695
1696
            # TODO(HandH1998): `get_device_capability` is not supported by `torch.hpu` for now.
            # Update this once the support is available.
            # major, minor = torch.hpu.get_device_capability(device_id)
            major, minor = None, None
1697
1698
1699
1700
1701
1702
1703
1704
        except Exception as e:
            raise RuntimeError(
                f"An error occurred while getting device capability of hpu: {e}."
            ) from e

    return major, minor


1705
1706
1707
1708
1709
1710
1711
1712
1713
def get_npu_compiler_config():
    config = {
        "frozen_parameter": True,
        "tiling_schedule_optimize": True,
        "topology_sorting_strategy": "StableRDFS",
    }
    return config


1714
1715
1716
1717
def get_compiler_backend() -> str:
    if hasattr(torch, "hpu") and torch.hpu.is_available():
        return "hpu_backend"

1718
    if hasattr(torch, "npu") and torch.npu.is_available():
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
        try:
            import torchair
            import torchair.ge_concrete_graph.ge_converter.experimental.patch_for_hcom_allreduce
            from torchair.configs.compiler_config import CompilerConfig
        except ImportError as e:
            raise ImportError(
                "NPU detected, but torchair package is not installed. "
                "Please install torchair for torch.compile support on NPU."
            )
        compiler_config = CompilerConfig()
        predefined_config = get_npu_compiler_config()
        for k, v in predefined_config.items():
            setattr(compiler_config.experimental_config, k, v)
1732

1733
        npu_backend = torchair.get_npu_backend(compiler_config=compiler_config)
1734
1735
        return npu_backend

1736
1737
1738
    return "inductor"


1739
1740
1741
sglang_lib = Library("sglang", "FRAGMENT")  # noqa


1742
1743
1744
1745
1746
1747
# 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")


1748
1749
1750
1751
1752
1753
1754
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,
):
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
    """
    `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.
    """
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
    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)

    my_lib = target_lib or sglang_lib
    my_lib.define(op_name + schema_str)
    my_lib.impl(op_name, op_func, "CUDA")
    if fake_impl is not None:
        my_lib._register_fake(op_name, fake_impl)
1785
1786


1787
def set_gpu_proc_affinity(
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
    tp_size: int,
    nnodes: int,
    gpu_id: int,
):
    # current process
    pid = os.getpid()
    p = psutil.Process(pid)

    tp_size_per_node = tp_size // nnodes

    # total physical cores
    total_pcores = psutil.cpu_count(logical=False)
    # physical cores per TP (N.B. more Cores than GPUs on node)
    num_cores_bind = total_pcores // tp_size_per_node

    # able to handle multiple DP per node
    start_cpu_id = (gpu_id * num_cores_bind) % total_pcores
    end_cpu_id = start_cpu_id + num_cores_bind

    if psutil.cpu_count() != psutil.cpu_count(logical=False):
        # HT on
Wang Ran (汪然)'s avatar
Wang Ran (汪然) committed
1809
1810
1811
        lower_cpu_ids = [id for id in range(start_cpu_id, end_cpu_id)]
        upper_cpu_ids = [id + total_pcores for id in range(start_cpu_id, end_cpu_id)]
        bind_cpu_ids = list(itertools.chain(lower_cpu_ids, upper_cpu_ids))
1812
1813
1814
1815
1816
1817
1818
    else:
        # HT off
        bind_cpu_ids = [id for id in range(start_cpu_id, end_cpu_id)]

    # set cpu_affinity to current process
    p.cpu_affinity(bind_cpu_ids)
    logger.info(f"Process {pid} gpu_id {gpu_id} is running on CPUs: {p.cpu_affinity()}")
1819
1820


1821
1822
1823
1824
1825
@lru_cache(maxsize=2)
def disable_request_logging() -> bool:
    return get_bool_env_var("SGLANG_DISABLE_REQUEST_LOGGING")


