Skip to content

启动器

lazyllm.LazyLLMLaunchersBase

Bases: object

用于统一管理外部进程或分布式作业（训练/推理等）生命周期的启动器抽象基类。不同平台（本地、SLURM、K8s、云资源等）的具体启动器应继承该类并实现核心接口。

Source code in lazyllm/launcher/base.py 
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
class LazyLLMLaunchersBase(object, metaclass=LazyLLMRegisterMetaClass):
    """用于统一管理外部进程或分布式作业（训练/推理等）生命周期的启动器抽象基类。不同平台（本地、SLURM、K8s、云资源等）的具体启动器应继承该类并实现核心接口。

Args:
    None.
"""
    Status = Status

    def __init__(self) -> None:
        self._id = str(uuid.uuid4().hex)

    def makejob(self, cmd):
        """根据给定命令创建并返回作业/进程句柄。需由子类实现。

Args:
    cmd: 用于创建作业的命令或配置（如字符串、参数列表或作业描述对象）。

Raises:
    NotImplementedError: 基类未实现，子类必须覆盖。
"""
        raise NotImplementedError

    def launch(self, *args, **kw):
        """启动一个或多个作业，并将其登记到 all_processes[self._id] 中。需由子类实现。

Args:
    *args: 与具体实现相关的位置参数。
    **kw: 与具体实现相关的关键字参数。

Raises:
    NotImplementedError: 基类未实现，子类必须覆盖。
"""
        raise NotImplementedError

    def cleanup(self):
        """停止并清理当前启动器登记的所有作业，从 all_processes 中移除相应记录，并在最后阻塞等待作业结束。

Args:
    None.
"""
        for k, v in self.all_processes[self._id]:
            v.stop()
            LOG.info(f'killed job:{k}')
        self.all_processes.pop(self._id)
        self.wait()

    @property
    def status(self):
        if len(self.all_processes[self._id]) == 1:
            return self.all_processes[self._id][0][1].status
        elif len(self.all_processes[self._id]) == 0:
            return Status.Cancelled
        raise RuntimeError('More than one tasks are found in one launcher!')

    @property
    def log_path(self):
        if len(self.all_processes[self._id]) == 1:
            return self.all_processes[self._id][0][1].log_path
        elif len(self.all_processes[self._id]) == 0:
            return None
        raise RuntimeError('More than one tasks are found in one launcher!')

    def wait(self):
        """阻塞等待当前启动器登记的所有作业结束。

Args:
    None.
"""
        for _, v in self.all_processes[self._id]:
            v.wait()

    def clone(self):
        """深拷贝当前启动器实例并分配新的唯一 _id，返回克隆后的实例。

Args:
    None.

**Returns:**

- LazyLLMLaunchersBase: 克隆出的启动器实例。
"""
        new = copy.deepcopy(self)
        new._id = str(uuid.uuid4().hex)
        return new

makejob(cmd)

根据给定命令创建并返回作业/进程句柄。需由子类实现。

Parameters:

  • cmd –

    用于创建作业的命令或配置（如字符串、参数列表或作业描述对象）。

Raises:

  • NotImplementedError –

    基类未实现，子类必须覆盖。

Source code in lazyllm/launcher/base.py 
42
43
44
45
46
47
48
49
50
51
    def makejob(self, cmd):
        """根据给定命令创建并返回作业/进程句柄。需由子类实现。

Args:
    cmd: 用于创建作业的命令或配置（如字符串、参数列表或作业描述对象）。

Raises:
    NotImplementedError: 基类未实现，子类必须覆盖。
"""
        raise NotImplementedError

launch(*args, **kw)

启动一个或多个作业，并将其登记到 all_processes[self._id] 中。需由子类实现。

Parameters:

  • *args –

    与具体实现相关的位置参数。

  • **kw –

    与具体实现相关的关键字参数。

Raises:

  • NotImplementedError –

    基类未实现，子类必须覆盖。

Source code in lazyllm/launcher/base.py 
53
54
55
56
57
58
59
60
61
62
63
    def launch(self, *args, **kw):
        """启动一个或多个作业，并将其登记到 all_processes[self._id] 中。需由子类实现。

Args:
    *args: 与具体实现相关的位置参数。
    **kw: 与具体实现相关的关键字参数。

Raises:
    NotImplementedError: 基类未实现，子类必须覆盖。
"""
        raise NotImplementedError

cleanup()

停止并清理当前启动器登记的所有作业，从 all_processes 中移除相应记录，并在最后阻塞等待作业结束。

Source code in lazyllm/launcher/base.py 
65
66
67
68
69
70
71
72
73
74
75
    def cleanup(self):
        """停止并清理当前启动器登记的所有作业，从 all_processes 中移除相应记录，并在最后阻塞等待作业结束。

Args:
    None.
"""
        for k, v in self.all_processes[self._id]:
            v.stop()
            LOG.info(f'killed job:{k}')
        self.all_processes.pop(self._id)
        self.wait()

wait()

阻塞等待当前启动器登记的所有作业结束。

Source code in lazyllm/launcher/base.py 
 93
 94
 95
 96
 97
 98
 99
100
    def wait(self):
        """阻塞等待当前启动器登记的所有作业结束。

Args:
    None.
"""
        for _, v in self.all_processes[self._id]:
            v.wait()

clone()

深拷贝当前启动器实例并分配新的唯一 _id，返回克隆后的实例。

Returns:

  • LazyLLMLaunchersBase: 克隆出的启动器实例。
Source code in lazyllm/launcher/base.py 
102
103
104
105
106
107
108
109
110
111
112
113
114
    def clone(self):
        """深拷贝当前启动器实例并分配新的唯一 _id，返回克隆后的实例。

Args:
    None.

**Returns:**

- LazyLLMLaunchersBase: 克隆出的启动器实例。
"""
        new = copy.deepcopy(self)
        new._id = str(uuid.uuid4().hex)
        return new

lazyllm.launcher.EmptyLauncher

Bases: LazyLLMLaunchersBase

此类是 LazyLLMLaunchersBase 的子类，作为一个本地的启动器。

Parameters:

  • subprocess (bool, default: False ) –

    是否使用子进程来启动。默认为 False。

  • sync (bool, default: True ) –

    是否同步执行作业。默认为 True，否则为异步执行。

Examples:

>>> import lazyllm
>>> launcher = lazyllm.launchers.empty()
Source code in lazyllm/launcher/base.py 
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
@final
class EmptyLauncher(LazyLLMLaunchersBase):
    """此类是 ``LazyLLMLaunchersBase`` 的子类，作为一个本地的启动器。

Args:
    subprocess (bool): 是否使用子进程来启动。默认为 `False`。
    sync (bool): 是否同步执行作业。默认为 `True`，否则为异步执行。


Examples:
    >>> import lazyllm
    >>> launcher = lazyllm.launchers.empty()
    """
    all_processes = defaultdict(list)

    @final
    class Job(Job):
        """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
        def __init__(self, cmd, launcher, *, sync=True):
            super(__class__, self).__init__(cmd, launcher, sync=sync)

        def _wrap_cmd(self, cmd):
            if self._launcher.ngpus == 0:
                return cmd
            gpus = self._launcher._get_idle_gpus()
            if gpus and lazyllm.config['cuda_visible']:
                if self._launcher.ngpus is None:
                    empty_cmd = f'export CUDA_VISIBLE_DEVICES={gpus[0]} && '
                elif self._launcher.ngpus <= len(gpus):
                    empty_cmd = 'export CUDA_VISIBLE_DEVICES=' + \
                                ','.join([str(n) for n in gpus[:self._launcher.ngpus]]) + ' && '
                else:
                    error_info = (f'Not enough GPUs available. Requested {self._launcher.ngpus} GPUs, '
                                  f'but only {len(gpus)} are available.')
                    LOG.error(error_info)
                    raise error_info
            else:
                empty_cmd = ''
            return empty_cmd + cmd

        def stop(self):
            if self.ps:
                try:
                    parent = psutil.Process(self.ps.pid)
                    for child in parent.children(recursive=True):
                        child.kill()
                    parent.kill()
                except psutil.NoSuchProcess:
                    LOG.warning(f'Process with PID {self.ps.pid} does not exist.')
                except psutil.AccessDenied:
                    LOG.warning(f'Permission denied when trying to kill process with PID {self.ps.pid}.')
                except Exception as e:
                    LOG.warning(f'An error occurred: {e}')

        @property
        def status(self):
            return_code = self.ps.poll()
            if return_code is None: job_status = Status.Running
            elif return_code == 0: job_status = Status.Done
            else: job_status = Status.Failed
            return job_status

        def _get_jobid(self):
            self.jobid = self.ps.pid if self.ps else None

        def get_jobip(self):
            return '127.0.0.1'

        def wait(self):
            if self.ps:
                self.ps.wait()

    def __init__(self, subprocess=False, ngpus=None, sync=True, **kwargs):
        super().__init__()
        self.subprocess = subprocess
        self.sync = sync
        self.ngpus = ngpus

    def makejob(self, cmd):
        return EmptyLauncher.Job(cmd, launcher=self, sync=self.sync)

    def launch(self, f, *args, **kw):
        if isinstance(f, EmptyLauncher.Job):
            f.start()
            return f.return_value
        elif callable(f):
            if not self.subprocess:
                return f(*args, **kw)
            else:
                LOG.info('Async execution of callable object is not supported currently.')
                p = multiprocessing.Process(target=f, args=args, kwargs=kw)
                p.start()
                p.join()
        else:
            raise RuntimeError('Invalid cmd given, please check the return value of cmd.')

    def _get_idle_gpus(self):
        try:
            order_list = subprocess.check_output(
                ['nvidia-smi', '--query-gpu=index,memory.free', '--format=csv,noheader,nounits'],
                encoding='utf-8'
            )
        except Exception as e:
            LOG.warning(f'Get idle gpus failed: {e}, if you have no gpu-driver, ignor it.')
            return []
        lines = order_list.strip().split('\n')

        str_num = os.getenv('CUDA_VISIBLE_DEVICES', None)
        if str_num:
            sub_gpus = [int(x) for x in str_num.strip().split(',')]

        gpu_info = []
        for line in lines:
            index, memory_free = line.split(', ')
            if not str_num or int(index) in sub_gpus:
                gpu_info.append((int(index), int(memory_free)))
        gpu_info.sort(key=lambda x: x[1], reverse=True)
        LOG.info('Memory left:\n' + '\n'.join([f'{item[0]} GPU, left: {item[1]} MiB' for item in gpu_info]))
        return [info[0] for info in gpu_info]

Job

Bases: Job

通用任务调度执行类。 该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Parameters:

  • cmd (LazyLLMCMD) –

    要执行的命令对象。

  • launcher (Any) –

    启动器实例，用于实际任务调度执行。

  • sync (bool, default: True ) –

    是否为同步执行，默认为 True。

Source code in lazyllm/launcher/base.py 
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
    @final
    class Job(Job):
        """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
        def __init__(self, cmd, launcher, *, sync=True):
            super(__class__, self).__init__(cmd, launcher, sync=sync)

        def _wrap_cmd(self, cmd):
            if self._launcher.ngpus == 0:
                return cmd
            gpus = self._launcher._get_idle_gpus()
            if gpus and lazyllm.config['cuda_visible']:
                if self._launcher.ngpus is None:
                    empty_cmd = f'export CUDA_VISIBLE_DEVICES={gpus[0]} && '
                elif self._launcher.ngpus <= len(gpus):
                    empty_cmd = 'export CUDA_VISIBLE_DEVICES=' + \
                                ','.join([str(n) for n in gpus[:self._launcher.ngpus]]) + ' && '
                else:
                    error_info = (f'Not enough GPUs available. Requested {self._launcher.ngpus} GPUs, '
                                  f'but only {len(gpus)} are available.')
                    LOG.error(error_info)
                    raise error_info
            else:
                empty_cmd = ''
            return empty_cmd + cmd

        def stop(self):
            if self.ps:
                try:
                    parent = psutil.Process(self.ps.pid)
                    for child in parent.children(recursive=True):
                        child.kill()
                    parent.kill()
                except psutil.NoSuchProcess:
                    LOG.warning(f'Process with PID {self.ps.pid} does not exist.')
                except psutil.AccessDenied:
                    LOG.warning(f'Permission denied when trying to kill process with PID {self.ps.pid}.')
                except Exception as e:
                    LOG.warning(f'An error occurred: {e}')

