worker.py 13.7 KB
Newer Older
1
"""A GPU worker class."""
2
import gc
3
import os
4
from typing import Dict, List, Optional, Set, Tuple
Woosuk Kwon's avatar
Woosuk Kwon committed
5
6

import torch
7
import torch.distributed
Woosuk Kwon's avatar
Woosuk Kwon committed
8

9
from vllm.config import (CacheConfig, DeviceConfig, LoRAConfig, ModelConfig,
10
                         ParallelConfig, SchedulerConfig, VisionLanguageConfig)
11
from vllm.lora.request import LoRARequest
12
from vllm.model_executor import set_random_seed
13
from vllm.model_executor.parallel_utils import pynccl_utils
14
from vllm.model_executor.parallel_utils.communication_op import (
15
    broadcast_tensor_dict)
16
from vllm.model_executor.parallel_utils.custom_all_reduce import init_custom_ar
Woosuk Kwon's avatar
Woosuk Kwon committed
17
from vllm.model_executor.parallel_utils.parallel_state import (
18
    ensure_model_parallel_initialized, init_distributed_environment)
19
from vllm.sequence import SamplerOutput, SequenceGroupMetadata
Woosuk Kwon's avatar
Woosuk Kwon committed
20
from vllm.worker.cache_engine import CacheEngine
21
from vllm.worker.model_runner import ModelRunner
22
from vllm.worker.worker_base import WorkerBase
Woosuk Kwon's avatar
Woosuk Kwon committed
23

24

25
class Worker(WorkerBase):
26
27
28
29
30
31
    """A worker class that executes (a partition of) the model on a GPU.

    Each worker is associated with a single GPU. The worker is responsible for
    maintaining the KV cache and executing the model on the GPU. In case of
    distributed inference, each worker is assigned a partition of the model.
    """
Woosuk Kwon's avatar
Woosuk Kwon committed
32
33
34

    def __init__(
        self,
35
36
37
        model_config: ModelConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
38
        device_config: DeviceConfig,
39
        cache_config: CacheConfig,
40
41
42
        local_rank: int,
        rank: int,
        distributed_init_method: str,
43
        lora_config: Optional[LoRAConfig] = None,
44
        vision_language_config: Optional[VisionLanguageConfig] = None,
45
        is_driver_worker: bool = False,
Woosuk Kwon's avatar
Woosuk Kwon committed
46
    ) -> None:
47
48
49
        self.model_config = model_config
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
50
        self.device_config = device_config
51
        self.cache_config = cache_config
52
        self.local_rank = local_rank
53
54
        self.rank = rank
        self.distributed_init_method = distributed_init_method
55
        self.lora_config = lora_config
56
57
58
        self.is_driver_worker = is_driver_worker
        if self.is_driver_worker:
            assert self.rank == 0, "The driver worker must have rank 0."
59

60
61
62
63
64
65
66
67
68
69
70
        self.vision_language_config = vision_language_config
        if self.vision_language_config:
            assert not self.lora_config, (
                "To be tested: vision language model with LoRA settings.")

        self.model_runner = ModelRunner(
            model_config,
            parallel_config,
            scheduler_config,
            device_config,
            lora_config=self.lora_config,
71
            kv_cache_dtype=self.cache_config.cache_dtype,
72
73
            is_driver_worker=is_driver_worker,
            vision_language_config=vision_language_config)
74
        # Uninitialized cache engine. Will be initialized by
75
        # initialize_cache.
76
77
78
        self.cache_engine = None
        self.gpu_cache = None

79
    def init_device(self) -> None:
80
81
82
83
84
85
86
87
        if self.device_config.device.type == "cuda":
            # torch.distributed.all_reduce does not free the input tensor until
            # the synchronization point. This causes the memory usage to grow
            # as the number of all_reduce calls increases. This env var disables
            # this behavior.
            # Related issue:
            # https://discuss.pytorch.org/t/cuda-allocation-lifetime-for-inputs-to-distributed-all-reduce/191573
            os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
88

