Move swa memory pool to a seperate file (#16347)
This commit is contained in:
@@ -10,7 +10,6 @@ from sglang.srt.distributed.parallel_state import get_world_group
|
||||
from sglang.srt.layers.dp_attention import get_attention_tp_size
|
||||
from sglang.srt.mem_cache.allocator import (
|
||||
PagedTokenToKVPoolAllocator,
|
||||
SWATokenToKVPoolAllocator,
|
||||
TokenToKVPoolAllocator,
|
||||
)
|
||||
from sglang.srt.mem_cache.memory_pool import (
|
||||
@@ -23,8 +22,8 @@ from sglang.srt.mem_cache.memory_pool import (
|
||||
MLATokenToKVPoolFP4,
|
||||
NSATokenToKVPool,
|
||||
ReqToTokenPool,
|
||||
SWAKVPool,
|
||||
)
|
||||
from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool, SWATokenToKVPoolAllocator
|
||||
from sglang.srt.utils.common import (
|
||||
get_available_gpu_memory,
|
||||
is_float4_e2m1fn_x2,
|
||||
|
||||
Reference in New Issue
Block a user