        @property
        def status(self):
            return_code = self.ps.poll()
            if return_code is None: job_status = Status.Running
            elif return_code == 0: job_status = Status.Done
            else: job_status = Status.Failed
            return job_status

        def _get_jobid(self):
            self.jobid = self.ps.pid if self.ps else None

        def get_jobip(self):
            return '127.0.0.1'

        def wait(self):
            if self.ps:
                self.ps.wait()

lazyllm.launcher.RemoteLauncher

Bases: LazyLLMLaunchersBase

此类是 LazyLLMLaunchersBase 的一个子类，它充当了一个远程启动器的代理。它根据配置文件中的 lazyllm.config['launcher'] 条目动态地创建并返回一个对应的启动器实例(例如：SlurmLauncher 或 ScoLauncher)。

Parameters:

  • *args –

    位置参数，将传递给动态创建的启动器构造函数。

  • sync (bool) –

    是否同步执行作业。默认为 False。

  • **kwargs –

    关键字参数，将传递给动态创建的启动器构造函数。

注意事项
  • RemoteLauncher 不是一个直接的启动器，而是根据配置动态创建一个启动器。
  • 配置文件中的 lazyllm.config['launcher'] 指定一个存在于 lazyllm.launchers 模块中的启动器类名。该配置可通过设置环境变量 LAZYLLM_DEFAULT_LAUNCHER 来设置。如：export LAZYLLM_DEFAULT_LAUNCHER=sco , export LAZYLLM_DEFAULT_LAUNCHER=slurm 。

Examples:

>>> import lazyllm
>>> launcher = lazyllm.launchers.remote(ngpus=1)
Source code in lazyllm/launcher/base.py 
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
class RemoteLauncher(LazyLLMLaunchersBase):
    """此类是 ``LazyLLMLaunchersBase`` 的一个子类，它充当了一个远程启动器的代理。它根据配置文件中的 ``lazyllm.config['launcher']`` 条目动态地创建并返回一个对应的启动器实例(例如：``SlurmLauncher`` 或 ``ScoLauncher``)。

Args:
    *args: 位置参数，将传递给动态创建的启动器构造函数。
    sync (bool): 是否同步执行作业。默认为 ``False``。
    **kwargs: 关键字参数，将传递给动态创建的启动器构造函数。

注意事项: 
    - ``RemoteLauncher`` 不是一个直接的启动器，而是根据配置动态创建一个启动器。 
    - 配置文件中的 ``lazyllm.config['launcher']`` 指定一个存在于 ``lazyllm.launchers`` 模块中的启动器类名。该配置可通过设置环境变量 ``LAZYLLM_DEFAULT_LAUNCHER`` 来设置。如：``export LAZYLLM_DEFAULT_LAUNCHER=sco`` , ``export LAZYLLM_DEFAULT_LAUNCHER=slurm`` 。


Examples:
    >>> import lazyllm
    >>> launcher = lazyllm.launchers.remote(ngpus=1)
    """
    def __new__(cls, *args, sync=False, ngpus=1, **kwargs):
        return getattr(lazyllm.launchers, lazyllm.config['launcher'])(*args, sync=sync, ngpus=ngpus, **kwargs)

lazyllm.launcher.SlurmLauncher

Bases: LazyLLMLaunchersBase

此类是 LazyLLMLaunchersBase 的子类，作为 Slurm 启动器。

具体而言，它提供了启动和配置 Slurm 作业的方法，包括指定分区、节点数量、进程数量、GPU 数量以及超时时间等参数。

Parameters:

  • partition (str, default: None ) –

    要使用的 Slurm 分区。默认为 None，此时将使用 lazyllm.config['partition'] 中的默认分区。 该配置可通过设置环境变量来生效，如 export LAZYLLM_SLURM_PART=a100。

  • nnode (int, default: 1 ) –

    要使用的节点数量。默认为 1。

  • nproc (int, default: 1 ) –

    每个节点要使用的进程数量。默认为 1。

  • ngpus (int, default: None ) –

    每个节点要使用的 GPU 数量。默认为 None，即不使用 GPU。

  • timeout (int, default: None ) –

    作业的超时时间（以秒为单位）。默认为 None，此时将不设置超时时间。

  • sync (bool, default: True ) –

    是否同步执行作业。默认为 True，否则为异步执行。

  • **kwargs –

    额外参数，其中支持： - num_can_use_nodes (int): 可使用的最大节点数。默认为 5。

Examples:

>>> import lazyllm
>>> launcher = lazyllm.launchers.slurm(partition='partition_name', nnode=1, nproc=1, ngpus=1, sync=False)
Source code in lazyllm/launcher/slurm.py 
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
@final
class SlurmLauncher(LazyLLMLaunchersBase):
    """此类是 ``LazyLLMLaunchersBase`` 的子类，作为 Slurm 启动器。

具体而言，它提供了启动和配置 Slurm 作业的方法，包括指定分区、节点数量、进程数量、GPU 数量以及超时时间等参数。

Args:
    partition (str): 要使用的 Slurm 分区。默认为 ``None``，此时将使用 ``lazyllm.config['partition']`` 中的默认分区。
                     该配置可通过设置环境变量来生效，如 ``export LAZYLLM_SLURM_PART=a100``。
    nnode (int): 要使用的节点数量。默认为 ``1``。
    nproc (int): 每个节点要使用的进程数量。默认为 ``1``。
    ngpus (int): 每个节点要使用的 GPU 数量。默认为 ``None``，即不使用 GPU。
    timeout (int): 作业的超时时间（以秒为单位）。默认为 ``None``，此时将不设置超时时间。
    sync (bool): 是否同步执行作业。默认为 ``True``，否则为异步执行。
    **kwargs: 额外参数，其中支持：
        - num_can_use_nodes (int): 可使用的最大节点数。默认为 ``5``。


Examples:
    >>> import lazyllm
    >>> launcher = lazyllm.launchers.slurm(partition='partition_name', nnode=1, nproc=1, ngpus=1, sync=False)
    """
    # In order to obtain the jobid to monitor and terminate the job more
    # conveniently, only one srun command is allowed in one Job
    all_processes = defaultdict(list)
    count = 0

    @final
    class Job(Job):
        """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
        def __init__(self, cmd, launcher, *, sync=True, **kw):
            super(__class__, self).__init__(cmd, launcher, sync=sync)
            self.name = self._generate_name()

        def _wrap_cmd(self, cmd):
            # Assemble the order
            slurm_cmd = f'srun -p {self._launcher.partition} -N {self._launcher.nnode} --job-name={self.name}'
            if self._launcher.nproc:
                slurm_cmd += f' -n{self._launcher.nproc}'
            if self._launcher.timeout:
                slurm_cmd += f' -t {self._launcher.timeout}'
            if self._launcher.ngpus:
                slurm_cmd += f' --gres=gpu:{self._launcher.ngpus}'
            return f'{slurm_cmd} bash -c "{cmd}"'

        def _get_jobid(self):
            time.sleep(0.5)  # Wait for cmd to be stably submitted to slurm
            id_str = subprocess.check_output(['squeue', '--name=' + self.name, '--noheader'])
            if id_str:
                id_list = id_str.decode().strip().split()
                self.jobid = id_list[0]

        def get_jobip(self):
            id_str = subprocess.check_output(['squeue', '--name=' + self.name, '--noheader'])
            id_list = id_str.decode().strip().split()
            self.ip = id_list[10]
            return self.ip

        def stop(self):
            if self.jobid:
                cmd = f'scancel --quiet {self.jobid}'
                subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
                                 encoding='utf-8', executable='/bin/bash')
                self.jobid = None

            if self.ps:
                self.ps.terminate()
                self.queue = Queue()
                self.output_thread_event.set()
                self.output_thread.join()

        def wait(self):
            if self.ps:
                self.ps.wait()

        @property
        def status(self):
            # lookup job
            if self.jobid:
                jobinfo = subprocess.check_output(['scontrol', 'show', 'job', str(self.jobid)])
                job_state = None
                job_state = None
                for line in jobinfo.decode().split('\n'):
                    if 'JobState' in line:
                        job_state = line.strip().split()[0].split('=')[1].strip().lower()
                        if job_state == 'running':
                            return Status.Running
                        elif job_state == 'tbsubmitted':
                            return Status.TBSubmitted
                        elif job_state == 'inqueue':
                            return Status.InQueue
                        elif job_state == 'pending':
                            return Status.Pending
                        elif job_state == 'done':
                            return Status.Done
                        elif job_state == 'cancelled':
                            return Status.Cancelled
                        else:
                            return Status.Failed
            else:
                return Status.Failed

    # TODO(wangzhihong): support configs; None -> lookup config
    def __init__(self, partition=None, nnode=1, nproc=1, ngpus=None, timeout=None, *, sync=True, **kwargs):
        super(__class__, self).__init__()
        # TODO: global config
        self.partition = partition if partition else lazyllm.config['partition']
        self.nnode, self.nproc, self.ngpus, self.timeout = nnode, nproc, ngpus, timeout
        self.sync = sync
        self.num_can_use_nodes = kwargs.get('num_can_use_nodes', 5)

    def makejob(self, cmd):
        """创建并返回一个 SlurmLauncher.Job 对象。

Args:
    cmd: 要执行的命令字符串。

**Returns:**

- SlurmLauncher.Job: 配置好的 Slurm 作业对象。
"""
        return SlurmLauncher.Job(cmd, launcher=self, sync=self.sync)

    def _add_dict(self, node_ip, used_gpus, node_dict):
        if node_ip not in node_dict:
            node_dict[node_ip] = 8 - used_gpus
        else:
            node_dict[node_ip] -= used_gpus

    def _expand_nodelist(self, nodes_str):
        pattern = r'\[(.*?)\]'
        matches = re.search(pattern, nodes_str)
        result = []
        if matches:
            nums = matches.group(1).split(',')
            base = nodes_str.split('[')[0]
            result = [base + str(x) for x in nums]
        return result

    def get_idle_nodes(self, partion=None):
        """
        Obtain the current number of available nodes based on the available number of GPUs.
        Return a dictionary with node IP as the key and the number of available GPUs as the value.
        """
        if not partion:
            partion = self.partition
        num_can_use_nodes = self.num_can_use_nodes

        # Query the number of available GPUs for applied nodes
        nodesinfo = subprocess.check_output(['squeue', '-p', partion, '--noheader'])
        node_dict = dict()

        for line in nodesinfo.decode().split('\n'):
            if 'gpu:' in line:
                node_info = line.strip().split()
                num_nodes = int(node_info[-3])
                num_gpus = int(node_info[-2].split(':')[-1])
                node_list = node_info[-1]
                if num_nodes == 1:
                    self._add_dict(node_list, num_gpus, node_dict)
                else:
                    avg_gpus = int(num_gpus / num_nodes)
                    result = self._expand_nodelist(node_list)
                    for x in result:
                        self._add_dict(x, avg_gpus, node_dict)

        # Obtain all available idle nodes in the specified partition
        idle_nodes = []
        nodesinfo = subprocess.check_output(['sinfo', '-p', partion, '--noheader'])
        for line in nodesinfo.decode().split('\n'):
            if 'idle' in line:
                node_info = line.strip().split()
                num_nodes = int(node_info[-3])
                node_list = node_info[-1]
                if num_nodes == 1:
                    idle_nodes.append(node_list)
                else:
                    idle_nodes += self._expand_nodelist(node_list)

        # Add idle nodes under resource constraints
        num_allocated_nodes = len(node_dict)
        num_append_nodes = num_can_use_nodes - num_allocated_nodes

        for i, node_ip in enumerate(idle_nodes):
            if i + 1 <= num_append_nodes:
                node_dict[node_ip] = 8

        # Remove nodes with depleted GPUs
        node_dict = {k: v for k, v in node_dict.items() if v != 0}
        return node_dict

    def launch(self, job) -> None:
        """启动 Slurm 作业并管理其执行。

该方法启动指定的 Slurm 作业，并根据同步设置决定是否等待作业完成。如果设置为同步执行，会持续监控作业状态直到完成，然后停止作业。

Args:
    job: 要启动的 SlurmLauncher.Job 对象。

**Returns:**

- 作业的返回值。

Raises:
    AssertionError: 如果传入的 job 不是 SlurmLauncher.Job 类型。
"""
        assert isinstance(job, SlurmLauncher.Job), 'Slurm launcher only support cmd'
        job.start()
        if self.sync:
            while job.status == Status.Running:
                time.sleep(10)
            job.stop()
        return job.return_value

makejob(cmd)

创建并返回一个 SlurmLauncher.Job 对象。

Parameters:

  • cmd –

    要执行的命令字符串。

Returns:

  • SlurmLauncher.Job: 配置好的 Slurm 作业对象。
Source code in lazyllm/launcher/slurm.py 
133
134
135
136
137
138
139
140
141
142
143
    def makejob(self, cmd):
        """创建并返回一个 SlurmLauncher.Job 对象。

Args:
    cmd: 要执行的命令字符串。

**Returns:**

- SlurmLauncher.Job: 配置好的 Slurm 作业对象。
"""
        return SlurmLauncher.Job(cmd, launcher=self, sync=self.sync)

get_idle_nodes(partion=None)

Obtain the current number of available nodes based on the available number of GPUs. Return a dictionary with node IP as the key and the number of available GPUs as the value.

