Mooncake store use adxl inferface (#3350)
Use adxl inferface in mooncake store, mooncake PR https://github.com/kvcache-ai/Mooncake/pull/929 - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 --------- Signed-off-by: LCAIZJ <leichao139636@163.com>
This commit is contained in:
28
vllm_ascend/distributed/mooncake/transfer_engine.py
Normal file
28
vllm_ascend/distributed/mooncake/transfer_engine.py
Normal file
@@ -0,0 +1,28 @@
|
||||
import threading
|
||||
from typing import Optional
|
||||
|
||||
from mooncake.engine import TransferEngine # type: ignore
|
||||
|
||||
_global_te = None
|
||||
_global_te_lock = threading.Lock()
|
||||
|
||||
|
||||
def get_global_te(hostname: str, device_name: Optional[str]):
|
||||
global _global_te
|
||||
if _global_te is None:
|
||||
with _global_te_lock:
|
||||
# Double-Checked Locking
|
||||
if _global_te is None:
|
||||
if TransferEngine is None:
|
||||
raise RuntimeError("mooncake is not available")
|
||||
transfer_engine = TransferEngine()
|
||||
device_name = device_name if device_name is not None else ""
|
||||
ret_value = transfer_engine.initialize(hostname,
|
||||
"P2PHANDSHAKE",
|
||||
"ascend", device_name)
|
||||
if ret_value != 0:
|
||||
raise RuntimeError(
|
||||
f"TransferEngine initialization failed with ret_value: {ret_value}"
|
||||
)
|
||||
_global_te = transfer_engine
|
||||
return _global_te
|
||||
Reference in New Issue
Block a user