biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][Frontend] Fix Issues Under High Load With zeromq Frontend (#7394)

Browse Source 
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
This commit is contained in:
Robert Shaw
2024-08-21 13:34:14 -04:00
committed by GitHub
parent d3c002eadc
commit f7e3b0c5aa
9 changed files with 322 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/entrypoints/openai/api_server.py
View File

@@ -135,6 +135,12 @@ async def build_async_engine_client(
logger.info("Multiprocessing frontend to use %s for RPC Path.",
rpc_path)
# Build RPCClient, which conforms to AsyncEngineClient Protocol.
# NOTE: Actually, this is not true yet. We still need to support
# embedding models via RPC (see TODO above)
rpc_client = AsyncEngineRPCClient(rpc_path)
async_engine_client = rpc_client # type: ignore
# Start RPCServer in separate process (holds the AsyncLLMEngine).
context = multiprocessing.get_context("spawn")
# the current process might have CUDA context,
@@ -145,11 +151,6 @@ async def build_async_engine_client(
rpc_server_process.start()
logger.info("Started engine process with PID %d",
rpc_server_process.pid)
# Build RPCClient, which conforms to AsyncEngineClient Protocol.
# NOTE: Actually, this is not true yet. We still need to support
# embedding models via RPC (see TODO above)
rpc_client = AsyncEngineRPCClient(rpc_path)
async_engine_client = rpc_client # type: ignore
try:
while True: