__init__.py 10.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
7
8
9
import logging
import traceback
from itertools import chain
from typing import TYPE_CHECKING, Optional

from vllm.plugins import load_plugins_by_group
10
from vllm.utils import resolve_obj_by_qualname, supports_xccl
11

12
from .interface import _Backend  # noqa: F401
13
from .interface import CpuArchEnum, Platform, PlatformEnum
zhuwenwen's avatar
zhuwenwen committed
14
import torch
15

16
logger = logging.getLogger(__name__)
17

18

19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
def vllm_version_matches_substr(substr: str) -> bool:
    """
    Check to see if the vLLM version matches a substring.
    """
    from importlib.metadata import PackageNotFoundError, version
    try:
        vllm_version = version("vllm")
    except PackageNotFoundError as e:
        logger.warning(
            "The vLLM package was not found, so its version could not be "
            "inspected. This may cause platform detection to fail.")
        raise e
    return substr in vllm_version


34
35
def tpu_platform_plugin() -> Optional[str]:
    is_tpu = False
36
    logger.debug("Checking if TPU platform is available.")
37
38
39
40
41
42
43
    try:
        # While it's technically possible to install libtpu on a
        # non-TPU machine, this is a very uncommon scenario. Therefore,
        # we assume that libtpu is installed if and only if the machine
        # has TPUs.
        import libtpu  # noqa: F401
        is_tpu = True
44
45
46
        logger.debug("Confirmed TPU platform is available.")
    except Exception as e:
        logger.debug("TPU platform is not available because: %s", str(e))
47

48
    return "vllm.platforms.tpu.TpuPlatform" if is_tpu else None
49
50


51
52
def cuda_platform_plugin() -> Optional[str]:
    is_cuda = False
53
    logger.debug("Checking if CUDA platform is available.")
54
    try:
55
56
        from vllm.utils import import_pynvml
        pynvml = import_pynvml()
57
58
        pynvml.nvmlInit()
        try:
59
60
61
62
63
64
            # NOTE: Edge case: vllm cpu build on a GPU machine.
            # Third-party pynvml can be imported in cpu build,
            # we need to check if vllm is built with cpu too.
            # Otherwise, vllm will always activate cuda plugin
            # on a GPU machine, even if in a cpu build.
            is_cuda = (pynvml.nvmlDeviceGetCount() > 0
65
                       and not vllm_version_matches_substr("cpu"))
66
67
68
69
70
71
72
73
            if pynvml.nvmlDeviceGetCount() <= 0:
                logger.debug(
                    "CUDA platform is not available because no GPU is found.")
            if vllm_version_matches_substr("cpu"):
                logger.debug("CUDA platform is not available because"
                             " vLLM is built with CPU.")
            if is_cuda:
                logger.debug("Confirmed CUDA platform is available.")
74
75
        finally:
            pynvml.nvmlShutdown()
76
    except Exception as e:
77
78
        logger.debug("Exception happens when checking CUDA platform: %s",
                     str(e))
79
80
81
82
        if "nvml" not in e.__class__.__name__.lower():
            # If the error is not related to NVML, re-raise it.
            raise e

83
84
85
86
87
88
89
90
        # CUDA is supported on Jetson, but NVML may not be.
        import os

        def cuda_is_jetson() -> bool:
            return os.path.isfile("/etc/nv_tegra_release") \
                or os.path.exists("/sys/class/tegra-firmware")

        if cuda_is_jetson():
91
            logger.debug("Confirmed CUDA platform is available on Jetson.")
92
            is_cuda = True
93
94
        else:
            logger.debug("CUDA platform is not available because: %s", str(e))
95

96
97
98
99
100
    return "vllm.platforms.cuda.CudaPlatform" if is_cuda else None


def rocm_platform_plugin() -> Optional[str]:
    is_rocm = False
101
    logger.debug("Checking if ROCm platform is available.")
102
103
104
105
106
107
    try:
        import amdsmi
        amdsmi.amdsmi_init()
        try:
            if len(amdsmi.amdsmi_get_processor_handles()) > 0:
                is_rocm = True
108
                logger.debug("Confirmed ROCm platform is available.")
109
110
111
            else:
                logger.debug("ROCm platform is not available because"
                             " no GPU is found.")
112
113
        finally:
            amdsmi.amdsmi_shut_down()
114
115
    except Exception as e:
        logger.debug("ROCm platform is not available because: %s", str(e))
116
117
118
119
120
121

    return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None


def xpu_platform_plugin() -> Optional[str]:
    is_xpu = False
122
    logger.debug("Checking if XPU platform is available.")
123
124
125
126
    try:
        # installed IPEX if the machine has XPUs.
        import intel_extension_for_pytorch  # noqa: F401
        import torch
127
128
129
130
131
132
        if supports_xccl():
            dist_backend = "xccl"
        else:
            dist_backend = "ccl"
            import oneccl_bindings_for_pytorch  # noqa: F401