89
90
91
92
            # This env var set by Ray causes exceptions with graph building.
            os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
            self.device = torch.device(f"cuda:{self.local_rank}")
            torch.cuda.set_device(self.device)
93

94
            _check_if_gpu_supports_dtype(self.model_config.dtype)
95
96
            torch.cuda.empty_cache()
            self.init_gpu_memory = torch.cuda.mem_get_info()[0]
97
98
99
        else:
            raise RuntimeError(
                f"Not support device type: {self.device_config.device}")
100
        # Initialize the distributed environment.
101
102
103
        init_worker_distributed_environment(self.parallel_config, self.rank,
                                            self.distributed_init_method,
                                            self.local_rank)
104
        # Set random seed.
105
        set_random_seed(self.model_config.seed)
106
107

    def load_model(self):
108
        self.model_runner.load_model()
109

110
    @torch.inference_mode()
111
112
113
114
115
116
117
118
119
120
121
    def determine_num_available_blocks(self) -> Tuple[int, int]:
        """Profiles the peak memory usage of the model to determine how many
        KV blocks may be allocated without OOMs.

        The engine will first conduct a profiling of the existing memory usage.
        Then, it calculate the maximum possible number of GPU and CPU blocks
        that can be allocated with the remaining free memory.

        .. tip::
            You may limit the usage of GPU memory
            by adjusting the `gpu_memory_utilization` parameter.
122
        """
123
124
125
126
        # Profile the memory usage of the model and get the maximum number of
        # cache blocks that can be allocated with the remaining free memory.
        torch.cuda.empty_cache()

127
128
129
        # Execute a forward pass with dummy inputs to profile the memory usage
        # of the model.
        self.model_runner.profile_run()
130
131
132
133

        # Calculate the number of blocks that can be allocated with the
        # profiled peak memory.
        torch.cuda.synchronize()
134
        free_gpu_memory, total_gpu_memory = torch.cuda.mem_get_info()
135
136
137
        # NOTE(woosuk): Here we assume that the other processes using the same
        # GPU did not change their memory usage during the profiling.
        peak_memory = self.init_gpu_memory - free_gpu_memory
138
139
140
        assert peak_memory > 0, (
            "Error in memory profiling. This happens when the GPU memory was "
            "not properly cleaned up before initializing the vLLM instance.")
141

142
        cache_block_size = self.get_cache_block_size_bytes()
143
        num_gpu_blocks = int(
144
145
146
147
            (total_gpu_memory * self.cache_config.gpu_memory_utilization -
             peak_memory) // cache_block_size)
        num_cpu_blocks = int(self.cache_config.swap_space_bytes //
                             cache_block_size)
148
149
        num_gpu_blocks = max(num_gpu_blocks, 0)
        num_cpu_blocks = max(num_cpu_blocks, 0)
150
151
152
        if self.model_runner.lora_manager:
            self.model_runner.remove_all_loras()
        gc.collect()
153
154
155
        torch.cuda.empty_cache()
        return num_gpu_blocks, num_cpu_blocks

156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
    def initialize_cache(self, num_gpu_blocks: int,
                         num_cpu_blocks: int) -> None:
        """Allocate GPU and CPU KV cache with the specified number of blocks.

