worker_base.py 18.7 KB
Newer Older
1
import dataclasses
2
import os
3
import time
4
from abc import ABC, abstractmethod
5
from typing import Any, Dict, List, Optional, Set, Tuple, Type, Union
6

7
8
import torch

9
from vllm.config import ObservabilityConfig, VllmConfig
10
from vllm.distributed import broadcast_tensor_dict, get_pp_group, get_tp_group
11
from vllm.logger import init_logger
12
from vllm.lora.request import LoRARequest
13
14
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.sequence import ExecuteModelRequest, IntermediateTensors
15
from vllm.utils import (enable_trace_function_call_for_thread,
16
                        resolve_obj_by_qualname, update_environment_variables)
17
18
19
from vllm.worker.model_runner_base import (BroadcastableModelInput,
                                           ModelRunnerBase,
                                           ModelRunnerInputBase)
20
21

logger = init_logger(__name__)
22
23
24
25


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

30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
    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.prompt_adapter_config = vllm_config.prompt_adapter_config
        self.observability_config = vllm_config.observability_config
45
        self.kv_transfer_config = vllm_config.kv_transfer_config
46
47
        from vllm.platforms import current_platform
        self.current_platform = current_platform
48

49
50
51
52
53
54
55
56
    @abstractmethod
    def init_device(self) -> None:
        """Initialize device state, such as loading the model or other on-device
        memory allocations.
        """
        raise NotImplementedError

    @abstractmethod
57
    def determine_num_available_blocks(self) -> Tuple[int, int]:
58
59
60
61
62
63
        """Determine the number of available blocks for the GPU KV cache and
        swappable CPU KV cache.

        The implementation may run profiling or other heuristics to determine
        the size of caches.

64
        Returns a Tuple[num_gpu_blocks, num_cpu_blocks], where num_gpu_blocks
65
66
67
68
69
70
71
72
73
74
75
76
77
        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.
        """
        raise NotImplementedError

    @abstractmethod
    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

78
79
80
81
82
83
    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.
        """
84
85
86
87
88
        with self.current_platform.inference_mode():
            while True:
                output = self.execute_model(execute_model_req=None)
                if output is None:
                    return None
89

90
    @abstractmethod
91
    def execute_model(
92
93
        self,
        execute_model_req: Optional[ExecuteModelRequest] = None
94
    ) -> Optional[List[SamplerOutput]]:
95
96
97
        raise NotImplementedError

    @abstractmethod
98
    def get_cache_block_size_bytes(self) -> int:
99
100
101
102
103
104
105
106
107
108
109
110
111
        """Return the size of a single cache block, in bytes. Used in
        speculative decoding.
        """
        raise NotImplementedError

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

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

112
113
114
115
    @abstractmethod
    def pin_lora(self, lora_id: int) -> bool:
        raise NotImplementedError

116
    @abstractmethod
117
    def list_loras(self) -> Set[int]:
118
119
120
121
122
123
124
125
126
127
128
129
130
131
        raise NotImplementedError


class LoraNotSupportedWorkerBase(WorkerBase):
    """Partial implementation of WorkerBase that raises exceptions when LoRA
    methods are invoked.
    """

    def add_lora(self, lora_request: LoRARequest) -> bool:
        raise ValueError(f"{type(self)} does not support LoRA")

    def remove_lora(self, lora_id: int) -> bool:
        raise ValueError(f"{type(self)} does not support LoRA")

132
133
134
135
    def pin_lora(self, lora_id: int) -> bool:
        return ValueError(
            f"{type(self)} does not support LoRA")  # type: ignore

136
    def list_loras(self) -> Set[int]:
137
        raise ValueError(f"{type(self)} does not support LoRA")
138
139


140
141
142
143
144
145
146
147
148
149
@dataclasses.dataclass(frozen=True)
class WorkerInput:
    """Local inputs to each worker. May contain device-specific data. These
    fields should be broadcastable to other workers.
    """

    num_seq_groups: Optional[int] = None
    blocks_to_swap_in: Optional[torch.Tensor] = None
    blocks_to_swap_out: Optional[torch.Tensor] = None
    blocks_to_copy: Optional[torch.Tensor] = None
150
    virtual_engine: int = 0
151
    num_steps: int = 1
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166

