[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
@@ -51,6 +51,9 @@ async def build_async_engine_client(
# Context manager to handle engine_client lifecycle
# Ensures everything is shutdown and cleaned up on error/exit
engine_args = AsyncEngineArgs.from_cli_args(args)
from ..cli_args import propagate_flash_late_interaction
propagate_flash_late_interaction(args, engine_args)
if client_config:
engine_args._api_process_count = client_config.get("client_count", 1)
engine_args._api_process_rank = client_config.get("client_index", 0)