[Frontend][RFC] Rust front-end integration (#40848)
Signed-off-by: Nick Hill <nickhill123@gmail.com> Signed-off-by: Bugen Zhao <i@bugenzhao.com> Co-authored-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
@@ -2,6 +2,7 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import argparse
|
||||
import signal
|
||||
|
||||
import uvloop
|
||||
|
||||
@@ -110,6 +111,14 @@ def cmd_init() -> list[CLISubcommand]:
|
||||
|
||||
async def run_launch_fastapi(args: argparse.Namespace) -> None:
|
||||
"""Run the online serving layer with FastAPI (no GPU inference)."""
|
||||
|
||||
# Interrupt initialization if SIGTERM arrives before uvicorn installs
|
||||
# its own signal handlers. Once uvicorn is running it replaces this.
|
||||
def _interrupt_init(*_) -> None:
|
||||
raise KeyboardInterrupt("terminated")
|
||||
|
||||
signal.signal(signal.SIGTERM, _interrupt_init)
|
||||
|
||||
# 1. Socket binding
|
||||
listen_address, sock = setup_server(args)
|
||||
|
||||
|
||||
@@ -21,7 +21,11 @@ from vllm.v1.engine.utils import CoreEngineProcManager, launch_core_engines
|
||||
from vllm.v1.executor import Executor
|
||||
from vllm.v1.executor.multiproc_executor import MultiprocExecutor
|
||||
from vllm.v1.metrics.prometheus import setup_multiprocess_prometheus
|
||||
from vllm.v1.utils import APIServerProcessManager, wait_for_completion_or_failure
|
||||
from vllm.v1.utils import (
|
||||
APIServerProcessManager,
|
||||
RustFrontendProcessManager,
|
||||
wait_for_completion_or_failure,
|
||||
)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -81,11 +85,12 @@ class ServeSubcommand(CLISubcommand):
|
||||
)
|
||||
|
||||
# Default api_server_count if not explicitly set.
|
||||
# - Rust frontend: Use 1 (not applicable as it's multithreaded)
|
||||
# - External LB: Leave as 1 (external LB handles distribution)
|
||||
# - Hybrid LB: Use local DP size (internal LB for local ranks only)
|
||||
# - Internal LB: Use full DP size
|
||||
if args.api_server_count is None:
|
||||
if is_external_lb:
|
||||
if is_external_lb or envs.VLLM_RUST_FRONTEND_PATH:
|
||||
args.api_server_count = 1
|
||||
elif is_hybrid_lb:
|
||||
args.api_server_count = args.data_parallel_size_local or 1
|
||||
@@ -102,6 +107,11 @@ class ServeSubcommand(CLISubcommand):
|
||||
"Defaulting api_server_count to data_parallel_size (%d).",
|
||||
args.api_server_count,
|
||||
)
|
||||
elif envs.VLLM_RUST_FRONTEND_PATH and args.api_server_count > 1:
|
||||
logger.warning(
|
||||
"Ignoring --api-server-count=%d when using rust front-end process"
|
||||
)
|
||||
args.api_server_count = 1
|
||||
|
||||
# Elastic EP currently only supports running with at most one API server.
|
||||
if getattr(args, "enable_elastic_ep", False) and args.api_server_count > 1:
|
||||
@@ -114,7 +124,7 @@ class ServeSubcommand(CLISubcommand):
|
||||
|
||||
if args.api_server_count < 1:
|
||||
run_headless(args)
|
||||
elif args.api_server_count > 1:
|
||||
elif args.api_server_count > 1 or envs.VLLM_RUST_FRONTEND_PATH:
|
||||
run_multi_api_server(args)
|
||||
else:
|
||||
# Single API server (this process).
|
||||
@@ -230,9 +240,15 @@ def run_headless(args: argparse.Namespace):
|
||||
|
||||
def run_multi_api_server(args: argparse.Namespace):
|
||||
assert not args.headless
|
||||
rust_frontend_path = envs.VLLM_RUST_FRONTEND_PATH
|
||||
num_api_servers: int = args.api_server_count
|
||||
assert num_api_servers > 0
|
||||
|
||||
if rust_frontend_path and num_api_servers > 1:
|
||||
raise ValueError(
|
||||
"VLLM_RUST_FRONTEND_PATH does not support api_server_count > 1"
|
||||
)
|
||||
|
||||
if num_api_servers > 1:
|
||||
setup_multiprocess_prometheus()
|
||||
|
||||
@@ -270,7 +286,9 @@ def run_multi_api_server(args: argparse.Namespace):
|
||||
dp_rank = parallel_config.data_parallel_rank
|
||||
assert parallel_config.local_engines_only or dp_rank == 0
|
||||
|
||||
api_server_manager: APIServerProcessManager | None = None
|
||||
api_server_manager: APIServerProcessManager | RustFrontendProcessManager | None = (
|
||||
None
|
||||
)
|
||||
|
||||
from vllm.v1.engine.utils import get_engine_zmq_addresses
|
||||
|
||||
@@ -279,23 +297,34 @@ def run_multi_api_server(args: argparse.Namespace):
|
||||
with launch_core_engines(
|
||||
vllm_config, executor_class, log_stats, addresses, num_api_servers
|
||||
) as (local_engine_manager, coordinator, addresses, tensor_queue):
|
||||
# Construct common args for the APIServerProcessManager up-front.
|
||||
stats_update_address = None
|
||||
if coordinator:
|
||||
stats_update_address = coordinator.get_stats_publish_address()