1826
1827
1828
1829
1830
def dataclass_to_string_truncated(
    data, max_length=2048, skip_names: Optional[Set[str]] = None
):
    if skip_names is None:
        skip_names = set()
1831
1832
1833
    if isinstance(data, str):
        if len(data) > max_length:
            half_length = max_length // 2
1834
            return f"{repr(data[:half_length])} ... {repr(data[-half_length:])}"
1835
        else:
1836
            return f"{repr(data)}"
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
    elif isinstance(data, (list, tuple)):
        if len(data) > max_length:
            half_length = max_length // 2
            return str(data[:half_length]) + " ... " + str(data[-half_length:])
        else:
            return str(data)
    elif isinstance(data, dict):
        return (
            "{"
            + ", ".join(
1847
                f"'{k}': {dataclass_to_string_truncated(v, max_length)}"
1848
                for k, v in data.items()
1849
                if k not in skip_names
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
            )
            + "}"
        )
    elif dataclasses.is_dataclass(data):
        fields = dataclasses.fields(data)
        return (
            f"{data.__class__.__name__}("
            + ", ".join(
                f"{f.name}={dataclass_to_string_truncated(getattr(data, f.name), max_length)}"
                for f in fields
1860
                if f.name not in skip_names
1861
1862
1863
            )
            + ")"
        )
1864
    else:
1865
        return str(data)
Tanjiro's avatar
Tanjiro committed
1866
1867


1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
def permute_weight(x: torch.Tensor) -> torch.Tensor:
    b_ = x.shape[0]
    n_ = x.shape[1]
    k_ = x.shape[2]

    x_ = x
    if x.dtype == torch.bfloat16 or x.dtype == torch.float16:
        x_ = x_.view(int(b_), int(n_ / 16), 16, int(k_ / 32), 4, 8)
    elif x.dtype == torch.float8_e4m3fnuz or x.dtype == torch.int8:
        x_ = x_.view(int(b_), int(n_ / 16), 16, int(k_ / 64), 4, 16)
    else:
1879
1880
        # return x_
        x_ = x_.view(int(b_), int(n_ / 16), 16, int(k_ / 8), 2, 4)
1881
1882
1883
1884
1885
1886
1887

    x_ = x_.permute(0, 1, 3, 4, 2, 5)
    x_ = x_.contiguous()
    x_ = x_.view(*x.shape)
    return x_


1888
1889
class MultiprocessingSerializer:
    @staticmethod
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
    def serialize(obj, output_str: bool = False):
        """
        Serialize a Python object using ForkingPickler.

        Args:
            obj: The object to serialize.
            output_str (bool): If True, return a base64-encoded string instead of raw bytes.

        Returns:
            bytes or str: The serialized object.
        """
1901
1902
1903
        buf = io.BytesIO()
        ForkingPickler(buf).dump(obj)
        buf.seek(0)
1904
1905
1906
1907
        output = buf.read()

        if output_str:
            # Convert bytes to base64-encoded string
1908
            output = pybase64.b64encode(output).decode("utf-8")
1909
1910

        return output
1911
1912
1913

    @staticmethod
    def deserialize(data):
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
        """
        Deserialize a previously serialized object.

        Args:
            data (bytes or str): The serialized data, optionally base64-encoded.

        Returns:
            The deserialized Python object.
        """
        if isinstance(data, str):
            # Decode base64 string to bytes
1925
            data = pybase64.b64decode(data, validate=True)
1926

1927
        return ForkingPickler.loads(data)
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938


def debug_timing(func):
    # todo: replace with a more organized instrumentation
    def wrapper(*args, **kwargs):
        if logger.isEnabledFor(logging.DEBUG):
            tic = torch.cuda.Event(enable_timing=True)
            toc = torch.cuda.Event(enable_timing=True)
            tic.record()
            result = func(*args, **kwargs)
            toc.record()
1939
            toc.synchronize()  # Wait for the function to complete without synchronizing all ops on the GPU
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
            elapsed = tic.elapsed_time(toc)
            indices = kwargs.get("indices", args[1] if len(args) > 1 else None)
            num_tokens = len(indices) if indices is not None else 0
            throughput = num_tokens / elapsed * 1000 if elapsed > 0 else 0
            logger.debug(
                f"Transfer time: {elapsed} ms, throughput: {throughput} tokens/s"
            )
            return result
        else:
            return func(*args, **kwargs)

    return wrapper
bjmsong's avatar
bjmsong committed
1952
1953
1954
1955
1956
1957


def nullable_str(val: str):
    if not val or val == "None":
        return None
    return val
1958
1959


1960
1961
1962
1963
1964
1965
1966
1967
1968
def pyspy_dump_schedulers():
    """py-spy dump on all scheduler in a local node."""
    try:
        pid = psutil.Process().pid
        # Command to run py-spy with the PID
        cmd = f"py-spy dump --pid {pid}"
        result = subprocess.run(
            cmd, shell=True, capture_output=True, text=True, check=True
        )
1969
        logger.error(f"Pyspy dump for PID {pid}:\n{result.stdout}")
1970
    except subprocess.CalledProcessError as e:
1971
        logger.error(f"Pyspy failed to dump PID {pid}. Error: {e.stderr}")
1972
1973
1974
1975
1976
1977
1978
1979
1980


def kill_itself_when_parent_died():
    if sys.platform == "linux":
        # sigkill this process when parent worker manager dies
        PR_SET_PDEATHSIG = 1
        libc = ctypes.CDLL("libc.so.6")
        libc.prctl(PR_SET_PDEATHSIG, signal.SIGKILL)
    else:
Wang Ran (汪然)'s avatar
Wang Ran (汪然) committed
1981
        logger.warning("kill_itself_when_parent_died is only supported in linux.")
1982
1983


