Spaces:
Sleeping
Sleeping
File size: 6,315 Bytes
ca1ecab |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 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 |
"""CacheEngine class for managing the KV cache."""
from typing import Dict, List, Tuple
import torch
from vllm._C import cache_ops
from vllm.config import CacheConfig, ModelConfig, ParallelConfig
from vllm.logger import init_logger
from vllm.utils import in_wsl, STR_DTYPE_TO_TORCH_DTYPE
logger = init_logger(__name__)
KVCache = Tuple[torch.Tensor, torch.Tensor]
class CacheEngine:
"""Manages the KV cache.
This class is responsible for initializing and managing the GPU and CPU KV
caches. It also provides methods for performing KV cache operations, such
as swapping and copying.
"""
def __init__(
self,
cache_config: CacheConfig,
model_config: ModelConfig,
parallel_config: ParallelConfig,
) -> None:
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
self.num_gpu_blocks = cache_config.num_gpu_blocks
self.num_cpu_blocks = cache_config.num_cpu_blocks
if cache_config.cache_dtype == "auto":
self.dtype = model_config.dtype
else:
self.dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
# Initialize the cache.
self.gpu_cache = self.allocate_gpu_cache()
self.cpu_cache = self.allocate_cpu_cache()
# Initialize the stream for caching operations.
self.cache_stream = torch.cuda.Stream()
assert self.cache_stream != torch.cuda.current_stream()
# Initialize the events for stream synchronization.
self.events = [torch.cuda.Event() for _ in range(self.num_layers)]
def get_key_block_shape(self) -> Tuple[int, int, int, int]:
element_size = torch.tensor([], dtype=self.dtype).element_size()
x = 16 // element_size
return (
self.num_heads,
self.head_size // x,
self.block_size,
x,
)
def get_value_block_shape(self) -> Tuple[int, int, int]:
return (
self.num_heads,
self.head_size,
self.block_size,
)
def allocate_gpu_cache(self) -> List[KVCache]:
gpu_cache: List[KVCache] = []
key_block_shape = self.get_key_block_shape()
value_block_shape = self.get_value_block_shape()
for _ in range(self.num_layers):
key_blocks = torch.empty(
size=(self.num_gpu_blocks, *key_block_shape),
dtype=self.dtype,
device="cuda",
)
value_blocks = torch.empty(
size=(self.num_gpu_blocks, *value_block_shape),
dtype=self.dtype,
device="cuda",
)
gpu_cache.append((key_blocks, value_blocks))
return gpu_cache
def allocate_cpu_cache(self) -> List[KVCache]:
cpu_cache: List[KVCache] = []
key_block_shape = self.get_key_block_shape()
value_block_shape = self.get_value_block_shape()
pin_memory = not in_wsl()
if not pin_memory:
# Pinning memory in WSL is not supported.
# https://docs.nvidia.com/cuda/wsl-user-guide/index.html#known-limitations-for-linux-cuda-applications
logger.warning("Using 'pin_memory=False' as WSL is detected. "
"This may slow down the performance.")
for _ in range(self.num_layers):
key_blocks = torch.empty(
size=(self.num_cpu_blocks, *key_block_shape),
dtype=self.dtype,
pin_memory=pin_memory,
)
value_blocks = torch.empty(
size=(self.num_cpu_blocks, *value_block_shape),
dtype=self.dtype,
pin_memory=pin_memory,
)
cpu_cache.append((key_blocks, value_blocks))
return cpu_cache
def _swap(
self,
src: List[KVCache],
dst: List[KVCache],
src_to_dst: Dict[int, int],
) -> None:
with torch.cuda.stream(self.cache_stream):
for i in range(self.num_layers):
src_key_cache, src_value_cache = src[i]
dst_key_cache, dst_value_cache = dst[i]
# Copy the key blocks.
cache_ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
# Copy the value blocks.
cache_ops.swap_blocks(src_value_cache, dst_value_cache,
src_to_dst)
event = self.events[i]
event.record(stream=self.cache_stream)
def swap_in(self, src_to_dst: Dict[int, int]) -> None:
self._swap(self.cpu_cache, self.gpu_cache, src_to_dst)
def swap_out(self, src_to_dst: Dict[int, int]) -> None:
self._swap(self.gpu_cache, self.cpu_cache, src_to_dst)
def copy(self, src_to_dsts: Dict[int, List[int]]) -> None:
key_caches = [key_cache for key_cache, _ in self.gpu_cache]
value_caches = [value_cache for _, value_cache in self.gpu_cache]
# NOTE(woosuk): This operation implicitly synchronizes the CPU and GPU.
cache_ops.copy_blocks(key_caches, value_caches, 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 = _get_dtype_size(dtype)
return dtype_size * total
def _get_dtype_size(dtype: torch.dtype) -> int:
return torch.tensor([], dtype=dtype).element_size()
|