[Perf] Integrate flash-maxsim Triton kernels for late-interaction scoring (#40337)

Signed-off-by: roi.pony <roi.pony@ibm.com>
Co-authored-by: roi.pony <roi.pony@ibm.com>
Co-authored-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
roipony
2026-10-01 14:06:43 -04:00
committed by GitHub
co-authored by roi.pony Taneem Ibrahim
parent 083060d046
commit e9be5323f0
10 changed files with 508 additions and 4 deletions
+3
View File
@@ -13,6 +13,7 @@ from vllm.entrypoints.cli.types import CLISubcommand
from vllm.entrypoints.launchers.api_server.entry import run_server, setup_server
from vllm.entrypoints.launchers.cli_args import (
make_arg_parser,
propagate_flash_late_interaction,
validate_parsed_serve_args,
)
from vllm.entrypoints.launchers.dp_supervisor import run_dp_supervisor
@@ -185,6 +186,7 @@ def run_headless(args: argparse.Namespace):
# Create the EngineConfig.
engine_args = vllm.AsyncEngineArgs.from_cli_args(args)
propagate_flash_late_interaction(args, engine_args)
usage_context = UsageContext.OPENAI_API_SERVER
vllm_config = engine_args.create_engine_config(
usage_context=usage_context, headless=True
@@ -295,6 +297,7 @@ def run_multi_api_server(args: argparse.Namespace):
listen_address, sock = setup_server(args, reuse_port=num_api_servers > 1)
engine_args = vllm.AsyncEngineArgs.from_cli_args(args)
propagate_flash_late_interaction(args, engine_args)
engine_args._api_process_count = num_api_servers
engine_args._api_process_rank = -1