[NPU] Support mamba cache transfer for NPU (#20364)
This commit is contained in:
@@ -34,6 +34,11 @@ class AscendKVManager(MooncakeKVManager):
|
|||||||
self.engine.batch_register(
|
self.engine.batch_register(
|
||||||
self.kv_args.aux_data_ptrs, self.kv_args.aux_data_lens
|
self.kv_args.aux_data_ptrs, self.kv_args.aux_data_lens
|
||||||
)
|
)
|
||||||
|
# Batch register state/extra pool data buffers
|
||||||
|
if self.kv_args.state_data_ptrs and self.kv_args.state_data_lens:
|
||||||
|
self.engine.batch_register(
|
||||||
|
self.kv_args.state_data_ptrs, self.kv_args.state_data_lens
|
||||||
|
)
|
||||||
|
|
||||||
def send_kvcache(
|
def send_kvcache(
|
||||||
self,
|
self,
|
||||||
|
|||||||
Reference in New Issue
Block a user