1984
def set_uvicorn_logging_configs():
1985
1986
    from uvicorn.config import LOGGING_CONFIG

1987
1988
1989
1990
1991
1992
1993
1994
    LOGGING_CONFIG["formatters"]["default"][
        "fmt"
    ] = "[%(asctime)s] %(levelprefix)s %(message)s"
    LOGGING_CONFIG["formatters"]["default"]["datefmt"] = "%Y-%m-%d %H:%M:%S"
    LOGGING_CONFIG["formatters"]["access"][
        "fmt"
    ] = '[%(asctime)s] %(levelprefix)s %(client_addr)s - "%(request_line)s" %(status_code)s'
    LOGGING_CONFIG["formatters"]["access"]["datefmt"] = "%Y-%m-%d %H:%M:%S"
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034


def get_ip() -> str:
    # SGLANG_HOST_IP env can be ignore
    host_ip = os.getenv("SGLANG_HOST_IP", "") or os.getenv("HOST_IP", "")
    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"
        " SGLANG_HOST_IP or HOST_IP.",
        stacklevel=2,
    )
    return "0.0.0.0"


def get_open_port() -> int:
    port = os.getenv("SGLANG_PORT")
    if port is not None:
2035
        port = int(port)
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
        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 is_valid_ipv6_address(address: str) -> bool:
    try:
        ipaddress.IPv6Address(address)
        return True
    except ValueError:
        return False
2062
2063


Vincent's avatar
Vincent committed
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
def configure_ipv6(dist_init_addr):
    addr = dist_init_addr
    end = addr.find("]")
    if end == -1:
        raise ValueError("invalid IPv6 address format: missing ']'")

    host = addr[: end + 1]

    # this only validates the address without brackets: we still need the below checks.
    # if it's invalid, immediately raise an error so we know it's not formatting issues.
    if not is_valid_ipv6_address(host[1:end]):
        raise ValueError(f"invalid IPv6 address: {host}")

    port_str = None
    if len(addr) > end + 1:
        if addr[end + 1] == ":":
            port_str = addr[end + 2 :]
        else:
            raise ValueError("received IPv6 address format: expected ':' after ']'")

    if not port_str:
        raise ValueError(
            "a port must be specified in IPv6 address (format: [ipv6]:port)"
        )

    try:
        port = int(port_str)
    except ValueError:
        raise ValueError(f"invalid port in IPv6 address: '{port_str}'")
    return port, host


2096
def rank0_log(msg: str):
2097
2098
2099
    from sglang.srt.distributed import get_tensor_model_parallel_rank

    if get_tensor_model_parallel_rank() == 0:
2100
        logger.info(msg)
2101
2102


2103
def launch_dummy_health_check_server(host, port):
2104
2105
    import asyncio

2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
    import uvicorn
    from fastapi import FastAPI, Response

    app = FastAPI()

    @app.get("/health")
    async def health():
        """Check the health of the http server."""
        return Response(status_code=200)

    @app.get("/health_generate")
    async def health_generate():
        """Check the health of the http server."""
        return Response(status_code=200)

2121
    config = uvicorn.Config(
2122
2123
2124
2125
        app,
        host=host,
        port=port,
        timeout_keep_alive=5,
2126
2127
2128
        loop="auto",
        log_config=None,
        log_level="warning",
2129
    )
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
    server = uvicorn.Server(config=config)

    try:
        loop = asyncio.get_running_loop()
        logger.info(
            f"Dummy health check server scheduled on existing loop at {host}:{port}"
        )
        loop.create_task(server.serve())

    except RuntimeError:
        logger.info(f"Starting dummy health check server at {host}:{port}")
        server.run()
2142
2143


2144
2145
2146
2147
def create_checksum(directory: str):
    raise NotImplementedError()


2148
2149
2150
2151
2152
def set_cuda_arch():
    if is_flashinfer_available():
        capability = torch.cuda.get_device_capability()
        arch = f"{capability[0]}.{capability[1]}"
        os.environ["TORCH_CUDA_ARCH_LIST"] = f"{arch}{'+PTX' if arch == '9.0' else ''}"
2153
2154


Lianmin Zheng's avatar
Lianmin Zheng committed
2155
2156
2157
2158
2159
2160
2161
def next_power_of_2(n: int):
    return 1 << (n - 1).bit_length() if n > 0 else 1


setattr(triton, "next_power_of_2", next_power_of_2)


2162
2163
2164
2165
2166
class EmptyContextManager:
    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_value, traceback):
2167
2168
2169
        pass


2170
2171
2172
2173
def empty_context(*args, **kwargs):
    return EmptyContextManager()


2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
def add_prefix(name: str, prefix: str) -> str:
    """Add a weight path prefix to a module name.