Source code in lazyllm/launcher/slurm.py 
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
def get_idle_nodes(self, partion=None):
    """
    Obtain the current number of available nodes based on the available number of GPUs.
    Return a dictionary with node IP as the key and the number of available GPUs as the value.
    """
    if not partion:
        partion = self.partition
    num_can_use_nodes = self.num_can_use_nodes

    # Query the number of available GPUs for applied nodes
    nodesinfo = subprocess.check_output(['squeue', '-p', partion, '--noheader'])
    node_dict = dict()

    for line in nodesinfo.decode().split('\n'):
        if 'gpu:' in line:
            node_info = line.strip().split()
            num_nodes = int(node_info[-3])
            num_gpus = int(node_info[-2].split(':')[-1])
            node_list = node_info[-1]
            if num_nodes == 1:
                self._add_dict(node_list, num_gpus, node_dict)
            else:
                avg_gpus = int(num_gpus / num_nodes)
                result = self._expand_nodelist(node_list)
                for x in result:
                    self._add_dict(x, avg_gpus, node_dict)

    # Obtain all available idle nodes in the specified partition
    idle_nodes = []
    nodesinfo = subprocess.check_output(['sinfo', '-p', partion, '--noheader'])
    for line in nodesinfo.decode().split('\n'):
        if 'idle' in line:
            node_info = line.strip().split()
            num_nodes = int(node_info[-3])
            node_list = node_info[-1]
            if num_nodes == 1:
                idle_nodes.append(node_list)
            else:
                idle_nodes += self._expand_nodelist(node_list)

    # Add idle nodes under resource constraints
    num_allocated_nodes = len(node_dict)
    num_append_nodes = num_can_use_nodes - num_allocated_nodes

    for i, node_ip in enumerate(idle_nodes):
        if i + 1 <= num_append_nodes:
            node_dict[node_ip] = 8

    # Remove nodes with depleted GPUs
    node_dict = {k: v for k, v in node_dict.items() if v != 0}
    return node_dict

launch(job)

启动 Slurm 作业并管理其执行。

该方法启动指定的 Slurm 作业，并根据同步设置决定是否等待作业完成。如果设置为同步执行，会持续监控作业状态直到完成，然后停止作业。

Parameters:

  • job –

    要启动的 SlurmLauncher.Job 对象。

Returns:

  • 作业的返回值。

Raises:

  • AssertionError –

    如果传入的 job 不是 SlurmLauncher.Job 类型。

Source code in lazyllm/launcher/slurm.py 
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
    def launch(self, job) -> None:
        """启动 Slurm 作业并管理其执行。

该方法启动指定的 Slurm 作业，并根据同步设置决定是否等待作业完成。如果设置为同步执行，会持续监控作业状态直到完成，然后停止作业。

Args:
    job: 要启动的 SlurmLauncher.Job 对象。

**Returns:**

- 作业的返回值。

Raises:
    AssertionError: 如果传入的 job 不是 SlurmLauncher.Job 类型。
"""
        assert isinstance(job, SlurmLauncher.Job), 'Slurm launcher only support cmd'
        job.start()
        if self.sync:
            while job.status == Status.Running:
                time.sleep(10)
            job.stop()
        return job.return_value

lazyllm.launcher.ScoLauncher

Bases: LazyLLMLaunchersBase

此类是 LazyLLMLaunchersBase 的子类，作为SCO (Sensecore)启动器。

具体而言，它提供了启动和配置 SCO 作业的方法，包括指定分区、工作空间名称、框架类型、节点数量、进程数量、GPU 数量以及是否使用 torchrun 等参数。

Parameters:

  • partition (str, default: None ) –

    要使用的分区。默认为 None，此时将使用 lazyllm.config['partition'] 中的默认分区。该配置可通过设置环境变量来生效，如 export LAZYLLM_SLURM_PART=a100 。

  • workspace_name (str, default: config['sco.workspace'] ) –

    SCO 上的工作空间名称。默认为 lazyllm.config['sco.workspace'] 中的配置。该配置可通过设置环境变量来生效，如 export LAZYLLM_SCO_WORKSPACE=myspace 。

  • framework (str, default: 'pt' ) –

    要使用的框架类型，例如 pt 代表 PyTorch。默认为 pt。

  • nnode (int, default: 1 ) –

    要使用的节点数量。默认为 1。

  • nproc (int, default: 1 ) –

    每个节点要使用的进程数量。默认为 1。

  • ngpus –

    (int): 每个节点要使用的 GPU 数量。默认为 1, 使用1块 GPU。

  • torchrun (bool, default: False ) –

    是否使用 torchrun 启动作业。默认为 False。

  • sync (bool, default: True ) –

    是否同步执行作业。默认为 True，否则为异步执行。

Examples:

>>> import lazyllm
>>> launcher = lazyllm.launchers.sco(partition='partition_name', nnode=1, nproc=1, ngpus=1, sync=False)
Source code in lazyllm/launcher/sco.py 
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
@final
class ScoLauncher(LazyLLMLaunchersBase):
    """此类是 ``LazyLLMLaunchersBase`` 的子类，作为SCO (Sensecore)启动器。

具体而言，它提供了启动和配置 SCO 作业的方法，包括指定分区、工作空间名称、框架类型、节点数量、进程数量、GPU 数量以及是否使用 torchrun 等参数。

Args:
    partition (str): 要使用的分区。默认为 ``None``，此时将使用 ``lazyllm.config['partition']`` 中的默认分区。该配置可通过设置环境变量来生效，如 ``export LAZYLLM_SLURM_PART=a100`` 。
    workspace_name (str): SCO 上的工作空间名称。默认为 ``lazyllm.config['sco.workspace']`` 中的配置。该配置可通过设置环境变量来生效，如 ``export LAZYLLM_SCO_WORKSPACE=myspace`` 。
    framework (str): 要使用的框架类型，例如 ``pt`` 代表 PyTorch。默认为 ``pt``。
    nnode  (int): 要使用的节点数量。默认为 ``1``。
    nproc (int): 每个节点要使用的进程数量。默认为 ``1``。
    ngpus: (int): 每个节点要使用的 GPU 数量。默认为 ``1``, 使用1块 GPU。
    torchrun (bool): 是否使用 ``torchrun`` 启动作业。默认为 ``False``。
    sync (bool): 是否同步执行作业。默认为 ``True``，否则为异步执行。


Examples:
    >>> import lazyllm
    >>> launcher = lazyllm.launchers.sco(partition='partition_name', nnode=1, nproc=1, ngpus=1, sync=False)
    """
    all_processes = defaultdict(list)

    @final
    class Job(Job):
        """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
        def __init__(self, cmd, launcher, *, sync=True):
            super(__class__, self).__init__(cmd, launcher, sync=sync)
            # SCO job name must start with a letter
            self.name = 's_flag_' + self._generate_name()
            self.workspace_name = launcher.workspace_name
            self.torchrun = launcher.torchrun
            self.output_hooks = [self.output_hook]

        def output_hook(self, line):
            if not self.ip and 'LAZYLLMIP' in line:
                self.ip = line.split()[-1]

        def _wrap_cmd(self, cmd):
            launcher = self._launcher
            # Assemble the cmd
            sco_cmd = f'srun -p {launcher.partition} --workspace-id {self.workspace_name} ' \
                      f'--job-name={self.name} -f {launcher.framework} ' \
                      f'-r {lazyllm.config["sco_resource_type"]}.{launcher.ngpus} ' \
                      f'-N {launcher.nnode} --priority normal '

            torchrun_cmd = f'python -m torch.distributed.run --nproc_per_node {launcher.nproc} '

            if launcher.nnode == 1:
                # SCO for mpi：supports multiple cards in a single machine
                torchrun_cmd += f'--nnodes {launcher.nnode} --node_rank 0 '
            else:
                # SCO for All Reduce-DDP: support multiple machines and multiple cards
                torchrun_cmd += '--nnodes ${WORLD_SIZE} --node_rank ${RANK} ' \
                                '--master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} '
            pythonpath = os.getenv('PYTHONPATH', '')
            precmd = (f'''export PYTHONPATH={os.getcwd()}:{pythonpath}:$PYTHONPATH '''
                      f'''&& export PATH={os.path.join(os.path.expanduser('~'), '.local/bin')}:$PATH && ''')
            if lazyllm.config['sco_env_name']:
                precmd = f'source activate {lazyllm.config["sco_env_name"]} && ' + precmd
            env_vars = os.environ
            lazyllm_vars = {k: v for k, v in env_vars.items() if k.startswith('LAZYLLM')}
            if lazyllm_vars:
                precmd += ' && '.join(f'export {k}={v}' for k, v in lazyllm_vars.items()) + ' && '
            # For SCO: bash -c 'ifconfig | grep "inet " | awk "{printf \"LAZYLLMIP %s\\n\", \$2}"'
            precmd += '''ifconfig | grep "inet " | awk "{printf \\"LAZYLLMIP %s\\\\n\\", \$2}" &&'''  # noqa W605

            # Delete 'python' in cmd
            if self.torchrun and cmd.strip().startswith('python'):
                cmd = cmd.strip()[6:]
            return f'{sco_cmd} \'{precmd} {torchrun_cmd if self.torchrun else ""} {cmd}\''

        def _get_jobid(self):
            for i in range(5):
                time.sleep(2)  # Wait for cmd to be stably submitted to sco
                try:
                    id_str = subprocess.check_output([
                        'squeue', f'--workspace-id={self.workspace_name}',
                        '-o', 'jobname,jobid']).decode('utf-8')
                except Exception:
                    LOG.warning(f'Failed to capture job_id, retry the {i}-th time.')
                    continue
                pattern = re.compile(rf'{re.escape(self.name)}\s+(\S+)')
                match = pattern.search(id_str)
                if match:
                    self.jobid = match.group(1).strip()
                    break
                else:
                    LOG.warning(f'Failed to capture job_id, retry the {i}-th time.')

        def get_jobip(self):
            if self.ip:
                return self.ip
            else:
                raise RuntimeError('Cannot get IP.', f'JobID: {self.jobid}')

        def _scancel_job(self, cmd, max_retries=3):
            retries = 0
            while retries < max_retries:
                if self.status in (Status.Failed, Status.Cancelled, Status.Done):
                    break
                ps = subprocess.Popen(
                    cmd, shell=True, stdout=subprocess.PIPE,
                    stderr=subprocess.STDOUT,
                    encoding='utf-8', executable='/bin/bash')
                try:
                    stdout, stderr = ps.communicate(timeout=3)
                    if stdout:
                        LOG.info(stdout)
                        if 'success scancel' in stdout:
                            break
                    if stderr:
                        LOG.error(stderr)
                except subprocess.TimeoutExpired:
                    ps.kill()
                    LOG.warning(f'Command timed out, retrying... (Attempt {retries + 1}/{max_retries})')
                except Exception as e:
                    LOG.error('Try to scancel, but meet: ', e)
                retries += 1
                time.sleep(0.5)
            if retries == max_retries:
                LOG.error(f'Command failed after {max_retries} attempts.')

        def stop(self):
            if self.jobid:
                cmd = f'scancel --workspace-id={self.workspace_name} {self.jobid}'
                if lazyllm.config['sco_keep_record']:
                    LOG.warning(
                        f'`sco_keep_record` is on, not executing scancel. '
                        f'You can now check the logs on the web. '
                        f'To delete by terminal, you can execute: `{cmd}`'
                    )
                else:
                    self._scancel_job(cmd)
                    time.sleep(0.5)  # Avoid the execution of scancel and scontrol too close together.

            n = 0
            while self.status not in (Status.Done, Status.Cancelled, Status.Failed):
                time.sleep(1)
                n += 1
                if n > 25:
                    break

            if self.ps:
                self.ps.terminate()
                self.queue = Queue()
                self.output_thread_event.set()
                self.output_thread.join()

            self.jobid = None

        def wait(self):
            if self.ps:
                self.ps.wait()

