[Distributed] Support pre-shared ncclUniqueId rendezvous for weight transfer (#53784)

Signed-off-by: Dharak Kharod <dharak.kharod@cohere.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
dharak-cohere
2026-09-01 14:55:57 +08:00
committed by GitHub
co-authored by Cursor
parent fa99a6fea6
commit d0e695a91b
6 changed files with 674 additions and 10 deletions
@@ -31,6 +31,7 @@ from vllm.distributed.weight_transfer.base import (
)
from vllm.distributed.weight_transfer.nccl_common import (
NCCLWeightTransferInitInfo,
worker_init_payload,
worker_init_process_group,
)
from vllm.distributed.weight_transfer.nccl_common import (
@@ -256,7 +257,8 @@ class SparseNCCLTrainerWeightTransferEngine(
# open the trainer endpoint (rank 0); both sides must rendezvous together.
with ThreadPoolExecutor(max_workers=1) as exe:
future = exe.submit(
engine.client.init_weight_transfer_engine, asdict(worker_init_info)
engine.client.init_weight_transfer_engine,
worker_init_payload(worker_init_info),
)
# Open the trainer endpoint as NCCL rank 0 on the current device
# (the init info satisfies the helper's rendezvous protocol).