[Feat][RL] IPC weight sync optimizations: multigpu support and chunked packed tensors (#37476)

Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Signed-off-by: hao-aaron <ahao@anyscale.com>
This commit is contained in:
Aaron Hao
2026-05-15 22:53:06 +08:00
committed by GitHub
parent 0fe7550254
commit e0a45f1455
10 changed files with 1591 additions and 455 deletions
@@ -21,7 +21,7 @@ from vllm.distributed.weight_transfer.base import (
from vllm.distributed.weight_transfer.packed_tensor import (
DEFAULT_PACKED_BUFFER_SIZE_BYTES,
DEFAULT_PACKED_NUM_BUFFERS,
packed_broadcast_consumer,
packed_nccl_broadcast_consumer,
)
@@ -184,7 +184,7 @@ class NCCLWeightTransferEngine(
dtype = getattr(torch, dtype_name)
yield (name, (shape, dtype))
packed_broadcast_consumer(
packed_nccl_broadcast_consumer(
iterator=state_dict_info_iterator(),
group=self.model_update_group,
src=0,
@@ -247,10 +247,10 @@ class NCCLWeightTransferEngine(
if args.packed:
# Use packed tensor broadcasting for efficiency
from vllm.distributed.weight_transfer.packed_tensor import (
packed_broadcast_producer,
packed_nccl_broadcast_producer,
)
packed_broadcast_producer(
packed_nccl_broadcast_producer(
iterator=iterator,
group=args.group,
src=args.src,