        @property
        def status(self):
            if self.jobid:
                try:
                    id_str = subprocess.check_output(['scontrol', f'--workspace-id={self.workspace_name}',
                                                      'show', 'job', str(self.jobid)]).decode('utf-8')
                    id_json = json.loads(id_str)
                    job_state = id_json['state'].strip().lower()
                    if job_state == 'running':
                        return Status.Running
                    elif job_state in ['tbsubmitted', 'suspending']:
                        return Status.TBSubmitted
                    elif job_state in ['waiting', 'init', 'queueing', 'creating',
                                       'restarting', 'recovering', 'starting']:
                        return Status.InQueue
                    elif job_state in ['suspended']:
                        return Status.Cancelled
                    elif job_state == 'succeeded':
                        return Status.Done
                except Exception as e:
                    lazyllm.LOG.error(f'Failed to get job status, reason is {str(e)}')
            return Status.Failed

    def __init__(self, partition=None, workspace_name=lazyllm.config['sco.workspace'],
                 framework='pt', nnode=1, nproc=1, ngpus=1, torchrun=False, sync=True, **kwargs):
        assert nnode >= 1, 'Use at least one node.'
        assert nproc >= 1, 'Start at least one process.'
        assert type(workspace_name) is str, f'"workspace_name" is {workspace_name}. Please set workspace_name.'
        self.partition = partition if partition else lazyllm.config['partition']
        self.workspace_name = workspace_name
        self.framework = framework
        self.nnode = nnode
        self.nproc = nproc
        self.ngpus = ngpus or 1
        self.torchrun = torchrun
        self.sync = sync
        super(__class__, self).__init__()

    def makejob(self, cmd):
        return ScoLauncher.Job(cmd, launcher=self, sync=self.sync)

    def launch(self, job) -> None:
        assert isinstance(job, ScoLauncher.Job), 'Sco launcher only support cmd'
        job.start()
        if self.sync:
            while job.status == Status.Running:
                time.sleep(10)
            job.stop()
        return job.return_value

Job

Bases: Job

通用任务调度执行类。 该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Parameters:

  • cmd (LazyLLMCMD) –

    要执行的命令对象。

  • launcher (Any) –

    启动器实例，用于实际任务调度执行。

  • sync (bool, default: True ) –

    是否为同步执行，默认为 True。

Source code in lazyllm/launcher/sco.py 
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
    @final
    class Job(Job):
        """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
        def __init__(self, cmd, launcher, *, sync=True):
            super(__class__, self).__init__(cmd, launcher, sync=sync)
            # SCO job name must start with a letter
            self.name = 's_flag_' + self._generate_name()
            self.workspace_name = launcher.workspace_name
            self.torchrun = launcher.torchrun
            self.output_hooks = [self.output_hook]

        def output_hook(self, line):
            if not self.ip and 'LAZYLLMIP' in line:
                self.ip = line.split()[-1]

        def _wrap_cmd(self, cmd):
            launcher = self._launcher
            # Assemble the cmd
            sco_cmd = f'srun -p {launcher.partition} --workspace-id {self.workspace_name} ' \
                      f'--job-name={self.name} -f {launcher.framework} ' \
                      f'-r {lazyllm.config["sco_resource_type"]}.{launcher.ngpus} ' \
                      f'-N {launcher.nnode} --priority normal '

            torchrun_cmd = f'python -m torch.distributed.run --nproc_per_node {launcher.nproc} '

            if launcher.nnode == 1:
                # SCO for mpi：supports multiple cards in a single machine
                torchrun_cmd += f'--nnodes {launcher.nnode} --node_rank 0 '
            else:
                # SCO for All Reduce-DDP: support multiple machines and multiple cards
                torchrun_cmd += '--nnodes ${WORLD_SIZE} --node_rank ${RANK} ' \
                                '--master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} '
            pythonpath = os.getenv('PYTHONPATH', '')
            precmd = (f'''export PYTHONPATH={os.getcwd()}:{pythonpath}:$PYTHONPATH '''
                      f'''&& export PATH={os.path.join(os.path.expanduser('~'), '.local/bin')}:$PATH && ''')
            if lazyllm.config['sco_env_name']:
                precmd = f'source activate {lazyllm.config["sco_env_name"]} && ' + precmd
            env_vars = os.environ
            lazyllm_vars = {k: v for k, v in env_vars.items() if k.startswith('LAZYLLM')}
            if lazyllm_vars:
                precmd += ' && '.join(f'export {k}={v}' for k, v in lazyllm_vars.items()) + ' && '
            # For SCO: bash -c 'ifconfig | grep "inet " | awk "{printf \"LAZYLLMIP %s\\n\", \$2}"'
            precmd += '''ifconfig | grep "inet " | awk "{printf \\"LAZYLLMIP %s\\\\n\\", \$2}" &&'''  # noqa W605

            # Delete 'python' in cmd
            if self.torchrun and cmd.strip().startswith('python'):
                cmd = cmd.strip()[6:]
            return f'{sco_cmd} \'{precmd} {torchrun_cmd if self.torchrun else ""} {cmd}\''

        def _get_jobid(self):
            for i in range(5):
                time.sleep(2)  # Wait for cmd to be stably submitted to sco
                try:
                    id_str = subprocess.check_output([
                        'squeue', f'--workspace-id={self.workspace_name}',
                        '-o', 'jobname,jobid']).decode('utf-8')
                except Exception:
                    LOG.warning(f'Failed to capture job_id, retry the {i}-th time.')
                    continue
                pattern = re.compile(rf'{re.escape(self.name)}\s+(\S+)')
                match = pattern.search(id_str)
                if match:
                    self.jobid = match.group(1).strip()
                    break
                else:
                    LOG.warning(f'Failed to capture job_id, retry the {i}-th time.')

        def get_jobip(self):
            if self.ip:
                return self.ip
            else:
                raise RuntimeError('Cannot get IP.', f'JobID: {self.jobid}')

        def _scancel_job(self, cmd, max_retries=3):
            retries = 0
            while retries < max_retries:
                if self.status in (Status.Failed, Status.Cancelled, Status.Done):
                    break
                ps = subprocess.Popen(
                    cmd, shell=True, stdout=subprocess.PIPE,
                    stderr=subprocess.STDOUT,
                    encoding='utf-8', executable='/bin/bash')
                try:
                    stdout, stderr = ps.communicate(timeout=3)
                    if stdout:
                        LOG.info(stdout)
                        if 'success scancel' in stdout:
                            break
                    if stderr:
                        LOG.error(stderr)
                except subprocess.TimeoutExpired:
                    ps.kill()
                    LOG.warning(f'Command timed out, retrying... (Attempt {retries + 1}/{max_retries})')
                except Exception as e:
                    LOG.error('Try to scancel, but meet: ', e)
                retries += 1
                time.sleep(0.5)
            if retries == max_retries:
                LOG.error(f'Command failed after {max_retries} attempts.')

        def stop(self):
            if self.jobid:
                cmd = f'scancel --workspace-id={self.workspace_name} {self.jobid}'
                if lazyllm.config['sco_keep_record']:
                    LOG.warning(
                        f'`sco_keep_record` is on, not executing scancel. '
                        f'You can now check the logs on the web. '
                        f'To delete by terminal, you can execute: `{cmd}`'
                    )
                else:
                    self._scancel_job(cmd)
                    time.sleep(0.5)  # Avoid the execution of scancel and scontrol too close together.

            n = 0
            while self.status not in (Status.Done, Status.Cancelled, Status.Failed):
                time.sleep(1)
                n += 1
                if n > 25:
                    break

            if self.ps:
                self.ps.terminate()
                self.queue = Queue()
                self.output_thread_event.set()
                self.output_thread.join()

            self.jobid = None

        def wait(self):
            if self.ps:
                self.ps.wait()

        @property
        def status(self):
            if self.jobid:
                try:
                    id_str = subprocess.check_output(['scontrol', f'--workspace-id={self.workspace_name}',
                                                      'show', 'job', str(self.jobid)]).decode('utf-8')
                    id_json = json.loads(id_str)
                    job_state = id_json['state'].strip().lower()
                    if job_state == 'running':
                        return Status.Running
                    elif job_state in ['tbsubmitted', 'suspending']:
                        return Status.TBSubmitted
                    elif job_state in ['waiting', 'init', 'queueing', 'creating',
                                       'restarting', 'recovering', 'starting']:
                        return Status.InQueue
                    elif job_state in ['suspended']:
                        return Status.Cancelled
                    elif job_state == 'succeeded':
                        return Status.Done
                except Exception as e:
                    lazyllm.LOG.error(f'Failed to get job status, reason is {str(e)}')
            return Status.Failed

lazyllm.launcher.Job

Bases: object

通用任务调度执行类。 该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Parameters:

  • cmd (LazyLLMCMD) –

    要执行的命令对象。

  • launcher (Any) –

    启动器实例，用于实际任务调度执行。

  • sync (bool, default: True ) –

    是否为同步执行，默认为 True。

Source code in lazyllm/launcher/base.py 
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
class Job(object):
    """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
    def __init__(self, cmd, launcher, *, sync=True):
        assert isinstance(cmd, LazyLLMCMD)
        self._origin_cmd = cmd
        self.sync = sync
        self._launcher = launcher
        self.queue, self.jobid, self.ip, self.ps = Queue(), None, None, None
        self.output_hooks = []

    def _set_return_value(self):
        cmd = getattr(self, '_fixed_cmd', None)
        if cmd and callable(cmd.return_value):
            self.return_value = cmd.return_value(self)
        elif cmd and cmd.return_value:
            self.return_value = cmd.return_value
        else:
            self.return_value = self

    def get_executable_cmd(self, *, fixed=False):
        """生成最终可执行命令。
如果已缓存固定命令（fixed），则直接返回。否则根据原始命令进行包裹（wrap）并缓存为 `_fixed_cmd`。

Args:
    fixed (bool): 是否使用已固定的命令对象（若已存在）。

**Returns:**

- LazyLLMCMD: 可直接执行的命令对象。
"""
        if fixed and hasattr(self, '_fixed_cmd'):
            LOG.info('Command is fixed!')
            return self._fixed_cmd
        cmd = self._origin_cmd
        if callable(cmd.cmd):
            cmd = cmd.with_cmd(cmd.cmd())
        self._fixed_cmd = cmd.with_cmd(self._wrap_cmd(cmd.cmd))
        return self._fixed_cmd

    # interfaces
    def stop(self):
        """停止当前作业。
该方法为接口定义，需子类实现，当前抛出 NotImplementedError。
"""
        raise NotImplementedError
    @property
    def status(self):
        """当前作业状态。
该属性为接口定义，需子类实现，当前抛出 NotImplementedError。
"""
        raise NotImplementedError
    def wait(self):
        """挂起当前线程，等待作业执行完成。当前实现为空方法（子类可重写）。
"""
        pass
    def _wrap_cmd(self, cmd): return cmd

    def _start(self, *, fixed):
        cmd = self.get_executable_cmd(fixed=fixed)
        LOG.info(f'Command: {cmd}')
        if lazyllm.config['mode'] == lazyllm.Mode.Display: return
        self.ps = subprocess.Popen(cmd.cmd, shell=True, stdout=subprocess.PIPE,
                                   stderr=subprocess.STDOUT)
        self._get_jobid()
        self._enqueue_subprocess_output(hooks=self.output_hooks)

        if self.sync:
            self.ps.wait()
        else:
            self._launcher.all_processes[self._launcher._id].append((self.jobid, self))
            n = 0
            while self.status in (Status.TBSubmitted, Status.InQueue, Status.Pending):
                time.sleep(2)
                n += 1
                if n > 1800:  # 3600s
                    self._launcher.all_processes[self._launcher._id].pop()
                    LOG.error('Launch failed: No computing resources are available.')
                    break

    def restart(self, *, fixed=False):
        """重新启动作业流程。
该函数会先停止已有进程，等待 2 秒后重新启动作业。

Args:
    fixed (bool): 是否使用固定后的命令。
"""
        self.stop()
        time.sleep(2)
        self._start(fixed=fixed)

    def start(self, *, restart=3, fixed=False):
        """对外接口：启动作业，并支持失败时的自动重试。
若作业执行失败，会根据 `restart` 参数控制重试次数。

