Support page size > 1 (#4356)
This commit is contained in:
@@ -103,6 +103,9 @@ class TpModelWorkerClient:
|
||||
self.worker.model_runner.token_to_kv_pool_allocator,
|
||||
)
|
||||
|
||||
def get_kv_cache(self):
|
||||
return self.worker.model_runner.token_to_kv_pool
|
||||
|
||||
def forward_thread_func(self):
|
||||
try:
|
||||
with torch.get_device_module(self.device).stream(self.forward_stream):
|
||||
@@ -203,7 +206,7 @@ class TpModelWorkerClient:
|
||||
-(self.future_token_ids_ct + 1),
|
||||
-(self.future_token_ids_ct + 1 + bs),
|
||||
-1,
|
||||
dtype=torch.int32,
|
||||
dtype=torch.int64,
|
||||
device=self.device,
|
||||
)
|
||||
self.future_token_ids_ct = (
|
||||
|
||||
Reference in New Issue
Block a user