    @classmethod
    def from_broadcasted_tensor_dict(
        cls: Type["WorkerInput"],
        tensor_dict: Dict[str, Any],
    ) -> "WorkerInput":
        """
        Pop fields from the given tensor_dict and populate a new instance of
        WorkerInput.
        """
        return cls(
            num_seq_groups=tensor_dict.pop("num_seq_groups"),
            blocks_to_swap_in=tensor_dict.pop("blocks_to_swap_in"),
            blocks_to_swap_out=tensor_dict.pop("blocks_to_swap_out"),
            blocks_to_copy=tensor_dict.pop("blocks_to_copy"),
167
            virtual_engine=tensor_dict["virtual_engine"],
168
            num_steps=tensor_dict.pop("num_steps"),
169
170
171
172
173
174
175
176
177
178
179
180
        )

    def as_broadcastable_tensor_dict(
            self) -> Dict[str, Union[int, torch.Tensor]]:
        """
        Extract broadcastable fields.
        """
        tensor_dict = {
            "num_seq_groups": self.num_seq_groups,
            "blocks_to_swap_in": self.blocks_to_swap_in,
            "blocks_to_swap_out": self.blocks_to_swap_out,
            "blocks_to_copy": self.blocks_to_copy,
181
            "virtual_engine": self.virtual_engine,
182
            "num_steps": self.num_steps,
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
        }

        return tensor_dict


class LocalOrDistributedWorkerBase(WorkerBase):
    """
    Partial implementation of WorkerBase that has a default `execute_model`
    definition to perform metadata transfer between workers when in distributed
    mode. Subclasses of this interface should use model runners that inherit
    from ModelRunnerBase, and should only need to implement worker-local logic.
    If custom control plane logic is needed to transfer metadata, or if the
    model runner cannot inherit from ModelRunnerBase, use WorkerBase instead.
    """
    is_driver_worker: bool
    model_runner: ModelRunnerBase
199
    observability_config: Optional[ObservabilityConfig] = None
200
201
202
203
204
205
206
207
208
209
210
211
212
213

    @property
    @abstractmethod
    def do_metadata_broadcast(self) -> bool:
        """
        Used by the default `execute_model` to check whether broadcast is
        needed to transfer request inputs from the driver worker to other
        workers in the TP group. If WorkerBase subclass only supports
        single-worker execution, then this method should return False.
        """
        raise NotImplementedError

    @property
    @abstractmethod
214
    def kv_cache(self) -> Optional[List[List[torch.Tensor]]]:
215
        """
216
217
218
219
220
        Gets the list of kv caches to pass to the worker's model runner. Each
        element in the list is a kv cache corresponding to a particular virtual
        engine (PP stream). Used by the default `execute_model`. If the worker's
        model runner does not follow the ModelRunnerBase interface, then inherit
        from WorkerBase instead.
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
        """
        raise NotImplementedError

    @abstractmethod
    def prepare_worker_input(
            self, execute_model_req: ExecuteModelRequest) -> WorkerInput:
        """
        Prepare the inputs to WorkerBase.execute_worker from an execution
        request. This method may move data to the worker's local device. It is
        not allowed to communicate with other workers or devices.
        """
        raise NotImplementedError

    @abstractmethod
    def execute_worker(self, worker_input: WorkerInput) -> None:
        """
        Process an execution request.
        """
        raise NotImplementedError

241
    def _get_worker_input_from_broadcast(
242
243
244
        self
    ) -> Optional[Tuple[BroadcastableModelInput, WorkerInput, Dict[
            str, torch.Tensor]]]:
245
246
247
248
249
250
251
252
253
254
255
256
        """ Get the worker input from the broadcasted tensor dict. """
        assert self.do_metadata_broadcast
        assert not self.is_driver_worker
        broadcast_data = broadcast_tensor_dict(src=0)
        if not broadcast_data:
            return None

        worker_input = WorkerInput.from_broadcasted_tensor_dict(broadcast_data)
        model_input = (
            self.model_runner.make_model_input_from_broadcasted_tensor_dict(
                broadcast_data))