|
||||
|
||||
# Start API servers.
|
||||
api_server_manager = APIServerProcessManager(
|
||||
listen_address=listen_address,
|
||||
sock=sock,
|
||||
args=args,
|
||||
num_servers=num_api_servers,
|
||||
input_addresses=addresses.inputs,
|
||||
output_addresses=addresses.outputs,
|
||||
stats_update_address=stats_update_address,
|
||||
tensor_queue=tensor_queue,
|
||||
stats_update_address = (
|
||||
coordinator.get_stats_publish_address() if coordinator else None
|
||||
)
|
||||
|
||||
if rust_frontend_path:
|
||||
# Start rust front-end process.
|
||||
api_server_manager = RustFrontendProcessManager(
|
||||
binary_path=rust_frontend_path,
|
||||
sock=sock,
|
||||
args=args,
|
||||
input_address=addresses.inputs[0],
|
||||
output_address=addresses.outputs[0],
|
||||
engine_count=parallel_config.data_parallel_size,
|
||||
stats_update_address=stats_update_address,
|
||||
)
|
||||
else:
|
||||
# Start API server(s).
|
||||
api_server_manager = APIServerProcessManager(
|
||||
listen_address=listen_address,
|
||||
sock=sock,
|
||||
args=args,
|
||||
num_servers=num_api_servers,
|
||||
input_addresses=addresses.inputs,
|
||||
output_addresses=addresses.outputs,
|
||||
stats_update_address=stats_update_address,
|
||||
tensor_queue=tensor_queue,
|
||||
)
|
||||
|
||||
# Wait for API servers.
|
||||
try:
|
||||
wait_for_completion_or_failure(
|
||||
|
||||
@@ -533,8 +533,7 @@ def validate_api_server_args(args):
|
||||
|
||||
@instrument(span_name="API server setup")
|
||||
def setup_server(args):
|
||||
"""Validate API server args, set up signal handler, create socket
|
||||
ready to serve."""
|
||||
"""Validate API server args and create the server socket."""
|
||||
|
||||
log_version_and_model(logger, VLLM_VERSION, args.model)
|
||||
log_non_default_args(args)
|
||||
@@ -560,12 +559,6 @@ def setup_server(args):
|
||||
# many concurrent requests active
|
||||
set_ulimit()
|
||||
|
||||
def signal_handler(*_) -> None:
|
||||
# Interrupt server on sigterm while initializing
|
||||
raise KeyboardInterrupt("terminated")
|
||||
|
||||
signal.signal(signal.SIGTERM, signal_handler)
|
||||
|
||||
if args.uds:
|
||||
listen_address = f"unix:{args.uds}"
|
||||
else:
|
||||
@@ -675,6 +668,13 @@ async def run_server(args, **uvicorn_kwargs) -> None:
|
||||
# Add process-specific prefix to stdout and stderr.
|
||||
decorate_logs("APIServer")
|
||||
|
||||
# Interrupt initialization if SIGTERM arrives before uvicorn installs its
|
||||
# own signal handlers. Once uvicorn is running it replaces this.
|
||||
def _interrupt_init(*_) -> None:
|
||||
raise KeyboardInterrupt("terminated")
|
||||
|
||||
signal.signal(signal.SIGTERM, _interrupt_init)
|
||||
|
||||
listen_address, sock = setup_server(args)
|
||||
await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)
|
||||
|
||||
|
||||
@@ -9,6 +9,7 @@ from argparse import Namespace
|
||||
from http import HTTPStatus
|
||||
from logging import Logger
|
||||
from string import Template
|
||||
from typing import Any
|
||||
|
||||
import regex as re
|
||||
from fastapi import Request
|
||||
@@ -210,7 +211,7 @@ def get_max_tokens(
|
||||
)
|
||||
|
||||
|
||||
def log_non_default_args(args: Namespace | EngineArgs):
|
||||
def get_non_default_args(args: Namespace | EngineArgs) -> dict[str, Any]:
|
||||
from vllm.entrypoints.openai.cli_args import make_arg_parser
|
||||
|
||||
non_default_args = {}
|
||||
@@ -237,6 +238,43 @@ def log_non_default_args(args: Namespace | EngineArgs):
|
||||
"Unsupported argument type. Must be Namespace or EngineArgs instance."
|
||||
)
|
||||
|
||||
return non_default_args
|
||||
|
||||
|
||||
def _jsonify_arg_value(value: Any) -> Any:
|
||||
if value is None or isinstance(value, bool | int | float | str):
|
||||
return value
|
||||
if dataclasses.is_dataclass(value) and not isinstance(value, type):
|
||||
return {
|
||||
key: _jsonify_arg_value(val)
|
||||
for key, val in dataclasses.asdict(value).items()
|
||||
}
|
||||
if isinstance(value, dict):
|
||||
return {str(key): _jsonify_arg_value(val) for key, val in value.items()}
|
||||
if isinstance(value, tuple | list):
|
||||
return [_jsonify_arg_value(item) for item in value]
|
||||
if (model_dump := getattr(value, "model_dump", None)) is not None:
|
||||
return _jsonify_arg_value(model_dump(mode="json"))
|
||||
if (to_dict := getattr(value, "dict", None)) is not None:
|
||||
return _jsonify_arg_value(to_dict())
|
||||
return repr(value)
|
||||
|
||||
|
||||
def jsonify_non_default_args(
|
||||
args: Namespace | EngineArgs,
|
||||
*,
|
||||
exclude: set[str] | None = None,
|
||||
) -> dict[str, Any]:
|
||||
non_default_args = get_non_default_args(args)
|
||||
if exclude is not None:
|
||||
for key in exclude:
|
||||
non_default_args.pop(key, None)
|
||||
|
||||
return {key: _jsonify_arg_value(value) for key, value in non_default_args.items()}
|
||||
|
||||
|
||||
def log_non_default_args(args: Namespace | EngineArgs):
|
||||
non_default_args = get_non_default_args(args)
|
||||
logger.info("non-default args: %s", non_default_args)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user