        This also warms up the model, which may record CUDA graphs.
        """
        raise_if_cache_size_invalid(num_gpu_blocks,
                                    self.cache_config.block_size,
                                    self.model_config.max_model_len)

        self.cache_config.num_gpu_blocks = num_gpu_blocks
        self.cache_config.num_cpu_blocks = num_cpu_blocks

        self._init_cache_engine()
        self._warm_up_model()

    def _init_cache_engine(self):
        assert self.cache_config.num_gpu_blocks is not None
174
175
        self.cache_engine = CacheEngine(self.cache_config, self.model_config,
                                        self.parallel_config)
Woosuk Kwon's avatar
Woosuk Kwon committed
176
        self.gpu_cache = self.cache_engine.gpu_cache
177
        self.model_runner.set_block_size(self.cache_engine.block_size)
Woosuk Kwon's avatar
Woosuk Kwon committed
178

179
    def _warm_up_model(self) -> None:
180
181
182
183
184
185
        if not self.model_config.enforce_eager:
            self.model_runner.capture_model(self.gpu_cache)
        # Reset the seed to ensure that the random state is not affected by
        # the model initialization and profiling.
        set_random_seed(self.model_config.seed)

186
    def cache_swap(
Woosuk Kwon's avatar
Woosuk Kwon committed
187
188
189
        self,
        blocks_to_swap_in: Dict[int, int],
        blocks_to_swap_out: Dict[int, int],
190
        blocks_to_copy: Dict[int, List[int]],
191
    ) -> None:
Woosuk Kwon's avatar
Woosuk Kwon committed
192
        # Issue cache operations.
193
        # TODO(woosuk): Profile swapping overhead and optimize if needed.
Woosuk Kwon's avatar
Woosuk Kwon committed
194
195
196
197
198
199
        if blocks_to_swap_in:
            self.cache_engine.swap_in(blocks_to_swap_in)
        if blocks_to_swap_out:
            self.cache_engine.swap_out(blocks_to_swap_out)
        if blocks_to_copy:
            self.cache_engine.copy(blocks_to_copy)
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214

    @torch.inference_mode()
    def execute_model(
        self,
        seq_group_metadata_list: Optional[List[SequenceGroupMetadata]] = None,
        blocks_to_swap_in: Optional[Dict[int, int]] = None,
        blocks_to_swap_out: Optional[Dict[int, int]] = None,
        blocks_to_copy: Optional[Dict[int, List[int]]] = None,
    ) -> Optional[SamplerOutput]:
        if self.is_driver_worker:
            assert seq_group_metadata_list is not None
            num_seq_groups = len(seq_group_metadata_list)
            assert blocks_to_swap_in is not None
            assert blocks_to_swap_out is not None
            assert blocks_to_copy is not None
215
216
217
218
219
220
221
            data = {
                "num_seq_groups": num_seq_groups,
                "blocks_to_swap_in": blocks_to_swap_in,
                "blocks_to_swap_out": blocks_to_swap_out,
                "blocks_to_copy": blocks_to_copy,
            }
            broadcast_tensor_dict(data, src=0)
222
        else:
223
224
225
226
227
228
229
            data = broadcast_tensor_dict(src=0)
            num_seq_groups = data["num_seq_groups"]
            blocks_to_swap_in = data["blocks_to_swap_in"]
            blocks_to_swap_out = data["blocks_to_swap_out"]
            blocks_to_copy = data["blocks_to_copy"]

        self.cache_swap(blocks_to_swap_in, blocks_to_swap_out, blocks_to_copy)
230

Woosuk Kwon's avatar
Woosuk Kwon committed
231
        # If there is no input, we don't need to execute the model.
232
        if num_seq_groups == 0:
Woosuk Kwon's avatar
Woosuk Kwon committed
233
234
            return {}

235
        output = self.model_runner.execute_model(seq_group_metadata_list,
236
                                                 self.gpu_cache)
Woosuk Kwon's avatar
Woosuk Kwon committed
237
238
        return output

239
240
241
242
243
244
245
246
247
    def add_lora(self, lora_request: LoRARequest) -> bool:
        return self.model_runner.add_lora(lora_request)

    def remove_lora(self, lora_id: int) -> bool:
        return self.model_runner.remove_lora(lora_id)

    def list_loras(self) -> Set[int]:
        return self.model_runner.list_loras()

248
249
250
251
252
253
254
255
    @property
    def max_model_len(self) -> int:
        return self.model_config.max_model_len

    @property
    def vocab_size(self) -> int:
        return self.model_runner.vocab_size

256
    def get_cache_block_size_bytes(self) -> int:
257
258
        """Get the size of the KV cache block size in bytes.
        """
259
        return CacheEngine.get_cache_block_size(self.cache_config,
260
261
262
                                                self.model_config,
                                                self.parallel_config)

Woosuk Kwon's avatar
Woosuk Kwon committed
263

264
def init_worker_distributed_environment(
265
266
    parallel_config: ParallelConfig,
    rank: int,
267
    distributed_init_method: Optional[str] = None,
268
    local_rank: int = -1,
269
270
) -> None:
    """Initialize the distributed environment."""
271
272
    init_distributed_environment(parallel_config.world_size, rank,
                                 distributed_init_method, local_rank)
273

274
275
276
    if pynccl_utils.is_initialized():
        pynccl_world_size = pynccl_utils.get_world_size()
        if pynccl_world_size != parallel_config.world_size:
Woosuk Kwon's avatar
Woosuk Kwon committed
277
            raise RuntimeError(
278
                "pynccl is already initialized but the pynccl world "
Woosuk Kwon's avatar
Woosuk Kwon committed
279
                "size does not match parallel_config.world_size "
280
281
282
                f"({pynccl_world_size} vs. {parallel_config.world_size}).")
    elif parallel_config.world_size > 1:
        # NOTE(woosuk): We don't initialize pynccl process group when world size
Woosuk Kwon's avatar
Woosuk Kwon committed
283
        # is 1.
284
        pynccl_utils.init_process_group(
Woosuk Kwon's avatar
Woosuk Kwon committed
285
            world_size=parallel_config.world_size,
286
            local_rank=local_rank,
Woosuk Kwon's avatar
Woosuk Kwon committed
287
            rank=rank,
288
            init_method=distributed_init_method,
Woosuk Kwon's avatar
Woosuk Kwon committed
289
290
        )

291
292
    ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
                                      parallel_config.pipeline_parallel_size)
293

294
295
296
297
    # Initialize a custom fast all-reduce implementation.
    if not parallel_config.disable_custom_all_reduce:
        init_custom_ar()

298
299
300
301
302
    # A small all_reduce for warmup.
    torch.distributed.all_reduce(torch.zeros(1).cuda())
    if pynccl_utils.is_initialized():
        pynccl_utils.all_reduce(torch.zeros(1).cuda())

303

304
305
306
307
308
309
310
311
312
def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):
    # Check if the GPU supports the dtype.
    if torch_dtype == torch.bfloat16:
        compute_capability = torch.cuda.get_device_capability()
        if compute_capability[0] < 8:
            gpu_name = torch.cuda.get_device_name()
            raise ValueError(
                "Bfloat16 is only supported on GPUs with compute capability "
                f"of at least 8.0. Your {gpu_name} GPU has compute capability "
Woosuk Kwon's avatar
Woosuk Kwon committed
313
314
315
                f"{compute_capability[0]}.{compute_capability[1]}. "
                "You can use float16 instead by explicitly setting the"
                "`dtype` flag in CLI, for example: --dtype=half.")
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331


def raise_if_cache_size_invalid(num_gpu_blocks, block_size,
                                max_model_len) -> None:
    if num_gpu_blocks <= 0:
        raise ValueError("No available memory for the cache blocks. "
                         "Try increasing `gpu_memory_utilization` when "
                         "initializing the engine.")
    max_seq_len = block_size * num_gpu_blocks
    if max_model_len > max_seq_len:
        raise ValueError(
            f"The model's max seq len ({max_model_len}) "
            "is larger than the maximum number of tokens that can be "
            f"stored in KV cache ({max_seq_len}). Try increasing "
            "`gpu_memory_utilization` or decreasing `max_model_len` when "
            "initializing the engine.")