257
258
259
        kwargs = extract_previous_hidden_states(broadcast_data)

        return model_input, worker_input, kwargs
260
261
262

    def _get_driver_input_and_broadcast(
        self, execute_model_req: ExecuteModelRequest
263
    ) -> Tuple[BroadcastableModelInput, WorkerInput, Dict[str, torch.Tensor]]:
264
265
266
267
268
269
270
271
272
273
274
        """ Get the driver input and broadcast it to other workers.  """
        assert self.is_driver_worker

        worker_input: WorkerInput = self.prepare_worker_input(
            execute_model_req=execute_model_req)
        model_input: ModelRunnerInputBase = (
            self.model_runner.prepare_model_input(
                execute_model_req.seq_group_metadata_list,
                execute_model_req.virtual_engine,
                execute_model_req.finished_requests_ids))

275
276
        kwargs = extract_previous_hidden_states(execute_model_req)

277
278
279
        if self.do_metadata_broadcast:
            broadcast_data = worker_input.as_broadcastable_tensor_dict()
            broadcast_data.update(model_input.as_broadcastable_tensor_dict())
280
            broadcast_data.update(kwargs)
281
282
            broadcast_tensor_dict(broadcast_data, src=0)

283
        if execute_model_req.async_callback:
284
285
            model_input = dataclasses.replace(  # type: ignore
                model_input,
286
                async_callback=execute_model_req.async_callback)
287

288
        return model_input, worker_input, kwargs
289
290

    def prepare_input(
291
292
        self,
        execute_model_req: Optional[ExecuteModelRequest] = None
293
294
    ) -> Optional[Tuple[BroadcastableModelInput, WorkerInput, Dict[
            str, torch.Tensor]]]:
295
296
297
        """
        Prepare the inputs to ModelRunner and workers.
        """
298
299
300
301
302
303
304
305
306
307
        if self.is_driver_worker:
            if execute_model_req is None:
                if self.do_metadata_broadcast:
                    # This signals that there's no more requests to process for
                    # now. All workers are running infinite loop with
                    # broadcast_tensor_dict, and it stops the loop when the
                    # driver broadcasts an empty input. Send an empty input to
                    # notify all other workers to stop their execution loop.
                    broadcast_tensor_dict({}, src=0)
                return None
308
            return self._get_driver_input_and_broadcast(execute_model_req)
309
        else:
310
311
312
313
            return self._get_worker_input_from_broadcast()

    def execute_model(
        self,
314
        execute_model_req: Optional[ExecuteModelRequest] = None,
315
316
317
318
319
320
321
322
    ) -> Optional[List[SamplerOutput]]:
        """Executes at least one model step on the given sequences, unless no
        sequences are provided."""
        start_time = time.perf_counter()

        inputs = self.prepare_input(execute_model_req)
        if inputs is None:
            return None
323

324
        model_input, worker_input, kwargs = inputs
325
        num_steps = worker_input.num_steps
326
327
328
329
330
331
332

        self.execute_worker(worker_input)

        # If there is no input, we don't need to execute the model.
        if worker_input.num_seq_groups == 0:
            return []

333
        intermediate_tensors = None
334
        orig_model_execute_time = 0.0
335
336
        if not get_pp_group().is_first_rank:
            intermediate_tensors = IntermediateTensors(
337
338
                get_pp_group().recv_tensor_dict(
                    all_gather_group=get_tp_group()))
339
340
341
342
            if (self.observability_config is not None
                    and self.observability_config.collect_model_execute_time):
                orig_model_execute_time = intermediate_tensors.tensors.get(
                    "model_execute_time", torch.tensor(0)).item()
343
344

        output = self.model_runner.execute_model(
345
346
347
348
349
350
351
352
            model_input=model_input,
            kv_caches=self.kv_cache[worker_input.virtual_engine]
            if self.kv_cache is not None else None,
            intermediate_tensors=intermediate_tensors,
            num_steps=num_steps,
            **kwargs,
        )

353
        model_execute_time = time.perf_counter() - start_time
354
        if not get_pp_group().is_last_rank:
355
            # output is IntermediateTensors
356
            assert isinstance(output, IntermediateTensors)
357
358
359
360
            if (self.observability_config is not None
                    and self.observability_config.collect_model_execute_time):
                output.tensors["model_execute_time"] = torch.tensor(
                    model_execute_time + orig_model_execute_time)
361
362
            get_pp_group().send_tensor_dict(output.tensors,
                                            all_gather_group=get_tp_group())
363
            return [None]
364
365
366
367
368
369
        if (self.observability_config is not None
                and self.observability_config.collect_model_execute_time
                and output is not None):
            for o in output:
                o.model_execute_time = (orig_model_execute_time +
                                        model_execute_time)
370