Args:
    restart (int): 重试次数。默认为 3。
    fixed (bool): 是否使用固定后的命令。用于避免多次构建。
"""
        self._start(fixed=fixed)
        if not (lazyllm.config['mode'] == lazyllm.Mode.Display or self._fixed_cmd.checkf(self)):
            if restart > 0:
                for ii in range(restart):
                    LOG.warning(f'Job failed, restarting... ({ii + 1}/{restart})')
                    self.restart(fixed=fixed)
                    if self._fixed_cmd.checkf(self): break
                else:
                    detail = self.queue.get_recent(join='\n', join_prefix=' Last logs:\n')
                    raise RuntimeError(f'Job failed after retrying {restart} times.{detail}')
            else:
                detail = self.queue.get_recent(join='\n', join_prefix=' Last logs:\n')
                raise RuntimeError(f'Job failed without retries.{detail}')
        self._set_return_value()

    def _enqueue_subprocess_output(self, hooks=None):
        self.output_thread_event = threading.Event()

        def impl(out, queue):
            for line in iter(out.readline, b''):
                try:
                    line = line.decode('utf-8')
                except Exception:
                    try:
                        line = line.decode('gb2312')
                    except Exception:
                        pass
                if isinstance(line, str):
                    queue.put(line)
                    if hooks:
                        hooks(line) if callable(hooks) else [hook(line) for hook in hooks]
                LOG.info(f'{line.lstrip("INFO:").rstrip()}', jobid=self.jobid, name='launcher')
                if self.output_thread_event.is_set():
                    break
            out.close()
        self.output_thread = threading.Thread(target=impl, args=(self.ps.stdout, self.queue))
        self.output_thread.daemon = True
        self.output_thread.start()

    def _generate_name(self):
        now = datetime.now()
        return str(hex(hash(now.strftime('%S%M') + str(random.randint(3, 2000)))))[2:10]

    def __deepcopy__(self, memo=None):
        raise RuntimeError('Cannot copy Job object')

    @property
    def log_path(self):
        match = re.search(r'tee\s+([^\s]+\.log)', self._origin_cmd.cmd)
        if match:
            return match.group(1)
        return None

status property

当前作业状态。 该属性为接口定义，需子类实现，当前抛出 NotImplementedError。

get_executable_cmd(*, fixed=False)

生成最终可执行命令。 如果已缓存固定命令（fixed），则直接返回。否则根据原始命令进行包裹（wrap）并缓存为 _fixed_cmd。

Parameters:

  • fixed (bool, default: False ) –

    是否使用已固定的命令对象（若已存在）。

Returns:

  • LazyLLMCMD: 可直接执行的命令对象。
Source code in lazyllm/launcher/base.py 
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
    def get_executable_cmd(self, *, fixed=False):
        """生成最终可执行命令。
如果已缓存固定命令（fixed），则直接返回。否则根据原始命令进行包裹（wrap）并缓存为 `_fixed_cmd`。

Args:
    fixed (bool): 是否使用已固定的命令对象（若已存在）。

**Returns:**

- LazyLLMCMD: 可直接执行的命令对象。
"""
        if fixed and hasattr(self, '_fixed_cmd'):
            LOG.info('Command is fixed!')
            return self._fixed_cmd
        cmd = self._origin_cmd
        if callable(cmd.cmd):
            cmd = cmd.with_cmd(cmd.cmd())
        self._fixed_cmd = cmd.with_cmd(self._wrap_cmd(cmd.cmd))
        return self._fixed_cmd

restart(*, fixed=False)

重新启动作业流程。 该函数会先停止已有进程，等待 2 秒后重新启动作业。

Parameters:

  • fixed (bool, default: False ) –

    是否使用固定后的命令。

Source code in lazyllm/launcher/base.py 
213
214
215
216
217
218
219
220
221
222
    def restart(self, *, fixed=False):
        """重新启动作业流程。
该函数会先停止已有进程，等待 2 秒后重新启动作业。

Args:
    fixed (bool): 是否使用固定后的命令。
"""
        self.stop()
        time.sleep(2)
        self._start(fixed=fixed)

start(*, restart=3, fixed=False)

对外接口：启动作业，并支持失败时的自动重试。 若作业执行失败，会根据 restart 参数控制重试次数。

Parameters:

  • restart (int, default: 3 ) –

    重试次数。默认为 3。

  • fixed (bool, default: False ) –

    是否使用固定后的命令。用于避免多次构建。

Source code in lazyllm/launcher/base.py 
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
    def start(self, *, restart=3, fixed=False):
        """对外接口：启动作业，并支持失败时的自动重试。
若作业执行失败，会根据 `restart` 参数控制重试次数。

Args:
    restart (int): 重试次数。默认为 3。
    fixed (bool): 是否使用固定后的命令。用于避免多次构建。
"""
        self._start(fixed=fixed)
        if not (lazyllm.config['mode'] == lazyllm.Mode.Display or self._fixed_cmd.checkf(self)):
            if restart > 0:
                for ii in range(restart):
                    LOG.warning(f'Job failed, restarting... ({ii + 1}/{restart})')
                    self.restart(fixed=fixed)
                    if self._fixed_cmd.checkf(self): break
                else:
                    detail = self.queue.get_recent(join='\n', join_prefix=' Last logs:\n')
                    raise RuntimeError(f'Job failed after retrying {restart} times.{detail}')
            else:
                detail = self.queue.get_recent(join='\n', join_prefix=' Last logs:\n')
                raise RuntimeError(f'Job failed without retries.{detail}')
        self._set_return_value()

stop()

停止当前作业。 该方法为接口定义，需子类实现，当前抛出 NotImplementedError。

Source code in lazyllm/launcher/base.py 
174
175
176
177
178
    def stop(self):
        """停止当前作业。
该方法为接口定义，需子类实现，当前抛出 NotImplementedError。
"""
        raise NotImplementedError

wait()

挂起当前线程，等待作业执行完成。当前实现为空方法（子类可重写）。

Source code in lazyllm/launcher/base.py 
185
186
187
188
    def wait(self):
        """挂起当前线程，等待作业执行完成。当前实现为空方法（子类可重写）。
"""
        pass

lazyllm.launcher.K8sLauncher

Bases: LazyLLMLaunchersBase

K8sLauncher是一个基于Kubernetes的部署启动器，用于在Kubernetes集群中部署和管理服务。

Parameters:

  • kube_config_path (str, default: None ) –

    Kubernetes配置文件路径。

  • resource_config_path (str) –

    资源配置文件路径。

  • image (str, default: None ) –

    容器镜像。

  • volume_configs (list, default: None ) –

    卷配置列表。

  • svc_type (str, default: None ) –

    服务类型，默认为"LoadBalancer"。

  • namespace (str, default: None ) –

    Kubernetes命名空间，默认为"default"。

  • gateway_name (str, default: None ) –

    网关名称，默认为"lazyllm-gateway"。

  • gateway_class_name (str, default: None ) –

    网关类名称，默认为"istio"。

  • host (str, default: None ) –

    HTTP主机名，默认为None。

  • path (str, default: None ) –

    HTTP路径，默认为'/generate'。

  • gateway_retry (int) –

    网关重试次数。

Source code in lazyllm/launcher/k8s.py 
  24
  25
  26
  27
  28
  29
  30
  31
  32
  33
  34
  35
  36
  37
  38
  39
  40
  41
  42
  43
  44
  45
  46
  47
  48
  49
  50
  51
  52
  53
  54
  55
  56
  57
  58
  59
  60
  61
  62
  63
  64
  65
  66
  67
  68
  69
  70
  71
  72
  73
  74
  75
  76
  77
  78
  79
  80
  81
  82
  83
  84
  85
  86
  87
  88
  89
  90
  91
  92
  93
  94
  95
  96
  97
  98
  99
 100
 101
 102
 103
 104
 105
 106
 107
 108
 109
 110
 111
 112
 113
 114
 115
 116
 117
 118
 119
 120
 121
 122
 123
 124
 125
 126
 127
 128
 129
 130
 131
 132
 133
 134
 135
 136
 137
 138
 139
 140
 141
 142
 143
 144
 145
 146
 147
 148
 149
 150
 151
 152
 153
 154
 155
 156
 157
 158
 159
 160
 161
 162
 163
 164
 165
 166
 167
 168
 169
 170
 171
 172
 173
 174
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
@final
class K8sLauncher(LazyLLMLaunchersBase):
    """K8sLauncher是一个基于Kubernetes的部署启动器，用于在Kubernetes集群中部署和管理服务。

Args:
    kube_config_path (str): Kubernetes配置文件路径。
    resource_config_path (str): 资源配置文件路径。
    image (str): 容器镜像。
    volume_configs (list): 卷配置列表。
    svc_type (str): 服务类型，默认为"LoadBalancer"。
    namespace (str): Kubernetes命名空间，默认为"default"。
    gateway_name (str): 网关名称，默认为"lazyllm-gateway"。
    gateway_class_name (str): 网关类名称，默认为"istio"。
    host (str): HTTP主机名，默认为None。
    path (str): HTTP路径，默认为'/generate'。
    gateway_retry (int): 网关重试次数。
"""
    all_processes = defaultdict(list)
    namespace = 'default'

    class Job(Job):
        """通用任务调度执行类。
该类用于封装一个通过启动器（launcher）调度执行的任务，支持命令包装、同步控制、返回值提取、命令固定等功能。

Args:
    cmd (LazyLLMCMD): 要执行的命令对象。
    launcher (Any): 启动器实例，用于实际任务调度执行。
    sync (bool): 是否为同步执行，默认为 True。
"""
        def __init__(self, cmd, launcher, *, sync=True):
            super().__init__(cmd, launcher, sync=sync)
            self.launch_type = launcher.launch_type
            prefix = 'deployment' if self.launch_type == 'inference' else 'job'
            self.deployment_name = f'{prefix}-{uuid.uuid4().hex[:8]}'
            self.ngpus = launcher.ngpus
            self.namespace = launcher.namespace
            self.volume_configs = launcher.volume_configs
            self.gateway_name = launcher.gateway_name
            self.gateway_class_name = launcher.gateway_class_name
            self.deployment_port = 8080
            self.host = launcher.http_host
            self.path = launcher.http_path
            self.svc_type = launcher.svc_type
            self.gateway_retry = launcher.gateway_retry
            self.on_gateway = launcher.on_gateway
            self.image = launcher.image
            self.resource_config = launcher.resource_config if launcher.resource_config else {}

        def _wrap_cmd(self, cmd):
            pythonpath = os.getenv('PYTHONPATH', '')
            precmd = (f'''export PYTHONPATH={os.getcwd()}:{pythonpath}:$PYTHONPATH '''
                      f'''&& export PATH={os.path.join(os.path.expanduser('~'), '.local/bin')}:$PATH &&''')
            if lazyllm.config['k8s_env_name']:
                precmd = f'source activate {lazyllm.config["k8s_env_name"]} && ' + precmd
            env_vars = os.environ
            lazyllm_vars = {k: v for k, v in env_vars.items() if k.startswith('LAZYLLM')}
            if lazyllm_vars:
                precmd += ' && '.join(f'export {k}={v}' for k, v in lazyllm_vars.items()) + ' && '
            precmd += '''ifconfig | grep "inet " | awk "{printf \\"LAZYLLMIP %s\\\\n\\", \$2}" &&'''  # noqa W605
            if self.launch_type == 'inference':
                port_match = re.search(r'--(?:open_)?port=(\d+)', cmd)
                if port_match:
                    port = port_match.group(1)
                    LOG.info(f'Port: {port}')
                    self.deployment_port = int(port)
                else:
                    LOG.info('Port not found')
                    raise ValueError('Failed to obtain application port.')
            return precmd + ' ' + cmd

        def _create_container_and_volumes(self, cmd, volume_configs=None):
            device_type = lazyllm.config['k8s_device_type']
            resource_config = self.resource_config.get('requests', {'cpu': '2', 'memory': '16Gi'})
            if device_type:
                resource_config[device_type] = self.ngpus

            container = k8s.client.V1Container(
                name=self.deployment_name,
                image=self.image,
                image_pull_policy='IfNotPresent',
                command=['bash', '-c', cmd],
                resources=k8s.client.V1ResourceRequirements(
                    requests=resource_config,
                    limits=resource_config
                ),
                volume_mounts=[] if not volume_configs else [
                    k8s.client.V1VolumeMount(
                        mount_path=vol_config['mount_path'] if '__CURRENT_DIR__' not in vol_config['mount_path']
                        else vol_config['mount_path'].replace('__CURRENT_DIR__', os.getcwd()),
                        name=vol_config['name']
                    ) for vol_config in volume_configs
                ]
            )

            volumes = []
            if volume_configs:
                for vol_config in volume_configs:
                    if 'nfs_server' in vol_config and 'nfs_path' in vol_config:
                        volumes.append(
                            k8s.client.V1Volume(
                                name=vol_config['name'],
                                nfs=k8s.client.V1NFSVolumeSource(
                                    server=vol_config['nfs_server'],
                                    path=vol_config['nfs_path'] if '__CURRENT_DIR__' not in vol_config['nfs_path']
                                    else vol_config['nfs_path'].replace('__CURRENT_DIR__', os.getcwd()),
                                    read_only=vol_config.get('read_only', False)
                                )
                            )
                        )
                    elif 'host_path' in vol_config:
                        volumes.append(
                            k8s.client.V1Volume(
                                name=vol_config['name'],
                                host_path=k8s.client.V1HostPathVolumeSource(
                                    path=vol_config['host_path'] if '__CURRENT_DIR__' not in vol_config['host_path']
                                    else vol_config['host_path'].replace('__CURRENT_DIR__', os.getcwd()),
                                    type='Directory'
                                )
                            )
                        )
                    else:
                        LOG.error(f'{vol_config} configuration error.')
                        raise

            return container, volumes

        def _create_deployment_spec(self, cmd, volume_configs=None):
            container, volumes = self._create_container_and_volumes(cmd, volume_configs)

            template = k8s.client.V1PodTemplateSpec(
                metadata=k8s.client.V1ObjectMeta(labels={'app': self.deployment_name}),
                spec=k8s.client.V1PodSpec(restart_policy='Always', containers=[container], volumes=volumes)
            )
            deployment_spec = k8s.client.V1DeploymentSpec(
                replicas=1,
                template=template,
                selector=k8s.client.V1LabelSelector(match_labels={'app': self.deployment_name})
            )
            return k8s.client.V1Deployment(
                api_version='apps/v1',
                kind='Deployment',
                metadata=k8s.client.V1ObjectMeta(name=self.deployment_name),
                spec=deployment_spec
            )

        def _create_job_spec(self, cmd, volume_configs=None):
            container, volumes = self._create_container_and_volumes(cmd, volume_configs)

