[Core] Refactor Attention Take 2 (#3462)
This commit is contained in:
+30
-83
@@ -1,16 +1,15 @@
|
||||
"""CacheEngine class for managing the KV cache."""
|
||||
from typing import Dict, List, Tuple
|
||||
from typing import Dict, List
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.attention import get_attn_backend
|
||||
from vllm.config import CacheConfig, ModelConfig, ParallelConfig
|
||||
from vllm.logger import init_logger
|
||||
from vllm.utils import is_pin_memory_available, STR_DTYPE_TO_TORCH_DTYPE
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
KVCache = Tuple[torch.Tensor, torch.Tensor]
|
||||
|
||||
|
||||
class CacheEngine:
|
||||
"""Manages the KV cache.
|
||||
@@ -43,95 +42,43 @@ class CacheEngine:
|
||||
else:
|
||||
self.dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
|
||||
|
||||
# Get attention backend.
|
||||
self.attn_backend = get_attn_backend(model_config.dtype)
|
||||
|
||||
# Initialize the cache.
|
||||
self.gpu_cache = self.allocate_gpu_cache()
|
||||
self.cpu_cache = self.allocate_cpu_cache()
|
||||
self.gpu_cache = self._allocate_kv_cache(self.num_gpu_blocks, "cuda")
|
||||
self.cpu_cache = self._allocate_kv_cache(self.num_cpu_blocks, "cpu")
|
||||
|
||||
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 = is_pin_memory_available()
|
||||
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,
|
||||
device="cpu",
|
||||
)
|
||||
value_blocks = torch.empty(
|
||||
size=(self.num_cpu_blocks, *value_block_shape),
|
||||
dtype=self.dtype,
|
||||
pin_memory=pin_memory,
|
||||
device="cpu",
|
||||
)
|
||||
cpu_cache.append((key_blocks, value_blocks))
|
||||
return cpu_cache
|
||||
|
||||
def _swap(
|
||||
def _allocate_kv_cache(
|
||||
self,
|
||||
src: List[KVCache],
|
||||
dst: List[KVCache],
|
||||
src_to_dst: Dict[int, int],
|
||||
) -> None:
|
||||
from vllm._C import cache_ops
|
||||
|
||||
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)
|
||||
num_blocks: int,
|
||||
device: str,
|
||||
) -> List[torch.Tensor]:
|
||||
"""Allocates KV cache on the specified device."""
|
||||
kv_cache_shape = self.attn_backend.get_kv_cache_shape(
|
||||
num_blocks, self.block_size, self.num_heads, self.head_size)
|
||||
pin_memory = is_pin_memory_available() if device == "cpu" else False
|
||||
kv_cache: List[torch.Tensor] = []
|
||||
for _ in range(self.num_layers):
|
||||
kv_cache.append(
|
||||
torch.empty(kv_cache_shape,
|
||||
dtype=self.dtype,
|
||||
pin_memory=pin_memory,
|
||||
device=device))
|
||||
return kv_cache
|
||||
|
||||
def swap_in(self, src_to_dst: Dict[int, int]) -> None:
|
||||
self._swap(self.cpu_cache, self.gpu_cache, src_to_dst)
|
||||
for i in range(self.num_layers):
|
||||
self.attn_backend.swap_blocks(self.cpu_cache[i], self.gpu_cache[i],
|
||||
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)
|
||||
for i in range(self.num_layers):
|
||||
self.attn_backend.swap_blocks(self.gpu_cache[i], self.cpu_cache[i],
|
||||
src_to_dst)
|
||||
|
||||
def copy(self, src_to_dsts: Dict[int, List[int]]) -> None:
|
||||
from vllm._C import cache_ops
|
||||
|
||||
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)
|
||||
self.attn_backend.copy_blocks(self.gpu_cache, src_to_dsts)
|
||||
|
||||
@staticmethod
|
||||
def get_cache_block_size(
|
||||
|
||||
Reference in New Issue
Block a user