[Misc] Remove cache stream and cache events (#3461)

This commit is contained in:
Woosuk Kwon
2024-03-20 00:38:53 -07:00
committed by GitHub
parent 4ad521d8b5
commit 5ee14494e4
3 changed files with 86 additions and 32 deletions
+8 -18
View File
@@ -38,7 +38,7 @@ class CacheEngine:
self.num_gpu_blocks = cache_config.num_gpu_blocks
self.num_cpu_blocks = cache_config.num_cpu_blocks
# Skip initializing CUDA stream and buffer for Neuron backend.
# Skip initializing KV cache for Neuron backend.
if is_neuron():
return
@@ -51,12 +51,6 @@ class CacheEngine:
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
@@ -126,17 +120,13 @@ class CacheEngine:
) -> None:
from vllm._C import cache_ops
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)
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)
def swap_in(self, src_to_dst: Dict[int, int]) -> None:
self._swap(self.cpu_cache, self.gpu_cache, src_to_dst)
+1 -14
View File
@@ -65,7 +65,6 @@ class Worker:
# self.init_cache_engine().
self.cache_config = None
self.cache_engine = None
self.cache_events = None
self.gpu_cache = None
def init_model(self, cupy_port: Optional[int] = None) -> None:
@@ -148,7 +147,6 @@ class Worker:
self.cache_config = cache_config
self.cache_engine = CacheEngine(self.cache_config, self.model_config,
self.parallel_config)
self.cache_events = self.cache_engine.events
self.gpu_cache = self.cache_engine.gpu_cache
self.model_runner.set_block_size(self.cache_engine.block_size)
@@ -166,24 +164,13 @@ class Worker:
blocks_to_copy: Dict[int, List[int]],
) -> None:
# Issue cache operations.
issued_cache_op = False
# TODO(woosuk): Profile swapping overhead and optimize if needed.
if blocks_to_swap_in:
self.cache_engine.swap_in(blocks_to_swap_in)
issued_cache_op = True
if blocks_to_swap_out:
self.cache_engine.swap_out(blocks_to_swap_out)
issued_cache_op = True
if blocks_to_copy:
self.cache_engine.copy(blocks_to_copy)
issued_cache_op = True
cache_events = self.cache_events if issued_cache_op else None
# Wait for cache operations to finish.
# TODO(woosuk): Profile swapping overhead and optimize if needed.
if cache_events is not None:
for event in cache_events:
event.wait()
@torch.inference_mode()
def execute_model(