worker_base.py 11.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import os
5
6
from typing import (Any, Callable, Dict, List, Optional, Set, Tuple, TypeVar,
                    Union)
7

8
import cloudpickle
9
import torch.nn as nn
10

11
from vllm.config import VllmConfig, set_current_vllm_config
12
from vllm.logger import init_logger
13
from vllm.lora.request import LoRARequest
14
from vllm.model_executor.layers.sampler import SamplerOutput
15
from vllm.sequence import ExecuteModelRequest
16
from vllm.utils import (enable_trace_function_call_for_thread,
17
                        resolve_obj_by_qualname, run_method,
18
19
                        update_environment_variables,
                        warn_for_unimplemented_methods)
20
21

logger = init_logger(__name__)
22

23
24
_R = TypeVar("_R")

25

26
27
@warn_for_unimplemented_methods
class WorkerBase:
28
    """Worker interface that allows vLLM to cleanly separate implementations for
29
30
    different hardware. Also abstracts control plane communication, e.g., to
    communicate request metadata to other workers.
31
32
    """

33
34
35
36
37
38
39
40
41
42
43
44
45
46
    def __init__(
        self,
        vllm_config: VllmConfig,
    ) -> None:
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        self.lora_config = vllm_config.lora_config
        self.load_config = vllm_config.load_config
        self.parallel_config = vllm_config.parallel_config
        self.scheduler_config = vllm_config.scheduler_config
        self.device_config = vllm_config.device_config
        self.speculative_config = vllm_config.speculative_config
        self.observability_config = vllm_config.observability_config
47
        self.kv_transfer_config = vllm_config.kv_transfer_config
48
        self.compilation_config = vllm_config.compilation_config
49
50
        from vllm.platforms import current_platform
        self.current_platform = current_platform
51

52
53
54
55
56
57
58
59
60
61
62
63
    def init_device(self) -> None:
        """Initialize device state, such as loading the model or other on-device
        memory allocations.
        """
        raise NotImplementedError

    def initialize_cache(self, num_gpu_blocks: int,
                         num_cpu_blocks: int) -> None:
        """Initialize the KV cache with the given size in blocks.
        """
        raise NotImplementedError

64
65
66
    def get_model(self) -> nn.Module:
        raise NotImplementedError

67
68
69
70
    def apply_model(self, fn: Callable[[nn.Module], _R]) -> _R:
        """Apply a function on the model inside this worker."""
        return fn(self.get_model())

71
72
73
74
75
76
77
78
79
80
    def load_model(self) -> None:
        """Load model onto target device."""
        raise NotImplementedError

    def execute_model(
        self,
        execute_model_req: Optional[ExecuteModelRequest] = None
    ) -> Optional[List[SamplerOutput]]:
        raise NotImplementedError

81
82
83
84
85
86
    def start_worker_execution_loop(self) -> None:
        """Execute model loop in parallel worker.

        You can stop the loop by executing a driver worker with an empty output.
        See `stop_remote_worker_execution_loop` for more details.
        """
87
88
89
90
91
        with self.current_platform.inference_mode():
            while True:
                output = self.execute_model(execute_model_req=None)
                if output is None:
                    return None
92

93
94
95
    def determine_num_available_blocks(self) -> Tuple[int, int]:
        """Determine the number of available blocks for the GPU KV cache and
        swappable CPU KV cache.
96

97
98
99
100
101
102
103
104
        The implementation may run profiling or other heuristics to determine
        the size of caches.

        Returns a Tuple[num_gpu_blocks, num_cpu_blocks], where num_gpu_blocks
        are blocks that are "active" on the device and can be appended to.
        num_cpu_blocks refers to "swapped" blocks in CPU memory and cannot be
        appended to.
        """
105
106
        raise NotImplementedError

107
    def get_cache_block_size_bytes(self) -> int:
108
109
110
111
112
113
114
115
116
117
118
        """Return the size of a single cache block, in bytes. Used in
        speculative decoding.
        """
        raise NotImplementedError

    def add_lora(self, lora_request: LoRARequest) -> bool:
        raise NotImplementedError

    def remove_lora(self, lora_id: int) -> bool:
        raise NotImplementedError

119
120
121
    def pin_lora(self, lora_id: int) -> bool:
        raise NotImplementedError

122
    def list_loras(self) -> Set[int]:
123
124
        raise NotImplementedError

125
126
127
128
129
    @property
    def vocab_size(self) -> int:
        """Get vocabulary size from model configuration."""
        return self.model_config.get_vocab_size()

130
131
132
133
    def shutdown(self) -> None:
        """Clean up resources held by the worker."""
        return

134

135
136
class WorkerWrapperBase:
    """
137
138
    This class represents one process in an executor/engine. It is responsible
    for lazily initializing the worker and handling the worker's lifecycle.
139
140
141
142
143
    We first instantiate the WorkerWrapper, which remembers the worker module
    and class name. Then, when we call `update_environment_variables`, and the
    real initialization happens in `init_worker`.
    """