    Args:
        name: base module name.
        prefix: weight prefix str to added to the front of `name` concatenated with `.`.

    Returns:
        The string `prefix.name` if prefix is non-empty, otherwise just `name`.
    """
    return name if not prefix else f"{prefix}.{name}"
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210


def is_remote_url(url: Union[str, Path]) -> bool:
    """
    Check if the URL is a remote URL of the format:
    <connector_type>://<host>:<port>/<model_name>
    """
    if isinstance(url, Path):
        return False

    pattern = r"(.+)://(.*)"
    m = re.match(pattern, url)
    return m is not None


def parse_connector_type(url: str) -> str:
    """
    Parse the connector type from the URL of the format:
    <connector_type>://<path>
    """
    pattern = r"(.+)://(.*)"
    m = re.match(pattern, url)
    if m is None:
        return ""

    return m.group(1)
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239


def retry(
    fn,
    max_retry: int,
    initial_delay: float = 2.0,
    max_delay: float = 60.0,
    should_retry: Callable[[Any], bool] = lambda e: True,
):
    for try_index in itertools.count():
        try:
            return fn()
        except Exception as e:
            if try_index >= max_retry:
                raise Exception(f"retry() exceed maximum number of retries.")

            if not should_retry(e):
                raise Exception(f"retry() observe errors that should not be retried.")

            delay = min(initial_delay * (2**try_index), max_delay) * (
                0.75 + 0.25 * random.random()
            )

            logger.warning(
                f"retry() failed once ({try_index}th try, maximum {max_retry} retries). Will delay {delay:.2f}s and retry. Error: {e}"
            )
            traceback.print_exc()

            time.sleep(delay)
Mick's avatar
Mick committed
2240
2241
2242
2243
2244
2245
2246
2247
2248


def flatten_nested_list(nested_list):
    if isinstance(nested_list, list):
        return [
            item for sublist in nested_list for item in flatten_nested_list(sublist)
        ]
    else:
        return [nested_list]
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261


class DeepEPMode(Enum):
    normal = "normal"
    low_latency = "low_latency"
    auto = "auto"

    def enable_normal(self):
        return self in [DeepEPMode.normal, DeepEPMode.auto]

    def enable_low_latency(self):
        return self in [DeepEPMode.low_latency, DeepEPMode.auto]

2262
    def resolve(self, is_extend_in_batch: bool):
2263
2264
2265
        if self != DeepEPMode.auto:
            return self

2266
        if is_extend_in_batch:
2267
            return DeepEPMode.normal
2268
2269
        else:
            return DeepEPMode.low_latency
2270
2271


2272
2273
2274
2275
2276
2277
2278
2279
def is_non_idle_and_non_empty(forward_mode, hidden_states):
    return (
        (forward_mode is not None)
        and not forward_mode.is_idle()
        and hidden_states.shape[0] > 0
    )


2280
2281
2282
2283
2284
2285
2286
def fast_topk(values, topk, dim):
    if topk == 1:
        # Use max along the specified dimension to get both value and index
        return torch.max(values, dim=dim, keepdim=True)
    else:
        # Use topk for efficiency with larger k values
        return torch.topk(values, topk, dim=dim)
2287
2288


2289
2290
2291
2292
2293
2294
def bind_or_assign(target, source):
    if target is not None:
        target.copy_(source)
        return target
    else:
        return source
2295
2296