371
        # output is List[SamplerOutput]
372
        return output
373

374
    def _execute_model_spmd(
375
376
377
        self,
        execute_model_req: ExecuteModelRequest,
        intermediate_tensors: Optional[IntermediateTensors] = None
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
    ) -> Optional[List[SamplerOutput]]:
        """
        Execute model in Single Program Multiple Data (SPMD) fashion.
        All workers take the same request, prepare the input and
        execute the model.
        """
        assert execute_model_req is not None, (
            "_execute_model_spmd() requires each worker to take in an "
            "ExecuteModelRequest")
        worker_input: WorkerInput = self.prepare_worker_input(
            execute_model_req=execute_model_req)
        model_input: ModelRunnerInputBase = (
            self.model_runner.prepare_model_input(
                execute_model_req.seq_group_metadata_list))

        self.execute_worker(worker_input)

        # If there is no input, we don't need to execute the model.
        if worker_input.num_seq_groups == 0:
            return []

399
400
        kwargs = extract_previous_hidden_states(execute_model_req)

401
        return self.model_runner.execute_model(
402
403
404
405
406
407
            model_input=model_input,
            kv_caches=self.kv_cache[worker_input.virtual_engine]
            if self.kv_cache is not None else None,
            intermediate_tensors=intermediate_tensors,
            **kwargs,
        )
408

409

410
411
412
413
414
415
416
417
class WorkerWrapperBase:
    """
    The whole point of this class is to lazily initialize the worker.
    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`.
    """

418
419
    def __init__(
        self,
420
421
422
423
        vllm_config: VllmConfig,
    ) -> None:
        self.vllm_config = vllm_config
        trust_remote_code = vllm_config.model_config.trust_remote_code
424
        self.worker: Optional[WorkerBase] = None
425
426
427
428
        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()
429

430
431
    @staticmethod
    def update_environment_variables(envs: Dict[str, str]) -> None:
432
433
434
435
436
437
438
439
440
        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)

    def init_worker(self, *args, **kwargs):
        """
441
        Here we inject some common logic before initializing the worker.
442
443
        Arguments are passed to the worker class constructor.
        """
444
        enable_trace_function_call_for_thread(self.vllm_config)
445

446
447
448
        # see https://github.com/NVIDIA/nccl/issues/1234
        os.environ['NCCL_CUMEM_ENABLE'] = '0'

449
450
451
        from vllm.plugins import load_general_plugins
        load_general_plugins()

452
453
        worker_class = resolve_obj_by_qualname(
            self.vllm_config.parallel_config.worker_cls)
454
        self.worker = worker_class(*args, **kwargs)
455
        assert self.worker is not None
456

457
    def execute_method(self, method: str, *args, **kwargs):
458
        try:
459
460
            target = self if self.worker is None else self.worker
            executor = getattr(target, method)
461
462
463
464
465
466
467
468
469
470
            return executor(*args, **kwargs)
        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
            msg = (f"Error executing method {method}. "
                   "This might cause deadlock in distributed execution.")
            logger.exception(msg)
            raise e
471

472
473
474
    def __getattr__(self, attr):
        return getattr(self.worker, attr)

475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493

def extract_previous_hidden_states(
        data: Union[ExecuteModelRequest, Dict[str, torch.Tensor]]) -> \
            Dict[str, torch.Tensor]:
    """If data contains previous_hidden_states, extract it. This returns a dict
    which can be used directly as additional kwargs in any following 
    execute_model calls. This is used in draft models like EAGLE."""
    output = {}

    # When called from non-driver worker, data is dict but when called from
    # driver worker, data is ExecuteModelRequest.
    if isinstance(data, dict):
        if "previous_hidden_states" in data:
            output["previous_hidden_states"] = data["previous_hidden_states"]
    elif data.previous_hidden_states is not None:
        output["previous_hidden_states"] = data.previous_hidden_states\
            .hidden_states

    return output