[Misc] Remove cache stream and cache events (#3461)
This commit is contained in:
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user