2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
def get_local_ip_auto() -> str:
    interface = os.environ.get("SGLANG_LOCAL_IP_NIC", None)
    return (
        get_local_ip_by_nic(interface)
        if interface is not None
        else get_local_ip_by_remote()
    )


def get_local_ip_by_nic(interface: str) -> str:
    try:
        import netifaces
    except ImportError as e:
        raise ImportError(
            "Environment variable SGLANG_LOCAL_IP_NIC requires package netifaces, please install it through 'pip install netifaces'"
        ) from e

    try:
        addresses = netifaces.ifaddresses(interface)
        if netifaces.AF_INET in addresses:
            for addr_info in addresses[netifaces.AF_INET]:
                ip = addr_info.get("addr")
                if ip and ip != "127.0.0.1" and ip != "0.0.0.0":
                    return ip
        if netifaces.AF_INET6 in addresses:
            for addr_info in addresses[netifaces.AF_INET6]:
                ip = addr_info.get("addr")
                if ip and not ip.startswith("fe80::") and ip != "::1":
                    return ip.split("%")[0]
    except (ValueError, OSError) as e:
        raise ValueError(
            "Can not get local ip from NIC. Please verify whether SGLANG_LOCAL_IP_NIC is set correctly."
        )

    # Fallback
    return get_local_ip_by_remote()


2335
2336
2337
2338
2339
2340
2341
2342
2343
def get_local_ip_by_remote() -> str:
    # 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

2344
2345
2346
2347
2348
2349
2350
2351
    try:
        hostname = socket.gethostname()
        ip = socket.gethostbyname(hostname)
        if ip and ip != "127.0.0.1" and ip != "0.0.0.0":
            return ip
    except Exception:
        pass

2352
2353
2354
2355
2356
2357
2358
2359
    # 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:
Lianmin Zheng's avatar
Lianmin Zheng committed
2360
        raise ValueError("Can not get local ip")
2361
2362
2363
2364
2365
2366


def is_page_size_one(server_args):
    return server_args.page_size == 1


2367
2368
# TODO(hebiao064): Accelerate FA3 Spec Decode with topk > 1.
# TODO(hebiao064): Improve the acc rate for FA3 Spec Decode with topk == 1 and page_size > 1.
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
def is_no_spec_infer_or_topk_one(server_args):
    return server_args.speculative_eagle_topk is None or (
        server_args.speculative_eagle_topk is not None
        and server_args.speculative_eagle_topk == 1
        and is_page_size_one(server_args)
    )


def is_fa3_default_architecture(hf_config):
    architectures = getattr(hf_config, "architectures", None)
    if not isinstance(architectures, list) or not architectures:
        return False
    default_archs = {
        "Qwen2ForCausalLM",
        "Llama4ForConditionalGeneration",
        "LlamaForCausalLM",
Yineng Zhang's avatar
Yineng Zhang committed
2385
        "Gemma2ForCausalLM",
2386
        "Gemma3ForConditionalGeneration",
2387
2388
        "Qwen3ForCausalLM",
        "Qwen3MoeForCausalLM",
2389
2390
    }
    return architectures[0] in default_archs
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403


# Can be more general if it is used in multiple places (keep it simple and thus not general now)
class BumpAllocator:
    def __init__(self, buffer_size: int, dtype, device):
        self._buffer = torch.zeros((buffer_size,), dtype=dtype, device=device)
        self._pointer = 0

    def allocate(self, size: int):
        assert self._pointer + size <= len(self._buffer)
        output = self._buffer[self._pointer : self._pointer + size]
        self._pointer += size
        return output
2404
2405
2406
2407
2408
2409
2410


def log_info_on_rank0(logger, msg):
    from sglang.srt.distributed import get_tensor_model_parallel_rank

    if get_tensor_model_parallel_rank() == 0:
        logger.info(msg)
fzyzcjy's avatar
fzyzcjy committed
2411
2412


2413
2414
2415
2416
2417
2418
2419
def load_json_config(data: str):
    try:
        return json.loads(data)
    except JSONDecodeError:
        return json.loads(Path(data).read_text())


fzyzcjy's avatar
fzyzcjy committed
2420
2421
def dispose_tensor(x: torch.Tensor):
    x.set_(torch.empty((0,), device=x.device, dtype=x.dtype))
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443


T = TypeVar("T")


class Withable(Generic[T]):
    def __init__(self):
        self._value: Optional[T] = None

    @property
    def value(self) -> T:
        return self._value

    @contextmanager
    def with_value(self, new_value: T):
        assert self._value is None
        self._value = new_value
        try:
            yield
        finally:
            assert self._value is new_value
            self._value = None
2444
2445