            # use OnFailure for job to avoid infinite restart
            template = k8s.client.V1PodTemplateSpec(
                metadata=k8s.client.V1ObjectMeta(labels={'app': self.deployment_name}),
                spec=k8s.client.V1PodSpec(restart_policy='OnFailure', containers=[container], volumes=volumes)
            )
            job_spec = k8s.client.V1JobSpec(
                template=template,
                backoff_limit=3
            )
            return k8s.client.V1Job(
                api_version='batch/v1',
                kind='Job',
                metadata=k8s.client.V1ObjectMeta(name=self.deployment_name),
                spec=job_spec
            )

        def _create_deployment(self, *, cmd):
            api_instance = k8s.client.AppsV1Api()
            deployment = self._create_deployment_spec(cmd.cmd, self.volume_configs)
            try:
                api_instance.create_namespaced_deployment(
                    body=deployment,
                    namespace=self.namespace
                )
                LOG.info(f'Kubernetes Deployment "{self.deployment_name}" created successfully.')
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when creating Kubernetes Deployment: {e}')
                raise

        def _create_job(self, *, cmd):
            api_instance = k8s.client.BatchV1Api()
            job = self._create_job_spec(cmd.cmd, self.volume_configs)
            try:
                api_instance.create_namespaced_job(
                    body=job,
                    namespace=self.namespace
                )
                LOG.info(f'Kubernetes Job "{self.deployment_name}" created successfully.')
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when creating Kubernetes Job: {e}')
                raise

        def _delete_deployment(self, wait_for_completion=True, timeout=60, check_interval=5):
            k8s.config.load_kube_config(self._launcher.kube_config_path)
            api_instance = k8s.client.AppsV1Api()
            try:
                api_instance.delete_namespaced_deployment(
                    name=self.deployment_name,
                    namespace=self.namespace,
                    body=k8s.client.V1DeleteOptions(propagation_policy='Foreground')
                )
                LOG.info(f'Kubernetes Deployment {self.deployment_name} deleted.')

                if wait_for_completion:
                    self._wait_for_deployment_deletion(timeout=timeout, check_interval=check_interval)
            except k8s.client.rest.ApiException as e:
                if e.status == 404:
                    LOG.info(f'Kubernetes Deployment "{self.deployment_name}" already deleted.')
                else:
                    LOG.error(f'Exception when deleting Kubernetes Deployment: {e}')
                    raise

        def _delete_job(self, wait_for_completion=True, timeout=60, check_interval=5):
            k8s.config.load_kube_config(self._launcher.kube_config_path)
            api_instance = k8s.client.BatchV1Api()
            try:
                api_instance.delete_namespaced_job(
                    name=self.deployment_name,
                    namespace=self.namespace,
                    body=k8s.client.V1DeleteOptions(propagation_policy='Foreground')
                )
                LOG.info(f'Kubernetes Job {self.deployment_name} deleted.')

                if wait_for_completion:
                    self._wait_for_job_deletion(timeout=timeout, check_interval=check_interval)
            except k8s.client.rest.ApiException as e:
                if e.status == 404:
                    LOG.info(f'Kubernetes Job "{self.deployment_name}" already deleted.')
                else:
                    LOG.error(f'Exception when deleting Kubernetes Job: {e}')
                    raise

        def _wait_for_deployment_deletion(self, timeout, check_interval):
            api_instance = k8s.client.AppsV1Api()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    api_instance.read_namespaced_deployment(name=self.deployment_name, namespace=self.namespace)
                    LOG.info(f'Waiting for Kubernetes Deployment "{self.deployment_name}" to be deleted...')
                except k8s.client.rest.ApiException as e:
                    if e.status == 404:
                        LOG.info(f'Kubernetes Deployment "{self.deployment_name}" successfully deleted.')
                        return
                    else:
                        LOG.error(f'Error while checking Deployment deletion status: {e}')
                        raise
                time.sleep(check_interval)
            LOG.warning(f'Timeout while waiting for Kubernetes Deployment "{self.deployment_name}" to be deleted.')

        def _wait_for_job_deletion(self, timeout, check_interval):
            api_instance = k8s.client.BatchV1Api()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    api_instance.read_namespaced_job(name=self.deployment_name, namespace=self.namespace)
                    LOG.info(f'Waiting for Kubernetes Job "{self.deployment_name}" to be deleted...')
                except k8s.client.rest.ApiException as e:
                    if e.status == 404:
                        LOG.info(f'Kubernetes Job "{self.deployment_name}" successfully deleted.')
                        return
                    else:
                        LOG.error(f'Error while checking Job deletion status: {e}')
                        raise
                time.sleep(check_interval)
            LOG.warning(f'Timeout while waiting for Kubernetes Job "{self.deployment_name}" to be deleted.')

        def _expose_deployment(self):
            api_instance = k8s.client.CoreV1Api()
            service = k8s.client.V1Service(
                api_version='v1',
                kind='Service',
                metadata=k8s.client.V1ObjectMeta(name=f'service-{self.deployment_name}'),
                spec=k8s.client.V1ServiceSpec(
                    selector={'app': self.deployment_name},
                    ports=[k8s.client.V1ServicePort(port=self.deployment_port, target_port=self.deployment_port)],
                    type='ClusterIP'
                )
            )
            try:
                api_instance.create_namespaced_service(
                    namespace=self.namespace,
                    body=service
                )
                LOG.info(f'Kubernetes Service "service-{self.deployment_name}" created and exposed successfully.')
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when creating Service: {e}')
                raise

        def _delete_service(self, wait_for_completion=True, timeout=60, check_interval=5):
            k8s.config.load_kube_config(self._launcher.kube_config_path)
            svc_instance = k8s.client.CoreV1Api()
            service_name = f'service-{self.deployment_name}'
            try:
                svc_instance.delete_namespaced_service(
                    name=service_name,
                    namespace=self.namespace,
                    body=k8s.client.V1DeleteOptions(propagation_policy='Foreground')
                )
                LOG.info(f'Kubernetes Service "{service_name}" deleted.')

                if wait_for_completion:
                    self._wait_for_service_deletion(service_name=service_name,
                                                    timeout=timeout,
                                                    check_interval=check_interval)
            except k8s.client.rest.ApiException as e:
                if e.status == 404:
                    LOG.info(f'Kubernetes Service "{service_name}" already deleted.')
                else:
                    LOG.error(f'Exception when deleting Kubernetes Service: {e}')
                    raise

        def _wait_for_service_deletion(self, service_name, timeout, check_interval):
            svc_instance = k8s.client.CoreV1Api()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    svc_instance.read_namespaced_service(name=service_name, namespace=self.namespace)
                    LOG.info(f'Waiting for kubernetes Service "{service_name}" to be deleted...')
                except k8s.client.rest.ApiException as e:
                    if e.status == 404:
                        LOG.info(f'Kubernetes Service "{service_name}" successfully deleted.')
                        return
                    else:
                        LOG.error(f'Error while checking Service deletion status: {e}')
                        raise
                time.sleep(check_interval)
            LOG.warning(f'Timeout while waiting for kubernetes Service "{service_name}" to be deleted.')

        def _create_or_update_gateway(self):
            networking_api = k8s.client.CustomObjectsApi()
            gateway_spec = {
                'apiVersion': 'gateway.networking.k8s.io/v1beta1',
                'kind': 'Gateway',
                'metadata': {
                    'name': self.gateway_name,
                    'namespace': self.namespace,
                    'annotations': {
                        'networking.istio.io/service-type': self.svc_type
                    }
                },
                'spec': {
                    'gatewayClassName': self.gateway_class_name,
                    'listeners': [
                        {
                            'name': f'httproute-{self.deployment_name}',
                            'port': self.deployment_port,
                            'protocol': 'HTTP',
                        }
                    ]
                }
            }

            try:
                existing_gateway = networking_api.get_namespaced_custom_object(
                    group='gateway.networking.k8s.io',
                    version='v1beta1',
                    namespace=self.namespace,
                    plural='gateways',
                    name=self.gateway_name
                )

                existing_gateway['spec']['listeners'].extend(gateway_spec['spec']['listeners'])
                networking_api.replace_namespaced_custom_object(
                    group='gateway.networking.k8s.io',
                    version='v1beta1',
                    namespace=self.namespace,
                    plural='gateways',
                    name=self.gateway_name,
                    body=existing_gateway
                )
                LOG.info(f'Kubernetes Gateway "{self.gateway_name}" updated successfully.')
            except k8s.client.rest.ApiException as e:
                if e.status == 404:
                    try:
                        networking_api.create_namespaced_custom_object(
                            group='gateway.networking.k8s.io',
                            version='v1beta1',
                            namespace=self.namespace,
                            plural='gateways',
                            body=gateway_spec
                        )
                        LOG.info(f'Kubernetes Gateway "{self.gateway_name}" created successfully.')
                    except k8s.client.rest.ApiException as e_create:
                        LOG.error(f'Exception when creating Gateway: {e_create}')
                        raise
                else:
                    LOG.error(f'Exception when updating Gateway: {e}')
                    raise

        def _delete_or_update_gateway(self, wait_for_completion=True, timeout=60, check_interval=5):
            k8s.config.load_kube_config(self._launcher.kube_config_path)
            gateway_instance = k8s.client.CustomObjectsApi()
            try:
                gateway = gateway_instance.get_namespaced_custom_object(
                    group='gateway.networking.k8s.io',
                    version='v1beta1',
                    namespace=self.namespace,
                    plural='gateways',
                    name=self.gateway_name
                )

                listeners = gateway['spec']['listeners']
                gateway['spec']['listeners'] = [
                    listener for listener in listeners if listener['name'] != f'httproute-{self.deployment_name}'
                ]

                if gateway['spec']['listeners']:
                    gateway_instance.replace_namespaced_custom_object(
                        group='gateway.networking.k8s.io',
                        version='v1beta1',
                        namespace=self.namespace,
                        plural='gateways',
                        name=self.gateway_name,
                        body=gateway
                    )
                    LOG.info(f'Kubernetes Gateway "{self.gateway_name}" deleted updated.')

                    if wait_for_completion:
                        self._wait_for_gateway_update(timeout=timeout, check_interval=check_interval)
                else:
                    gateway_instance.delete_namespaced_custom_object(
                        group='gateway.networking.k8s.io',
                        version='v1beta1',
                        namespace=self.namespace,
                        plural='gateways',
                        name=self.gateway_name
                    )
                    LOG.info(f'Kubernetes Gateway "{self.gateway_name}" deleted.')

