cpu_worker.py 14.6 KB
Newer Older
1
"""A CPU worker class."""
2
from typing import Dict, List, Optional, Tuple, Type
3
4
5
6

import torch
import torch.distributed

7
import vllm.envs as envs
8
from vllm.attention import get_attn_backend
9
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
10
11
                         ModelConfig, ParallelConfig, PromptAdapterConfig,
                         SchedulerConfig)
12
from vllm.distributed import (ensure_model_parallel_initialized,
13
                              init_distributed_environment)
14
15
from vllm.logger import init_logger
from vllm.model_executor import set_random_seed
16
from vllm.sequence import ExecuteModelRequest
17
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
18
from vllm.worker.cpu_enc_dec_model_runner import CPUEncoderDecoderModelRunner
19
from vllm.worker.cpu_model_runner import CPUModelRunner
20
21
from vllm.worker.worker_base import (LocalOrDistributedWorkerBase,
                                     LoraNotSupportedWorkerBase, WorkerInput)
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

logger = init_logger(__name__)


class CPUCacheEngine:
    """Manages the KV cache for CPU backend.

    This class is responsible for initializing and managing CPU KV
    caches. It also provides methods for performing KV cache operations, such
    as copying.
    """

    def __init__(self, cache_config: CacheConfig, model_config: ModelConfig,
                 parallel_config: ParallelConfig,
                 device_config: DeviceConfig) -> None:
        assert device_config.device_type == "cpu"
        self.cache_config = cache_config
        self.model_config = model_config
        self.parallel_config = parallel_config

        self.head_size = model_config.get_head_size()
        self.num_layers = model_config.get_num_layers(parallel_config)
        self.num_heads = model_config.get_num_kv_heads(parallel_config)

        self.block_size = cache_config.block_size
        # Note: In CacheConfig, num_gpu_blocks actual is num_cpu_blocks
        # for CPU backend, because we want to reuse KV cache management
        # in the scheduler.
        self.num_cpu_blocks = cache_config.num_gpu_blocks

        if cache_config.cache_dtype == "auto":
            self.dtype = model_config.dtype
        else:
            self.dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]

        # Get attention backend.
58
59
60
61
62
63
64
65
66
        self.attn_backend = get_attn_backend(
            self.model_config.get_num_attention_heads(self.parallel_config),
            self.model_config.get_head_size(),
            self.model_config.get_num_kv_heads(self.parallel_config),
            self.model_config.get_sliding_window(),
            self.model_config.dtype,
            cache_config.cache_dtype,
            self.block_size,
        )
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

        # Initialize the cache.
        self.cpu_cache = self._allocate_kv_cache(self.num_cpu_blocks)

    def _allocate_kv_cache(
        self,
        num_blocks: int,
    ) -> List[torch.Tensor]:
        """Allocates KV cache on CPU."""
        kv_cache_shape = self.attn_backend.get_kv_cache_shape(
            num_blocks, self.block_size, self.num_heads, self.head_size)
        kv_cache: List[torch.Tensor] = []
        for _ in range(self.num_layers):
            kv_cache.append(
                torch.empty(kv_cache_shape, dtype=self.dtype, device="cpu"))
        return kv_cache

    def swap_in(self, src_to_dst: Dict[int, int]) -> None:
        raise NotImplementedError("Swap is not supported in CPUCacheEngine.")

    def swap_out(self, src_to_dst: Dict[int, int]) -> None:
        raise NotImplementedError("Swap is not supported in CPUCacheEngine.")

    def copy(self, src_to_dsts: Dict[int, List[int]]) -> None:
        self.attn_backend.copy_blocks(self.cpu_cache, src_to_dsts)

    @staticmethod
    def get_cache_block_size(
        block_size: int,
        cache_dtype: str,
        model_config: ModelConfig,
        parallel_config: ParallelConfig,
    ) -> int:
        head_size = model_config.get_head_size()
        num_heads = model_config.get_num_kv_heads(parallel_config)
        num_layers = model_config.get_num_layers(parallel_config)

        key_cache_block = block_size * num_heads * head_size
        value_cache_block = key_cache_block
        total = num_layers * (key_cache_block + value_cache_block)
        if cache_dtype == "auto":
            dtype = model_config.dtype
        else:
            dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_dtype]
        dtype_size = torch.tensor([], dtype=dtype).element_size()
        return dtype_size * total


115
class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
116
117
118
119
120
121
122
123
124
125
126
127
128
129
    """A worker class that executes (a partition of) the model on a CPU socket.