2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
def require_mlp_tp_gather(server_args):
    """
    Check if the input of MLP is obtained by all-gather rather than all-reduce. This only happens when each MLP TP group contains multiple attention DP groups.
    """
    if server_args.enable_dp_attention:
        assert server_args.dp_size > 1, "dp_size must be greater than 1"
        if (
            server_args.moe_dense_tp_size is None
        ):  # TODO(ch-wan): some MoE models do not have dense layers
            return True
        elif not server_args.enable_dp_lm_head:
            return True
        elif not server_args.enable_deepep_moe:
            return True
        else:
            return (
                server_args.moe_dense_tp_size
                > server_args.tp_size // server_args.dp_size
            )
    else:
        return False


def require_attn_tp_gather(server_args):
    """
    Check if the input of attention is scattered.
    """
    assert server_args.moe_dense_tp_size in [1, None]
    if server_args.enable_deepep_moe or server_args.moe_dense_tp_size == 1:
        if server_args.enable_dp_attention:
            return server_args.dp_size < server_args.tp_size
        else:
            return True
    else:
        return False


def require_gathered_buffer(server_args):
    return require_mlp_tp_gather(server_args) or require_attn_tp_gather(server_args)


def require_mlp_sync(server_args):
    return server_args.enable_dp_attention or require_gathered_buffer(server_args)


2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
def find_local_repo_dir(repo_id: str, revision: Optional[str] = None) -> Optional[str]:
    import huggingface_hub as hf

    # Build cache path
    cache_path = os.path.join(
        hf.constants.HF_HUB_CACHE,
        hf.constants.REPO_ID_SEPARATOR.join(["models", *repo_id.split("/")]),
    )

    # Get revision from main ref if not specified
    if not revision:
        ref_path = os.path.join(cache_path, "refs", "main")
        if os.path.isfile(ref_path):
            with open(ref_path) as f:
                revision = f.read().strip()

    # List files from revision directory
    if revision:
        rev_dir = os.path.join(cache_path, "snapshots", revision)
        if os.path.isdir(rev_dir):
            return rev_dir

    return None


def read_system_prompt_from_file(model_name: str) -> str:
    """Read system prompt from a file in the HuggingFace cache directory.

    Args:
        model_name: The model name to construct the file path

    Returns:
        The system prompt content from the file, or empty string if file not found
    """
    try:
        local_repo_dir = find_local_repo_dir(model_name)
        if local_repo_dir:
            system_prompt_file = os.path.join(local_repo_dir, "SYSTEM_PROMPT.txt")
            if os.path.exists(system_prompt_file):
                with open(system_prompt_file, "r", encoding="utf-8") as f:
                    return f.read()

        return ""
    except Exception:
        # If anything fails, return empty string
        return ""
2537
2538
2539
2540
2541
2542
2543
2544


def bind_or_assign(target, source):
    if target is not None:
        target.copy_(source)
        return target
    else:
        return source
2545
2546


2547
2548
2549
2550
2551
def prepack_weight_if_needed(weight):
    if weight.device != torch.device("cpu"):
        return weight
    if not cpu_has_amx_support():
        return weight
2552

2553
    return torch.ops.sgl_kernel.convert_weight_packed(weight)
2554
2555


2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
# TODO: currently gemm kernel has the below requirements:
# OC % TILE_N == 0, where TILE_N = 16
# IC % TILE_K == 0, where TILE_K = 32
def dim_is_supported(weight):
    return weight.size(0) % 16 == 0 and weight.size(1) % 32 == 0


def _process_weight_after_loading(module, weight_names, transpose_dims=None) -> None:
    # Pack weight for get better performance on CPU
    devices = {getattr(module, weight_name).device for weight_name in weight_names}
    assert len(devices) == 1, f"Expects all weights to be on the same device"
    device = devices.pop()

    if transpose_dims:
        assert len(weight_names) == len(
            transpose_dims
        ), "len(weight_names) should be equal to len(transpose_dims)"

    for i, weight_name in enumerate(weight_names):
        weight_tensor = getattr(module, weight_name)

        # We don't pack weight or use intel amx backend if any weight of this module has unsupported dim.
        if not dim_is_supported(weight_tensor):
            logger.warning(
                f"Expects weight.size(0) % 16 == 0 and weight.size(1) % 32 == 0 "
                f"but {weight_tensor.size(0)=} and {weight_tensor.size(1)=} in {module}. "
                f"{module} won't use intel amx backend."
            )
            module.use_intel_amx_backend = False
            return

        if transpose_dims and transpose_dims[i]:
            weight_tensor = weight_tensor.transpose(*transpose_dims[i])

        packed_weight = torch.nn.Parameter(
            prepack_weight_if_needed(weight_tensor),
            requires_grad=False,
        )
        packed_weight.__dict__ = weight_tensor.__dict__
        setattr(module, weight_name, packed_weight)

    module.use_intel_amx_backend = (
        device == torch.device("cpu") and cpu_has_amx_support()
2599
2600
    )