                    if wait_for_completion:
                        self._wait_for_gateway_deletion(timeout=timeout, check_interval=check_interval)
            except k8s.client.rest.ApiException as e:
                if e.status == 404:
                    LOG.info(f'Gateway "{self.gateway_name}" already deleted.')
                else:
                    LOG.error(f'Exception when deleting or updating Gateway: {e}')
                    raise

        def _wait_for_gateway_deletion(self, timeout, check_interval):
            gateway_instance = k8s.client.CustomObjectsApi()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    gateway_instance.get_namespaced_custom_object(
                        group='gateway.networking.k8s.io',
                        version='v1beta1',
                        namespace=self.namespace,
                        plural='gateways',
                        name=self.gateway_name
                    )
                    LOG.info(f'Waiting for Gateway "{self.gateway_name}" to be deleted...')
                except k8s.client.rest.ApiException as e:
                    if e.status == 404:
                        LOG.info(f'Gateway "{self.gateway_name}" successfully deleted.')
                        return
                    else:
                        LOG.error(f'Error while checking Gateway deletion status: {e}')
                        raise
                time.sleep(check_interval)
            LOG.warning(f'Timeout while waiting for Gateway "{self.gateway_name}" to be deleted.')

        def _wait_for_gateway_update(self, timeout, check_interval):
            gateway_instance = k8s.client.CustomObjectsApi()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    gateway_instance.get_namespaced_custom_object(
                        group='gateway.networking.k8s.io',
                        version='v1beta1',
                        namespace=self.namespace,
                        plural='gateways',
                        name=self.gateway_name
                    )
                    LOG.info(f'Gateway "{self.gateway_name}" status check passed.')
                    return
                except k8s.client.rest.ApiException as e:
                    LOG.error(f'Error while checking Gateway update status: {e}')
                    raise
                time.sleep(check_interval)
            LOG.warning(f'Timeout while waiting for Gateway "{self.gateway_name}" update.')

        def _create_httproute(self):
            custom_api = k8s.client.CustomObjectsApi()

            httproute_name = f'httproute-{self.deployment_name}'
            httproute_spec = {
                'apiVersion': 'gateway.networking.k8s.io/v1beta1',
                'kind': 'HTTPRoute',
                'metadata': {
                    'name': httproute_name,
                    'namespace': self.namespace
                },
                'spec': {
                    'parentRefs': [{
                        'name': self.gateway_name,
                        'port': self.deployment_port,
                        'sectionName': httproute_name
                    }],
                    'rules': [{
                        'matches': [{
                            'path': {
                                'type': 'PathPrefix',
                                'value': self.path
                            }
                        }],
                        'backendRefs': [{
                            'name': f'service-{self.deployment_name}',
                            'port': self.deployment_port
                        }]
                    }]
                }
            }

            if self.host:
                httproute_spec['spec']['hostnames'] = [self.host]

            try:
                custom_api.create_namespaced_custom_object(
                    group='gateway.networking.k8s.io',
                    version='v1beta1',
                    namespace=self.namespace,
                    plural='httproutes',
                    body=httproute_spec
                )
                LOG.info(f'Kubernetes HTTPRoute "{httproute_name}" created successfully.')
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when creating HTTPRoute: {e}')
                raise

        def _delete_httproute(self, wait_for_deletion=True, timeout=60, check_interval=5):
            k8s.config.load_kube_config(self._launcher.kube_config_path)
            httproute_instance = k8s.client.CustomObjectsApi()
            httproute_name = f'httproute-{self.deployment_name}'
            try:
                httproute_instance.delete_namespaced_custom_object(
                    group='gateway.networking.k8s.io',
                    version='v1beta1',
                    namespace=self.namespace,
                    plural='httproutes',
                    name=httproute_name
                )
                LOG.info(f'Kubernetes HTTPRoute "{httproute_name}" delete initiated.')
            except k8s.client.rest.ApiException as e:
                if e.status == 404:
                    LOG.info(f'HTTPRoute "{httproute_name}" already deleted.')
                    return
                else:
                    LOG.error(f'Exception when deleting HTTPRoute: {e}')
                    raise

            if wait_for_deletion:
                start_time = time.time()
                while time.time() - start_time < timeout:
                    try:
                        httproute_instance.get_namespaced_custom_object(
                            group='gateway.networking.k8s.io',
                            version='v1beta1',
                            namespace=self.namespace,
                            plural='httproutes',
                            name=httproute_name
                        )
                        LOG.info(f'Waiting for HTTPRoute "{httproute_name}" to be deleted...')
                    except k8s.client.rest.ApiException as e:
                        if e.status == 404:
                            LOG.info(f'HTTPRoute "{httproute_name}" successfully deleted.')
                            return
                        else:
                            LOG.error(f'Error while checking HTTPRoute status: {e}')
                            raise
                    time.sleep(check_interval)
                LOG.warning(f'Timeout while waiting for HTTPRoute "{httproute_name}" to be deleted.')

        def _start(self, *, fixed=False):
            cmd = self.get_executable_cmd(fixed=fixed)
            if self.launch_type == 'inference':
                self._create_deployment(cmd=cmd)
                self._expose_deployment()
                if self.on_gateway:
                    self._create_or_update_gateway()
                    self._create_httproute()
            else:
                self._create_job(cmd=cmd)

            self.jobid = self._get_jobid()
            self._launcher.all_processes[self._launcher._id].append((self.jobid, self))
            ret = self.wait()
            LOG.info(ret)

        def stop(self):
            if self.launch_type == 'inference':
                if self.on_gateway:
                    self._delete_or_update_gateway()
                    self._delete_httproute()
                self._delete_service()
                self._delete_deployment()
            else:
                self._delete_job()

        def _get_jobid(self):
            return f'service-{self.deployment_name}' if self.launch_type == 'inference' \
                else f'job-{self.deployment_name}'

        def _get_gateway_service_name(self):
            core_api = k8s.client.CoreV1Api()
            try:
                services = core_api.list_namespaced_service(namespace=self.namespace)

                for service in services.items:
                    labels = service.metadata.labels
                    if labels and ('gateway' in labels.get('app', '') or self.gateway_name in service.metadata.name):
                        LOG.info(f'Kubernetes Gateway service name: {service.metadata.name}')
                        return service.metadata.name

                LOG.warning('No Service was found corresponding to the specified Gateway.')
                return None
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when retrieving Gateway Service: {e}')
                return None

        def _get_gateway_deployment_name(self):  # noqa: C901
            core_api = k8s.client.CoreV1Api()
            apps_v1 = k8s.client.AppsV1Api()

            gateway_service_name = self._get_gateway_service_name()
            try:
                service = core_api.read_namespaced_service(gateway_service_name, self.namespace)
                selector = service.spec.selector
                if selector:
                    label_selector = ','.join(f'{k}={v}' for k, v in selector.items())
                    pods = core_api.list_namespaced_pod(self.namespace, label_selector=label_selector).items
                    if not pods:
                        LOG.warning(f'No Pods found for Service "{gateway_service_name}" in namespace '
                                    f'"{self.namespace}".')
                        return None

                    deployments = set()
                    for pod in pods:
                        for owner in pod.metadata.owner_references:
                            if owner.kind == 'ReplicaSet':
                                rs = apps_v1.read_namespaced_replica_set(owner.name, self.namespace)
                                for rs_owner in rs.metadata.owner_references:
                                    if rs_owner.kind == 'Deployment':
                                        deployments.add(rs_owner.name)

                    if deployments:
                        for deployment_name in deployments:
                            isRestart = False
                            deployment = apps_v1.read_namespaced_deployment(deployment_name, self.namespace)
                            for container in deployment.spec.template.spec.containers:
                                if container.name == 'istio-proxy' and container.image_pull_policy == 'Always':
                                    container.image_pull_policy = 'IfNotPresent'
                                    isRestart = True
                            if isRestart:
                                apps_v1.replace_namespaced_deployment(name=deployment_name, namespace=self.namespace,
                                                                      body=deployment)
                                LOG.info(f'Updated {deployment_name} with imagePullPolicy "IfNotPresent"')
                        return list(deployments)
                    else:
                        LOG.warning(f'No Deployment found for Gateway "{self.gateway_name}" in namespace '
                                    f'"{self.namespace}".')
                        return None
                else:
                    LOG.warning(f'Kubernetes Service "{gateway_service_name}" does not have a selector.')
                    return None
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Error fetching Service "{gateway_service_name}": {e}')
                return None

        def _get_gateway_ip(self):
            core_api = k8s.client.CoreV1Api()
            gateway_service_name = self._get_gateway_service_name()
            if gateway_service_name is None:
                raise ValueError('Kubernetes Gateway service name not found.')
            try:
                service = core_api.read_namespaced_service(
                    name=gateway_service_name,
                    namespace=self.namespace
                )

                if service.spec.type == 'LoadBalancer':
                    if service.status.load_balancer.ingress:
                        ip = service.status.load_balancer.ingress[0].ip
                        return ip
                    else:
                        LOG.warning('The LoadBalancer IP has not been assigned yet.')
                        return None
                elif service.spec.type == 'NodePort':
                    nodes = core_api.list_node()
                    node_ip = nodes.items[0].status.addresses[0].address
                    return node_ip
                elif service.spec.type == 'ClusterIP':
                    return service.spec.cluster_ip
                else:
                    LOG.warning('Unsupported Service type.')
                    return None
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when retrieving gateway IP: {e}')
                return None

        def _get_httproute_host(self):
            custom_api = k8s.client.CustomObjectsApi()
            try:
                httproute = custom_api.get_namespaced_custom_object(
                    group='gateway.networking.k8s.io',
                    version='v1beta1',
                    namespace=self.namespace,
                    plural='httproutes',
                    name=f'httproute-{self.deployment_name}'
                )

                hostnames = httproute.get('spec', {}).get('hostnames', [])
                if hostnames:
                    return hostnames[0]
                else:
                    LOG.warning('Kubernetes HTTPRoute has no configured hostnames.')
                    return None
            except k8s.client.rest.ApiException as e:
                LOG.error(f'Exception when retrieving HTTPRoute host: {e}')
                return None

        def get_jobip(self):
            if not self.on_gateway: return f'service-{self.deployment_name}'
            host = self._get_httproute_host()
            ip = self._get_gateway_ip()
            LOG.info(f'gateway ip: {ip}, hostname: {host}')
            return host if host else ip

        def wait_for_deployment_ready(self, timeout=300):
            api_instance = k8s.client.AppsV1Api()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    deployment_status = api_instance.read_namespaced_deployment_status(
                        name=self.deployment_name,
                        namespace=self.namespace
                    ).status
                    if deployment_status.available_replicas and deployment_status.available_replicas > 0:
                        LOG.info(f'Kubernetes Deployment "{self.deployment_name}" is running.')
                        return True
                    time.sleep(2)
                except k8s.client.rest.ApiException as e:
                    LOG.error(f'Exception when reading Deployment status: {e}')
                    raise
            LOG.warning(f'Timed out waiting for Deployment "{self.deployment_name}" to be ready.')
            return False

        def _is_service_ready(self, timeout):
            if self.on_gateway: return True
            url = f'http://service-{self.deployment_name}:{self.deployment_port}{self.path}'
            for i in range(self.gateway_retry):
                try:
                    response = requests.get(url, timeout=timeout)
                    if response.status_code != 503:
                        LOG.info(f'Kubernetes Service is ready at "{url}"')
                        self.queue.put(f'Uvicorn running on {url}')
                        return True
                    else:
                        LOG.info(f'Kubernetes Service at "{url}" returned status code {response.status_code}')
                except requests.RequestException as e:
                    LOG.error(f'Failed to access service at "{url}": {e}, retry: {i}/{self.gateway_retry}')
                    # raise
                time.sleep(timeout)

            self.queue.put(f'ERROR: Kubernetes Service failed to start on "{url}".')
            return False

        def _wait_for_service_ready(self, timeout=300):
            svc_instance = k8s.client.CoreV1Api()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    service = svc_instance.read_namespaced_service(
                        name=f'service-{self.deployment_name}',
                        namespace=self.namespace
                    )
                    if service.spec.type == 'LoadBalancer' and service.status.load_balancer.ingress:
                        ip = service.status.load_balancer.ingress[0].ip
                        LOG.info(f'Kubernetes Service "service-{self.deployment_name}" is ready with IP: {ip}')
                        return ip
                    elif service.spec.cluster_ip:
                        LOG.info(f'Kubernetes Service "service-{self.deployment_name}" is ready with ClusterIP: '
                                 f'{service.spec.cluster_ip}')
                        return service.spec.cluster_ip
                    elif service.spec.type == 'NodePort':
                        node_ports = [p.node_port for p in service.spec.ports]
                        if node_ports:
                            nodes = svc_instance.list_node()
                            for node in nodes.items:
                                for address in node.status.addresses:
                                    if address.type == 'InternalIP':
                                        node_ip = address.address
                                        LOG.info(f'Kubernetes Service "service-{self.deployment_name}" is ready on '
                                                 f'NodePort(s): {node_ports} at Node IP: {node_ip}')
                                        return {'ip': node_ip, 'ports': node_ports}
                                    elif address.type == 'ExternalIP':
                                        node_ip = address.address
                                        LOG.info(f'Kubernetes Service "service-{self.deployment_name}" is ready on '
                                                 f'NodePort(s): {node_ports} at External Node IP: {node_ip}')
                                        return {'ip': node_ip, 'ports': node_ports}
                    LOG.info(f'Kubernetes Service "service-{self.deployment_name}" is not ready yet. Retrying...')
                    time.sleep(2)
                except k8s.client.rest.ApiException as e:
                    LOG.error(f'Exception when reading Service status: {e}')
                    raise
            LOG.warning(f'Timed out waiting for Service "service-{self.deployment_name}" to be ready.')
            return None

