[Mini-vLLM] SM86-only cut: models, backends, Rust, tests

- Keep only Qwen3.5, Qwen3.5-MTP and Qwen4Exp model paths; trim the model registry to 41 entries
- Delete idle attention backends (MLA, FlashInfer, ROCm-only, quant backends) and dead GDN CuTeDSL kernels
- Remove inc/quark quant loaders; keep compressed-tensors and Marlin
- Drop the Rust frontend, pooling entrypoints, and idle test suites
- Rebuild kernels for sm_80;sm_86 and strip the extensions
- Verified: full import sweep clean (optional deps only) and a dummy-load Qwen3.5 engine run on RTX 3090 via the Triton GDN path

Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
2026-10-18 19:19:56 +07:00
co-authored by Hinata
parent c5ab42bc0f
commit e61dcb9143
2728 changed files with 159 additions and 950401 deletions
+13 -65
View File
@@ -17,7 +17,6 @@ from vllm.entrypoints.launchers.cli_args import (
validate_parsed_serve_args,
)
from vllm.entrypoints.launchers.dp_supervisor import run_dp_supervisor
from vllm.entrypoints.launchers.launcher import create_server_socket
from vllm.entrypoints.serve.utils.api_utils import VLLM_SUBCMD_PARSER_EPILOG
from vllm.logger import init_logger
from vllm.reasoning import ReasoningParserManager
@@ -30,7 +29,6 @@ from vllm.v1.executor.multiproc_executor import MultiprocExecutor
from vllm.v1.metrics.prometheus import setup_multiprocess_prometheus
from vllm.v1.utils import (
APIServerProcessManager,
RustFrontendProcessManager,
wait_for_completion_or_failure,
)
@@ -62,10 +60,6 @@ class ServeSubcommand(CLISubcommand):
uvloop.run(serve_grpc(args))
return
rust_frontend_path = (
envs.VLLM_RUST_FRONTEND_PATH if envs.VLLM_USE_RUST_FRONTEND else None
)
if args.headless:
if args.api_server_count is not None and args.api_server_count > 0:
raise ValueError(
@@ -106,12 +100,11 @@ class ServeSubcommand(CLISubcommand):
# Default api_server_count if not explicitly set.
# - Multi-port: 1 (supervisor spawns one server per local DP rank)
# - Rust frontend: 1 (not applicable as it's multithreaded)
# - External LB: 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_multi_port or is_external_lb or rust_frontend_path:
if is_multi_port or is_external_lb:
args.api_server_count = 1
elif is_hybrid_lb:
args.api_server_count = args.data_parallel_size_local or 1
@@ -128,12 +121,6 @@ class ServeSubcommand(CLISubcommand):
"Defaulting api_server_count to data_parallel_size (%d).",
args.api_server_count,
)
elif 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,
)
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:
@@ -148,7 +135,7 @@ class ServeSubcommand(CLISubcommand):
run_dp_supervisor(args)
elif args.api_server_count < 1:
run_headless(args)
elif args.api_server_count > 1 or rust_frontend_path:
elif args.api_server_count > 1:
run_multi_api_server(args)
else:
# Single API server (this process).
@@ -268,17 +255,9 @@ 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 if envs.VLLM_USE_RUST_FRONTEND else None
)
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()
@@ -296,13 +275,6 @@ def run_multi_api_server(args: argparse.Namespace):
signal.signal(signal.SIGINT, signal_handler)
listen_address, sock = setup_server(args, reuse_port=num_api_servers > 1)
# `--grpc-port` is only accepted for the Rust frontend, which inherits this
# listener like the HTTP one. gRPC follows the HTTP TCP host, or IPv4
# loopback when HTTP uses a Unix socket.
grpc_sock = None
if args.grpc_port is not None:
grpc_host = "127.0.0.1" if args.uds else (args.host or "")
grpc_sock = create_server_socket((grpc_host, args.grpc_port), reuse_port=False)
engine_args = vllm.AsyncEngineArgs.from_cli_args(args)
propagate_flash_late_interaction(args, engine_args)
@@ -324,9 +296,7 @@ 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 | RustFrontendProcessManager | None = (
None
)
api_server_manager: APIServerProcessManager | None = None
from vllm.v1.engine.utils import bind_engine_zmq_listeners
@@ -343,38 +313,16 @@ def run_multi_api_server(args: argparse.Namespace):
coordinator.get_stats_publish_address() if coordinator else None
)
if rust_frontend_path:
if parallel_config.local_engines_only:
expected_engine_start_index = parallel_config.data_parallel_rank
expected_engine_count = parallel_config.data_parallel_size_local
else:
expected_engine_start_index = 0
expected_engine_count = parallel_config.data_parallel_size
# Start rust front-end process.
api_server_manager = RustFrontendProcessManager(
binary_path=rust_frontend_path,
sock=sock,
grpc_sock=grpc_sock,
args=args,
input_listener=zmq_listeners.inputs[0],
output_listener=zmq_listeners.outputs[0],
engine_start_index=expected_engine_start_index,
engine_count=expected_engine_count,
data_parallel_size=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_listeners=zmq_listeners.inputs,
output_listeners=zmq_listeners.outputs,
stats_update_address=stats_update_address,
tensor_queue=engine_launch.tensor_queue,
)
api_server_manager = APIServerProcessManager(
listen_address=listen_address,
sock=sock,
args=args,
num_servers=num_api_servers,
input_listeners=zmq_listeners.inputs,
output_listeners=zmq_listeners.outputs,
stats_update_address=stats_update_address,
tensor_queue=engine_launch.tensor_queue,
)
# Set frontend processes to watch during engine startup.
# If any of these processes exit before the engines are up, the engine startup