2601
2602
2603
2604
2605
2606
    if (
        module.use_intel_amx_backend
        and hasattr(module, "bias")
        and module.bias is not None
    ):
        module.bias = torch.nn.Parameter(module.bias.data.float(), requires_grad=False)
2607

2608

2609
2610
2611
2612
class PackWeightMethod:
    def __init__(self, weight_names, transpose_dims=None):
        self.weight_names = weight_names
        self.transpose_dims = transpose_dims
2613

2614
2615
    def process_weights_after_loading(self, module) -> None:
        _process_weight_after_loading(module, self.weight_names, self.transpose_dims)
2616
2617


2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
class LazyValue:
    def __init__(self, creator: Callable):
        self._creator = creator
        self._value = None

    @property
    def value(self):
        if self._creator is not None:
            self._value = self._creator()
            self._creator = None
        return self._value
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641


def dynamic_import(func_path: str):
    parts = func_path.split(".")
    if len(parts) < 2:
        raise ValueError(
            "func_path should contain both module name and func name (such as 'module.func')"
        )
    module_path = ".".join(parts[:-1])
    func_name = parts[-1]
    module = importlib.import_module(module_path)
    func = getattr(module, func_name)
    return func
fzyzcjy's avatar
fzyzcjy committed
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666


def configure_gc_logger():
    logger.info("Enable GC Logger")

    import gc

    gc_start_time = {}

    def gc_callback(phase, info):
        gen = info.get("generation", "?")
        if phase == "start":
            gc_start_time[gen] = time.time()
            logger.info(f"GC start: Time {time.time()} | Generation {gen}")
        elif phase == "stop":
            duration = time.time() - gc_start_time.get(gen, time.time())
            collected = info.get("collected", "?")
            uncollectable = info.get("uncollectable", "?")
            logger.info(
                f"GC end: Time {time.time()} | Generation {gen} | "
                f"Duration: {duration:.4f}s | Collected: {collected} | Uncollectable: {uncollectable} "
                f'{"(LONG GC)" if duration > 0.1 else ""}'
            )

    gc.callbacks.append(gc_callback)
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676


# COPIED FROM DeepGEMM
def align(x: int, y: int) -> int:
    return ceil_div(x, y) * y


# COPIED FROM DeepGEMM
def ceil_div(x: int, y: int) -> int:
    return (x + y - 1) // y
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742


def parse_lscpu_topology():
    try:
        # Get CPU topology: CPU,Core,Socket,Node
        output = subprocess.check_output(
            ["lscpu", "-p=CPU,Core,Socket,Node"], text=True
        )
    except Exception as e:
        raise RuntimeError(f"Unexpected error running 'lscpu': {e}")

    # Parse only data lines (skip comments)
    cpu_info = []
    for line in output.splitlines():
        if not line.startswith("#"):
            cpu, core, socket, node = map(int, line.strip().split(","))
            cpu_info.append((cpu, core, socket, node))

    # [(0,0,0,0),(1,1,0,0),...,(43,43,0,1),...,(256,0,0,0),...]
    return cpu_info


def get_physical_cpus_by_numa():
    cpu_info = parse_lscpu_topology()

    # Map NUMA node -> set of (core_id, socket) to avoid duplicates
    # 0: {(0,0): 0, (1, 0): 1,...}
    # ...
    # 5: {(214,1): 214, (215,1): 215}
    physical_by_node = defaultdict(dict)  # node -> core_id -> cpu_id

    for cpu, core, socket, node in cpu_info:
        key = (core, socket)
        if key not in physical_by_node[node]:
            physical_by_node[node][
                key
            ] = cpu  # pick first CPU seen for that physical core

    # Retrieves CPUs that the current process is allowed to run on
    cpus_allowed_list = psutil.Process().cpu_affinity()

    # Convert to list of physical CPUs per node
    # 0: [0,1,2,...,42]
    # ...
    # 2: [86,87,...,127]
    # ...
    # 5: [214,215,...,255]
    node_to_cpus = {}
    for node, core_to_cpu in physical_by_node.items():
        cpus = sorted(core_to_cpu.values())
        allowed_cpus = set(cpus).intersection(cpus_allowed_list)
        node_to_cpus[node] = allowed_cpus

    return node_to_cpus


# Only physical cores are used. Logical cores are excluded.
def get_cpu_ids_by_node():
    node_to_cpus = get_physical_cpus_by_numa()
    # Sort by NUMA node index
    cpu_ids = [
        ",".join(map(str, sorted(node_to_cpus[node]))) for node in sorted(node_to_cpus)
    ]