    Each worker is associated with a single CPU socket. The worker is 
    responsible for maintaining the KV cache and executing the model on the 
    CPU. In case of distributed inference, each worker is assigned a partition
    of the model.
    """

    def __init__(
        self,
        model_config: ModelConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
        device_config: DeviceConfig,
130
        cache_config: CacheConfig,
131
        load_config: LoadConfig,
132
133
134
135
136
        local_rank: int,
        rank: int,
        distributed_init_method: str,
        lora_config: Optional[LoRAConfig] = None,
        kv_cache_dtype: Optional[str] = "auto",
137
        prompt_adapter_config: Optional[PromptAdapterConfig] = None,
138
139
140
141
142
143
        is_driver_worker: bool = False,
    ) -> None:
        self.model_config = model_config
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
        self.device_config = device_config
144
        self.cache_config = cache_config
145
        self.load_config = load_config
146
147
148
149
        self.local_rank = local_rank
        self.rank = rank
        self.distributed_init_method = distributed_init_method
        self.lora_config = lora_config
150
        self.prompt_adapter_config = prompt_adapter_config
151
152
153
        self.is_driver_worker = is_driver_worker
        if self.is_driver_worker:
            assert self.rank == 0, "The driver worker must have rank 0."
154

155
156
157
158
        if self.model_config.trust_remote_code:
            # note: lazy import to avoid importing torch before initializing
            from vllm.utils import init_cached_hf_modules
            init_cached_hf_modules()
159
160
161
162
163
164
165
166

        # Setup OpenMP threads affinity.
        omp_cpuids = envs.VLLM_CPU_OMP_THREADS_BIND
        if omp_cpuids == "all":
            self.local_omp_cpuid = "all"
        else:
            self.local_omp_cpuid = omp_cpuids.split("|")[rank]

167
168
169
170
        ModelRunnerClass: Type[CPUModelRunner] = CPUModelRunner
        if self._is_encoder_decoder_model():
            ModelRunnerClass = CPUEncoderDecoderModelRunner
        self.model_runner: CPUModelRunner = ModelRunnerClass(
171
172
173
174
            model_config,
            parallel_config,
            scheduler_config,
            device_config,
175
            cache_config,
176
177
178
            load_config=self.load_config,
            lora_config=self.lora_config,
            kv_cache_dtype=kv_cache_dtype,
179
            prompt_adapter_config=self.prompt_adapter_config,
180
            is_driver_worker=is_driver_worker)
181
        # Uninitialized cache engine. Will be initialized by
182
        # initialize_cache.
183
184
        self.cache_engine: List[CPUCacheEngine]
        self.cpu_cache: List[List[torch.Tensor]]
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
        # Torch profiler. Enabled and configured through env vars:
        # VLLM_TORCH_PROFILER_DIR=/path/to/save/trace
        if envs.VLLM_TORCH_PROFILER_DIR:
            torch_profiler_trace_dir = envs.VLLM_TORCH_PROFILER_DIR
            logger.info("Profiling enabled. Traces will be saved to: %s",
                        torch_profiler_trace_dir)
            self.profiler = torch.profiler.profile(
                activities=[
                    torch.profiler.ProfilerActivity.CPU,
                ],
                with_stack=True,
                on_trace_ready=torch.profiler.tensorboard_trace_handler(
                    torch_profiler_trace_dir, use_gzip=True))
        else:
            self.profiler = None

    def start_profile(self):
        if self.profiler is None:
            raise RuntimeError("Profiler is not enabled.")
        self.profiler.start()

    def stop_profile(self):
        if self.profiler is None:
            raise RuntimeError("Profiler is not enabled.")
        self.profiler.stop()

212
213
214
    def _is_encoder_decoder_model(self):
        return self.model_config.is_encoder_decoder_model

215
    def init_device(self) -> None:
216
        if self.local_omp_cpuid != "all":
217
            ret = torch.ops._C_utils.init_cpu_threads_env(self.local_omp_cpuid)
218
219
            if ret:
                logger.info(ret)
220

221
222
223
224
225
226
227
        self.init_distributed_environment()
        # Set random seed.
        set_random_seed(self.model_config.seed)

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

228
    def determine_num_available_blocks(self) -> Tuple[int, int]:
229
230
231
232
233
234
235
236
237
        """Determine the number of blocks available for the KV cache.

        This determines how many KV blocks can fit into the configured CPU
        KV cache space.

        Note that since vLLM assumes a block resides on GPU if it can be
        modified, we return num_gpu_blocks=num_cpu_blocks and num_cpu_blocks=0.
        This allows us to reuse the scheduler of vLLM without generalizing it
        to different devices.
238
239
240
        """
        # For CPU device, the block number will be calculated based on the
        # cpu_kvcache_space.
241
242
243
        cache_block_size = self.get_cache_block_size_bytes()
        num_cpu_blocks = int(self.cache_config.cpu_kvcache_space_bytes //
                             cache_block_size)
244
245
        num_cpu_blocks = max(num_cpu_blocks, 0)

246
247
248
249
250
        # Note: To reuse the cache management procedure,
        # use cpu cache as 'gpu cache'.
        num_gpu_blocks = num_cpu_blocks
        num_cpu_blocks = 0
        return num_gpu_blocks, num_cpu_blocks
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
    def initialize_cache(self, num_gpu_blocks: int,
                         num_cpu_blocks: int) -> None:
        """Initialize the KV cache. Currently, swappable CPU memory is not
        supported.