        def wait_for_service_ready(self, timeout=300, interval=5):
            _service_ready = self._wait_for_service_ready(timeout=timeout)
            return _service_ready if _service_ready and self._is_service_ready(timeout=interval) else None

        def _is_gateway_ready(self, timeout):
            url = f'http://{self.get_jobip()}:{self.deployment_port}{self.path}'
            for _ in range(self.gateway_retry):
                try:
                    response = requests.get(url, timeout=timeout)
                    if response.status_code != 503:
                        LOG.info(f'Kubernetes Service is ready at "{url}"')
                        self.queue.put(f'Uvicorn running on {url}')
                        return True
                    else:
                        LOG.info(f'Kubernetes Service at "{url}" returned status code {response.status_code}')
                except requests.RequestException as e:
                    LOG.error(f'Failed to access service at "{url}": {e}')
                    raise
                time.sleep(timeout)

            self.queue.put(f'ERROR: Kubernetes Service failed to start on "{url}".')
            return False

        def wait_for_gateway(self, timeout=300, interval=5):  # noqa: C901
            core_v1 = k8s.client.CoreV1Api()
            apps_v1 = k8s.client.AppsV1Api()
            gateway_service_name = self._get_gateway_service_name()
            gateway_deployment_names = self._get_gateway_deployment_name()
            service_ready = False
            deployment_ready = False

            start_time = time.time()
            while time.time() - start_time < timeout:
                if not service_ready:
                    try:
                        service = core_v1.read_namespaced_service(gateway_service_name, self.namespace)
                        if service.spec.type in ['NodePort', 'LoadBalancer']:
                            if service.spec.type == 'LoadBalancer':
                                if service.status.load_balancer.ingress:
                                    LOG.info(f'Kubernetes Service "{gateway_service_name}" is ready with '
                                             'LoadBalancer IP.')
                                    service_ready = True
                                else:
                                    LOG.info(f'Kubernetes Service "{gateway_service_name}" LoadBalancer IP '
                                             'not available yet.')
                                    service_ready = False
                            else:
                                if any(port.node_port for port in service.spec.ports):
                                    LOG.info(f'Kubernetes Service "{gateway_service_name}" is ready with '
                                             'NodePort configuration.')
                                    service_ready = True
                                else:
                                    LOG.info(f'Kubernetes Service "{gateway_service_name}" NodePort not assigned yet.')
                                    service_ready = False
                        else:
                            LOG.error(f'Unexpected Kubernetes Service type: {service.spec.type}.')
                            service_ready = False
                    except k8s.client.rest.ApiException as e:
                        LOG.error(f'Kubernetes Service "{gateway_service_name}" not found yet: {e}')
                        service_ready = False
                if not deployment_ready:
                    for deployment_name in gateway_deployment_names:
                        try:
                            deployment = apps_v1.read_namespaced_deployment(deployment_name, self.namespace)
                            if deployment.status.available_replicas and deployment.status.available_replicas > 0:
                                LOG.info(f'Kubernetes Deployment "{deployment_name}" is ready with '
                                         f'{deployment.status.available_replicas} replicas.')
                                deployment_ready = True
                                break
                            else:
                                LOG.info(f'Kubernetes Deployment "{deployment_name}" is not fully ready yet.')
                                deployment_ready = False
                        except k8s.client.rest.ApiException as e:
                            LOG.warning(f'Kubernetes Deployment "{deployment_name}" not found yet: {e}')
                            deployment_ready = False

                if service_ready and deployment_ready and self._is_gateway_ready(timeout=interval):
                    LOG.info(f'Kubernetes Gateway "{self.gateway_name}" is fully ready.')
                    return True

                time.sleep(interval)

            LOG.error(f'Kubernetes Gateway "{self.gateway_name}" failed to become ready with {timeout} seconds.')
            return False

        def wait_for_httproute(self, timeout=300):
            custom_api = k8s.client.CustomObjectsApi()
            start_time = time.time()
            while time.time() - start_time < timeout:
                try:
                    httproutes = custom_api.list_namespaced_custom_object(
                        group='gateway.networking.k8s.io',
                        version='v1beta1',
                        namespace=self.namespace,
                        plural='httproutes'
                    ).get('items', [])

                    for httproute in httproutes:
                        if httproute['metadata']['name'] == f'httproute-{self.deployment_name}':
                            LOG.info(f'Kubernetes HTTPRoute "httproute-{self.deployment_name}" is ready.')
                            return True
                    LOG.info(f'Waiting for HTTPRoute "httproute-{self.deployment_name}" to be ready...')
                except k8s.client.rest.ApiException as e:
                    LOG.error(f'Exception when checking HTTPRoute status: {e}')
                    raise

                time.sleep(2)
            LOG.warning(f'Timeout waiting for HTTPRoute "httproute-{self.deployment_name}" to be ready.')
            return False

        def wait(self):
            if self.launch_type == 'inference':
                deployment_ready = self.wait_for_deployment_ready()
                if not deployment_ready:
                    raise TimeoutError('Kubernetes Deployment did not become ready in time.')

                service_ip = self.wait_for_service_ready(interval=10)
                if not service_ip:
                    raise TimeoutError('Kubernetes Service did not become ready in time.')

                httproute_ready = True if not self.on_gateway else self.wait_for_httproute()
                if not httproute_ready:
                    raise TimeoutError('Kubernetes Httproute did not become ready in time.')

                gateway_ready = True if not self.on_gateway else self.wait_for_gateway()
                if not gateway_ready:
                    raise TimeoutError('Kubernetes Gateway did not become ready in time.')

                return {'deployment': Status.Running, 'service_ip': service_ip,
                        'gateway': Status.Running, 'httproute': Status.Running}
            else:
                return {'job': Status.Running}

        @property
        def status(self):
            if self.launch_type == 'inference':
                api_instance = k8s.client.AppsV1Api()
                try:
                    deployment_status = api_instance.read_namespaced_deployment_status(
                        name=self.deployment_name,
                        namespace=self.namespace
                    ).status
                    if deployment_status.available_replicas and deployment_status.available_replicas > 0:
                        return Status.Running
                    else:
                        return Status.Pending
                except k8s.client.rest.ApiException as e:
                    LOG.error(f'Exception when reading Deployment status: {e}')
                    return Status.Failed
            else:
                api_instance = k8s.client.BatchV1Api()
                try:
                    job_status = api_instance.read_namespaced_job_status(
                        name=self.deployment_name,
                        namespace=self.namespace
                    ).status
                    if getattr(job_status, 'succeeded', 0) and job_status.succeeded >= 1:
                        return Status.Done
                    elif getattr(job_status, 'active', 0) and job_status.active >= 1:
                        return Status.Running
                    elif getattr(job_status, 'failed', 0) and job_status.failed >= 1:
                        return Status.Failed
                    else:
                        return Status.Pending
                except k8s.client.rest.ApiException as e:
                    LOG.error(f'Exception when reading Job status: {e}')
                    return Status.Failed

    def __init__(self, kube_config_path=None, volume_configs=None, image=None, resource_config=None,
                 namespace=None, on_gateway=None, gateway_name=None, gateway_class_name=None, host=None, path=None,
                 svc_type: Literal['LoadBalancer', 'NodePort', 'ClusterIP'] = None, retry=3,
                 sync=True, ngpus=None, **kwargs):
        super().__init__()
        self.gateway_retry = retry
        self.sync = sync
        self.ngpus = ngpus
        self.launch_type = kwargs.get('launch_type', 'inference')
        config_data = self._read_config_file(lazyllm.config['k8s_config_path']) if lazyllm.config['k8s_config_path'] \
            else {}
        self.volume_configs = volume_configs if volume_configs else config_data.get('volume', [])
        self.image = image if image else config_data.get('container_image', 'lazyllm/lazyllm:k8s_launcher')
        self.resource_config = resource_config if resource_config else config_data.get('resource', {})
        self.kube_config_path = kube_config_path if kube_config_path \
            else config_data.get('kube_config_path', '~/.kube/config')
        self.svc_type = svc_type if svc_type else config_data.get('svc_type', 'LoadBalancer')
        self.namespace = namespace if namespace else config_data.get('namespace', 'default')
        self.on_gateway = on_gateway if on_gateway else config_data.get('on_gateway', False)
        self.gateway_name = gateway_name if gateway_name else config_data.get('gateway_name', 'lazyllm-gateway')
        self.gateway_class_name = gateway_class_name if gateway_class_name \
            else config_data.get('gateway_class_name', 'istio')
        self.http_host = host if host else config_data.get('host', None)
        self.http_path = path if path else config_data.get('path', '/generate')

    def _read_config_file(self, file_path):
        assert os.path.isabs(file_path), 'Resource config file path must be an absolute path.'
        with open(file_path, 'r') as fp:
            try:
                data = yaml.safe_load(fp)
                return data
            except yaml.YAMLError as e:
                LOG.error(f'Exception when reading resource configuration file: {e}')
                raise ValueError('Kubernetes resource configuration file format error.')

    def makejob(self, cmd):
        """创建一个Kubernetes作业实例。

Args:
    cmd (str): 要执行的命令。

**Returns:**

- K8sLauncher.Job: 一个新的Kubernetes作业实例。
"""
        # TODO(wangzhihong): support thread-local kube config by `client = config.new_client_from_config`
        k8s.config.load_kube_config(self.kube_config_path)
        return K8sLauncher.Job(cmd, launcher=self, sync=self.sync)

    def launch(self, f, *args, **kw):
        """启动一个Kubernetes作业或可调用对象。

Args:
    f (K8sLauncher.Job): 要启动的Kubernetes作业实例。
    *args: 位置参数。
    **kw: 关键字参数。

**Returns:**

- Any: 作业的返回值。

Raises:
    RuntimeError: 当提供的不是Deployment对象时抛出。
"""
        if isinstance(f, K8sLauncher.Job):
            f.start()
            LOG.info('Launcher started successfully.')
            self.job = f
            return f.return_value
        elif callable(f):
            LOG.info('Async execution in Kubernetes is not supported currently.')
            raise RuntimeError('Kubernetes launcher requires a Deployment object.')

makejob(cmd)

创建一个Kubernetes作业实例。

Parameters:

  • cmd (str) –

    要执行的命令。

Returns:

  • K8sLauncher.Job: 一个新的Kubernetes作业实例。
Source code in lazyllm/launcher/k8s.py 
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
    def makejob(self, cmd):
        """创建一个Kubernetes作业实例。

Args:
    cmd (str): 要执行的命令。

**Returns:**

- K8sLauncher.Job: 一个新的Kubernetes作业实例。
"""
        # TODO(wangzhihong): support thread-local kube config by `client = config.new_client_from_config`
        k8s.config.load_kube_config(self.kube_config_path)
        return K8sLauncher.Job(cmd, launcher=self, sync=self.sync)

launch(f, *args, **kw)

启动一个Kubernetes作业或可调用对象。

Parameters:

  • f (Job) –

    要启动的Kubernetes作业实例。

  • *args –

    位置参数。

  • **kw –

    关键字参数。

Returns:

  • Any: 作业的返回值。

Raises:

  • RuntimeError –

    当提供的不是Deployment对象时抛出。

Source code in lazyllm/launcher/k8s.py 
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
    def launch(self, f, *args, **kw):
        """启动一个Kubernetes作业或可调用对象。

Args:
    f (K8sLauncher.Job): 要启动的Kubernetes作业实例。
    *args: 位置参数。
    **kw: 关键字参数。

**Returns:**

- Any: 作业的返回值。

Raises:
    RuntimeError: 当提供的不是Deployment对象时抛出。
"""
        if isinstance(f, K8sLauncher.Job):
            f.start()
            LOG.info('Launcher started successfully.')
            self.job = f
            return f.return_value
        elif callable(f):
            LOG.info('Async execution in Kubernetes is not supported currently.')
            raise RuntimeError('Kubernetes launcher requires a Deployment object.')