    # ['0,1,2,3', '4,5,6,7', '8,9,10,11', '12,13,14,15', '16,17,18,19', '20,21,22,23']
    return cpu_ids
2743
2744
2745
2746
2747
2748
2749
2750
2751


def is_shm_available(dtype, world_size, local_size):
    return (
        cpu_has_amx_support()
        and dtype in [torch.bfloat16, torch.float]
        and world_size >= 1
        and world_size == local_size
    )
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796


def lru_cache_frozenset(maxsize=128):
    def _to_hashable(o):
        try:
            hash(o)
            return o
        except TypeError:
            # Not hashable; convert based on type
            if isinstance(o, (dict)):
                return frozenset(
                    (_to_hashable(k), _to_hashable(v)) for k, v in o.items()
                )
            elif isinstance(o, set):
                return frozenset(_to_hashable(v) for v in o)
            elif isinstance(o, (list, tuple)) or (
                isinstance(o, Sequence) and not isinstance(o, (str, bytes))
            ):
                return tuple(_to_hashable(v) for v in o)
            else:
                raise TypeError(f"Cannot make hashable: {type(o)}")

    def decorator(func):
        cache = OrderedDict()

        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            h_args = tuple(_to_hashable(a) for a in args)
            h_kwargs = frozenset(
                (_to_hashable(k), _to_hashable(v)) for k, v in kwargs.items()
            )
            key = (h_args, h_kwargs)
            if key in cache:
                cache.move_to_end(key)
                return cache[key]
            result = func(*args, **kwargs)
            cache[key] = result
            if maxsize is not None and len(cache) > maxsize:
                cache.popitem(last=False)
            return result

        wrapper.cache_clear = cache.clear  # For manual cache clearing
        return wrapper

    return decorator
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894


def apply_module_patch(target_module, target_function, wrappers):
    original_module, original_function = parse_module_path(
        target_module, target_function, False
    )

    original_function_id = id(original_function)

    candidate = original_function
    for wrapper in wrappers:
        candidate = wrapper(candidate)
    if target_function is not None:
        setattr(original_module, target_function, candidate)

    for key, value in sys.modules.copy().items():
        if (
            target_function is not None
            and hasattr(value, target_function)
            and id(getattr(value, target_function)) == original_function_id
        ):
            setattr(value, target_function, candidate)


def parse_module_path(module_path, function_name, create_dummy):
    from importlib.machinery import ModuleSpec

    def create_dummy_module(full_path, parent=None):
        """Create and register a placeholder module"""
        dummy = types.ModuleType(full_path)
        dummy.__file__ = "vllm_ascend.dummy_module.py"
        dummy.__spec__ = ModuleSpec(full_path, None)
        sys.modules[full_path] = dummy
        if parent:
            setattr(parent, full_path.split(".")[-1], dummy)
        return dummy

    def create_placeholder_function(func_name):
        """Create dummy function that raises when called"""

        def placeholder(*args, **kwargs):
            raise NotImplementedError(f"Function {func_name} is a placeholder")

        placeholder.__name__ = func_name
        return placeholder

    modules = module_path.split(".")
    current_module = None
    processed_path = []

    for idx, part in enumerate(modules):
        current_path = ".".join(modules[: idx + 1])
        parent_path = ".".join(modules[:idx]) if idx > 0 else None

        try:
            current_module = importlib.import_module(current_path)
        except ModuleNotFoundError:
            # Handle missing module
            parent = importlib.import_module(parent_path) if parent_path else None
            if parent and hasattr(parent, part):
                # Use existing attribute from parent
                current_module = getattr(parent, part)
                # Check for early function resolution
                if function_name and hasattr(current_module, function_name):
                    return current_module, getattr(current_module, function_name)
                if function_name and create_dummy:
                    ph_func = create_placeholder_function(function_name)
                    setattr(current_module, function_name, ph_func)
                    return current_module, ph_func
                if function_name:
                    raise AttributeError(
                        f"Function {function_name} missing in {current_path}"
                    )
            else:
                if not create_dummy:
                    raise
                # Create and register dummy module
                current_module = create_dummy_module(
                    current_path,
                    parent=(
                        importlib.import_module(parent_path) if parent_path else None
                    ),
                )

        processed_path.append(part)

    # Final function handling
    final_module = sys.modules[module_path]
    if function_name is not None:
        if not hasattr(final_module, function_name):
            if create_dummy:
                ph_func = create_placeholder_function(function_name)
                setattr(final_module, function_name, ph_func)
            else:
                setattr(final_module, function_name, None)
        return final_module, getattr(final_module, function_name)

    return final_module, None