48 lines
1.3 KiB
Python
48 lines
1.3 KiB
Python
"""Launch the inference server."""
|
|
|
|
import asyncio
|
|
import os
|
|
import sys
|
|
|
|
from sglang.srt.server_args import prepare_server_args
|
|
from sglang.srt.utils import kill_process_tree
|
|
from sglang.srt.utils.common import suppress_noisy_warnings
|
|
|
|
suppress_noisy_warnings()
|
|
|
|
|
|
def run_server(server_args):
|
|
"""Run the server based on server_args.grpc_mode and server_args.encoder_only."""
|
|
if server_args.grpc_mode:
|
|
from sglang.srt.entrypoints.grpc_server import serve_grpc
|
|
|
|
asyncio.run(serve_grpc(server_args))
|
|
elif server_args.encoder_only:
|
|
from sglang.srt.disaggregation.encode_server import launch_server
|
|
|
|
launch_server(server_args)
|
|
else:
|
|
# Default mode: HTTP mode.
|
|
from sglang.srt.entrypoints.http_server import launch_server
|
|
|
|
launch_server(server_args)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
import warnings
|
|
|
|
warnings.warn(
|
|
"'python -m sglang.launch_server' is still supported, but "
|
|
"'sglang serve' is the recommended entrypoint.\n"
|
|
" Example: sglang serve --model-path <model> [options]",
|
|
UserWarning,
|
|
stacklevel=1,
|
|
)
|
|
|
|
server_args = prepare_server_args(sys.argv[1:])
|
|
|
|
try:
|
|
run_server(server_args)
|
|
finally:
|
|
kill_process_tree(os.getpid(), include_parent=False)
|