        Since this worker does not support GPUs, we use the num_gpu_blocks to
        determine how many non-swappable CPU blocks to allocate.
        """
        assert (num_cpu_blocks == 0
                ), f"{type(self)} does not support swappable cache"

        # Note: To reuse the cache management procedure,
        # use cpu cache as 'gpu cache'.
        num_cpu_blocks = num_gpu_blocks

        self._validate_num_cpu_blocks(num_cpu_blocks)
        self.cache_config.num_gpu_blocks = num_cpu_blocks
        self.cache_config.num_cpu_blocks = 0

        # Initialize the cache.
        self._init_cache_engine()

    def _validate_num_cpu_blocks(self, num_cpu_blocks: int) -> None:
        """Raise errors if the num_cpu_blocks is invalid.
        """
        if num_cpu_blocks <= 0:
            raise ValueError("No available memory for the cache blocks. "
                             "Try increasing `VLLM_CPU_KVCACHE_SPACE` when "
                             "initializing the engine.")

        max_seq_len = self.cache_config.block_size * num_cpu_blocks
        if self.model_config.max_model_len > max_seq_len:
            raise ValueError(
                f"The model's max seq len ({self.model_config.max_model_len}) "
                "is larger than the maximum number of tokens that can be "
                f"stored in KV cache ({max_seq_len}). Try increasing "
                "`VLLM_CPU_KVCACHE_SPACE` or decreasing `max_model_len` when "
                "initializing the engine.")

    def _init_cache_engine(self) -> None:
292
293
294
295
296
297
298
299
300
301
302
303
304
305
        self.cache_engine = [
            CPUCacheEngine(self.cache_config, self.model_config,
                           self.parallel_config, self.device_config)
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]
        self.cpu_cache = [
            self.cache_engine[ve].cpu_cache
            for ve in range(self.parallel_config.pipeline_parallel_size)
        ]
        self.model_runner.block_size = self.cache_engine[0].block_size

        assert all(
            self.cpu_cache[ve] is not None
            for ve in range(self.parallel_config.pipeline_parallel_size))
306
307

        # Populate the cache to warmup the memory
308
309
310
        for ve in range(self.parallel_config.pipeline_parallel_size):
            for layer_cache in self.cpu_cache[ve]:
                layer_cache.fill_(0)
311

312
313
314
315
316
    @property
    def do_metadata_broadcast(self) -> bool:
        return self.parallel_config.tensor_parallel_size > 1

    @property
317
    def kv_cache(self) -> Optional[List[List[torch.Tensor]]]:
318
319
320
        return self.cpu_cache

    def execute_worker(
321
        self,
322
        worker_input: WorkerInput,
323
    ) -> None:
324
325
        if (worker_input.blocks_to_copy is not None
                and worker_input.blocks_to_copy.numel() > 0):
326
327
            self.cache_engine[worker_input.virtual_engine].copy(
                worker_input.blocks_to_copy)
328
329

    @torch.inference_mode()
330
331
332
    def prepare_worker_input(
            self, execute_model_req: ExecuteModelRequest) -> WorkerInput:
        assert execute_model_req is not None
333
        virtual_engine = execute_model_req.virtual_engine
334
335
336
337
338
339
340
341
342
343
        num_seq_groups: int = len(execute_model_req.seq_group_metadata_list)
        blocks_to_copy = execute_model_req.blocks_to_copy
        blocks_to_copy = torch.tensor(execute_model_req.blocks_to_copy,
                                      device="cpu",
                                      dtype=torch.int64).view(-1, 2)
        assert len(execute_model_req.blocks_to_swap_in) == 0
        assert len(execute_model_req.blocks_to_swap_out) == 0
        return WorkerInput(
            num_seq_groups=num_seq_groups,
            blocks_to_copy=blocks_to_copy,
344
            virtual_engine=virtual_engine,
345
        )
346
347
348
349
350
351
352

    def init_distributed_environment(self) -> None:
        """Initialize the distributed environment."""

        parallel_config = self.parallel_config
        rank = self.rank
        distributed_init_method = self.distributed_init_method
353
354
355
356
357
358
        init_distributed_environment(
            world_size=parallel_config.world_size,
            rank=rank,
            distributed_init_method=distributed_init_method,
            backend="gloo",
        )
359
360
361
362
363
364
365

        # A small all_reduce for warmup.
        torch.distributed.all_reduce(torch.zeros(1).cpu())

        ensure_model_parallel_initialized(
            parallel_config.tensor_parallel_size,
            parallel_config.pipeline_parallel_size)
366
367
368
369
370
371
372

    def get_cache_block_size_bytes(self) -> int:
        """Return the size in bytes of a single KV cache block.
        """
        return CPUCacheEngine.get_cache_block_size(
            self.cache_config.block_size, self.cache_config.cache_dtype,
            self.model_config, self.parallel_config)