Mooncake store get zero copy meta optimization (#9857)
This commit is contained in:
@@ -467,6 +467,7 @@ class MHATokenToKVPoolHost(HostKVCache):
|
|||||||
ptr_list = []
|
ptr_list = []
|
||||||
key_list = []
|
key_list = []
|
||||||
kv_buffer_data_ptr = self.kv_buffer.data_ptr()
|
kv_buffer_data_ptr = self.kv_buffer.data_ptr()
|
||||||
|
indices = indices.tolist()
|
||||||
v_offset = (
|
v_offset = (
|
||||||
self.layer_num
|
self.layer_num
|
||||||
* self.size
|
* self.size
|
||||||
@@ -706,6 +707,7 @@ class MLATokenToKVPoolHost(HostKVCache):
|
|||||||
ptr_list = []
|
ptr_list = []
|
||||||
key_list = []
|
key_list = []
|
||||||
kv_buffer_data_ptr = self.kv_buffer.data_ptr()
|
kv_buffer_data_ptr = self.kv_buffer.data_ptr()
|
||||||
|
indices = indices.tolist()
|
||||||
for index in range(0, len(indices), self.page_size):
|
for index in range(0, len(indices), self.page_size):
|
||||||
k_ptr = (
|
k_ptr = (
|
||||||
kv_buffer_data_ptr
|
kv_buffer_data_ptr
|
||||||
|
|||||||
Reference in New Issue
Block a user