144
145
    def __init__(
        self,
146
        vllm_config: VllmConfig,
147
        rpc_rank: int = 0,
148
    ) -> None:
149
150
151
152
153
154
155
156
157
158
159
        """
        Initialize the worker wrapper with the given vllm_config and rpc_rank.
        Note: rpc_rank is the rank of the worker in the executor. In most cases,
        it is also the rank of the worker in the distributed group. However,
        when multiple executors work together, they can be different.
        e.g. in the case of SPMD-style offline inference with TP=2,
        users can launch 2 engines/executors, each with only 1 worker.
        All workers have rpc_rank=0, but they have different ranks in the TP
        group.
        """
        self.rpc_rank = rpc_rank
160
        self.worker: Optional[WorkerBase] = None
161
        self.vllm_config: Optional[VllmConfig] = None
162
163
164
165
        # do not store this `vllm_config`, `init_worker` will set the final
        # one. TODO: investigate if we can remove this field in
        # `WorkerWrapperBase`, `init_cached_hf_modules` should be
        # unnecessary now.
166
167
168
169
170
171
172
173
        if vllm_config.model_config is not None:
            # it can be None in tests
            trust_remote_code = vllm_config.model_config.trust_remote_code
            if trust_remote_code:
                # note: lazy import to avoid importing torch before initializing
                from vllm.utils import init_cached_hf_modules
                init_cached_hf_modules()

174
175
176
177
    def shutdown(self) -> None:
        if self.worker is not None:
            self.worker.shutdown()

178
179
    def adjust_rank(self, rank_mapping: Dict[int, int]) -> None:
        """
180
        Adjust the rpc_rank based on the given mapping.
181
        It is only used during the initialization of the executor,
182
        to adjust the rpc_rank of workers after we create all workers.
183
        """
184
185
        if self.rpc_rank in rank_mapping:
            self.rpc_rank = rank_mapping[self.rpc_rank]
186

187
188
    def update_environment_variables(self, envs_list: List[Dict[str,
                                                                str]]) -> None:
189
        envs = envs_list[self.rpc_rank]
190
191
192
193
194
195
196
        key = 'CUDA_VISIBLE_DEVICES'
        if key in envs and key in os.environ:
            # overwriting CUDA_VISIBLE_DEVICES is desired behavior
            # suppress the warning in `update_environment_variables`
            del os.environ[key]
        update_environment_variables(envs)

197
    def init_worker(self, all_kwargs: List[Dict[str, Any]]) -> None:
198
        """
199
        Here we inject some common logic before initializing the worker.
200
201
        Arguments are passed to the worker class constructor.
        """
202
        kwargs = all_kwargs[self.rpc_rank]
203
        self.vllm_config = kwargs.get("vllm_config")
204
205
        assert self.vllm_config is not None, (
            "vllm_config is required to initialize the worker")
206
        enable_trace_function_call_for_thread(self.vllm_config)
207

208
209
210
        from vllm.plugins import load_general_plugins
        load_general_plugins()

211
212
213
214
        if isinstance(self.vllm_config.parallel_config.worker_cls, str):
            worker_class = resolve_obj_by_qualname(
                self.vllm_config.parallel_config.worker_cls)
        else:
215
216
217
218
219
220
            logger.warning(
                "passing worker_cls as a class object is strongly deprecated,"
                " as the serialization of class objects can be tricky and"
                " error-prone. To be safe, please keep the class in a separate"
                " module and pass the qualified name of the class as a string."
            )
221
222
223
224
            assert isinstance(self.vllm_config.parallel_config.worker_cls,
                              bytes)
            worker_class = cloudpickle.loads(
                self.vllm_config.parallel_config.worker_cls)
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
        if self.vllm_config.parallel_config.worker_extension_cls:
            worker_extension_cls = resolve_obj_by_qualname(
                self.vllm_config.parallel_config.worker_extension_cls)
            extended_calls = []
            if worker_extension_cls not in worker_class.__bases__:
                # check any conflicts between worker and worker_extension_cls
                for attr in dir(worker_extension_cls):
                    if attr.startswith("__"):
                        continue
                    assert not hasattr(worker_class, attr), (
                        f"Worker class {worker_class} already has an attribute"
                        f" {attr}, which conflicts with the worker"
                        f" extension class {worker_extension_cls}.")
                    if callable(getattr(worker_extension_cls, attr)):
                        extended_calls.append(attr)
                # dynamically inherit the worker extension class
                worker_class.__bases__ = worker_class.__bases__ + (
                    worker_extension_cls, )
                logger.info(
                    "Injected %s into %s for extended collective_rpc calls %s",
                    worker_extension_cls, worker_class, extended_calls)
246
247
248
249
        with set_current_vllm_config(self.vllm_config):
            # To make vLLM config available during worker initialization
            self.worker = worker_class(**kwargs)
            assert self.worker is not None
250

251
252
    def initialize_from_config(self, kv_cache_configs: List[Any]) -> None:
        kv_cache_config = kv_cache_configs[self.rpc_rank]
253
254
        with set_current_vllm_config(self.vllm_config):
            self.worker.initialize_from_config(kv_cache_config)  # type: ignore
255

256
257
258
259
260
    def init_device(self):
        with set_current_vllm_config(self.vllm_config):
            # To make vLLM config available during device initialization
            self.worker.init_device()  # type: ignore

261
    def execute_method(self, method: Union[str, bytes], *args, **kwargs):
262
        try:
263
264
265
266
267
            # method resolution order:
            # if a method is defined in this class, it will be called directly.
            # otherwise, since we define `__getattr__` and redirect attribute
            # query to `self.worker`, the method will be called on the worker.
            return run_method(self, method, args, kwargs)
268
269
270
271
272
        except Exception as e:
            # if the driver worker also execute methods,
            # exceptions in the rest worker may cause deadlock in rpc like ray
            # see https://github.com/vllm-project/vllm/issues/3455
            # print the error and inform the user to solve the error
273
            msg = (f"Error executing method {method!r}. "
274
275
276
                   "This might cause deadlock in distributed execution.")
            logger.exception(msg)
            raise e
277

278
279
    def __getattr__(self, attr):
        return getattr(self.worker, attr)