133
134
        if hasattr(torch, 'xpu') and torch.xpu.is_available():
            is_xpu = True
135
136
137
138
            from vllm.platforms.xpu import XPUPlatform
            XPUPlatform.dist_backend = dist_backend
            logger.debug("Confirmed %s backend is available.",
                         XPUPlatform.dist_backend)
139
140
141
            logger.debug("Confirmed XPU platform is available.")
    except Exception as e:
        logger.debug("XPU platform is not available because: %s", str(e))
142
143
144
145
146
147

    return "vllm.platforms.xpu.XPUPlatform" if is_xpu else None


def cpu_platform_plugin() -> Optional[str]:
    is_cpu = False
148
    logger.debug("Checking if CPU platform is available.")
149
    try:
150
        is_cpu = vllm_version_matches_substr("cpu")
151
152
153
        if is_cpu:
            logger.debug("Confirmed CPU platform is available because"
                         " vLLM is built with CPU.")
154
        if not is_cpu:
155
156
            import sys
            is_cpu = sys.platform.startswith("darwin")
157
158
            if is_cpu:
                logger.debug("Confirmed CPU platform is available"
159
                             " because the machine is MacOS.")
160

161
162
    except Exception as e:
        logger.debug("CPU platform is not available because: %s", str(e))
163
164
165
166
167

    return "vllm.platforms.cpu.CpuPlatform" if is_cpu else None


def neuron_platform_plugin() -> Optional[str]:
168
169
    tnx_installed = False
    nxd_installed = False
170
    logger.debug("Checking if Neuron platform is available.")
171
172
    try:
        import transformers_neuronx  # noqa: F401
173
        tnx_installed = True
174
175
        logger.debug("Confirmed Neuron platform is available because"
                     " transformers_neuronx is found.")
176
    except ImportError:
177
        pass
178

179
180
181
182
183
184
    try:
        import neuronx_distributed_inference  # noqa: F401
        nxd_installed = True
        logger.debug("Confirmed Neuron platform is available because"
                     " neuronx_distributed_inference is found.")
    except ImportError:
185
        pass
186

187
    is_neuron = tnx_installed or nxd_installed
188
    return "vllm.platforms.neuron.NeuronPlatform" if is_neuron else None
189
190


191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
builtin_platform_plugins = {
    'tpu': tpu_platform_plugin,
    'cuda': cuda_platform_plugin,
    'rocm': rocm_platform_plugin,
    'xpu': xpu_platform_plugin,
    'cpu': cpu_platform_plugin,
    'neuron': neuron_platform_plugin,
}


def resolve_current_platform_cls_qualname() -> str:
    platform_plugins = load_plugins_by_group('vllm.platform_plugins')

    activated_plugins = []

    for name, func in chain(builtin_platform_plugins.items(),
                            platform_plugins.items()):
        try:
            assert callable(func)
            platform_cls_qualname = func()
            if platform_cls_qualname is not None:
                activated_plugins.append(name)
        except Exception:
            pass

    activated_builtin_plugins = list(
        set(activated_plugins) & set(builtin_platform_plugins.keys()))
    activated_oot_plugins = list(
        set(activated_plugins) & set(platform_plugins.keys()))

    if len(activated_oot_plugins) >= 2:
        raise RuntimeError(
            "Only one platform plugin can be activated, but got: "
            f"{activated_oot_plugins}")
    elif len(activated_oot_plugins) == 1:
        platform_cls_qualname = platform_plugins[activated_oot_plugins[0]]()
        logger.info("Platform plugin %s is activated",
                    activated_oot_plugins[0])
    elif len(activated_builtin_plugins) >= 2:
        raise RuntimeError(
            "Only one platform plugin can be activated, but got: "
            f"{activated_builtin_plugins}")
    elif len(activated_builtin_plugins) == 1:
        platform_cls_qualname = builtin_platform_plugins[
            activated_builtin_plugins[0]]()
        logger.info("Automatically detected platform %s.",
                    activated_builtin_plugins[0])
    else:
239
        platform_cls_qualname = "vllm.platforms.interface.UnspecifiedPlatform"
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
        logger.info(
            "No platform detected, vLLM is running on UnspecifiedPlatform")
    return platform_cls_qualname


_current_platform = None
_init_trace: str = ''

if TYPE_CHECKING:
    current_platform: Platform


def __getattr__(name: str):
    if name == 'current_platform':
        # lazy init current_platform.
        # 1. out-of-tree platform plugins need `from vllm.platforms import
        #    Platform` so that they can inherit `Platform` class. Therefore,
        #    we cannot resolve `current_platform` during the import of
        #    `vllm.platforms`.
        # 2. when users use out-of-tree platform plugins, they might run
        #    `import vllm`, some vllm internal code might access
        #    `current_platform` during the import, and we need to make sure
        #    `current_platform` is only resolved after the plugins are loaded
        #    (we have tests for this, if any developer violate this, they will
        #    see the test failures).
        global _current_platform
        if _current_platform is None:
            platform_cls_qualname = resolve_current_platform_cls_qualname()
            _current_platform = resolve_obj_by_qualname(
                platform_cls_qualname)()
            global _init_trace
            _init_trace = "".join(traceback.format_stack())
        return _current_platform
273
    elif name in globals():
274
        return globals()[name]
275
276
277
    else:
        raise AttributeError(
            f"No attribute named '{name}' exists in {__name__}.")
278
279
280
281
282
283


__all__ = [
    'Platform', 'PlatformEnum', 'current_platform', 'CpuArchEnum',
    "_init_trace"
]