[Bugfix] Import each packed IPC export once on the consumer side (#51259)

Signed-off-by: acmore <hezj07+github@gmail.com>
This commit is contained in:
acmore
2026-08-11 10:05:27 +08:00
committed by GitHub
parent 33584901ba
commit f37dff22fc
2 changed files with 59 additions and 4 deletions
+12 -1
View File
@@ -27,6 +27,7 @@ if TYPE_CHECKING:
from vllm.config import VllmConfig
from vllm.distributed.weight_transfer.packed_tensor import (
DEFAULT_PACKED_BUFFER_SIZE_BYTES,
PackedBufferImporter,
packed_ipc_consumer,
packed_ipc_producer,
)
@@ -143,6 +144,9 @@ class IPCWeightTransferEngine(
# Set from the trainer-supplied init info at the handshake; defaults are
# only for the (unreachable) receive-before-init case.
self.packed = False
# Shared across all chunks of every packed transfer this engine
# receives; see PackedBufferImporter for the refcount contract.
self._packed_importer = PackedBufferImporter()
def init_transfer_engine(self, init_info: IPCWeightTransferInitInfo) -> None:
"""
@@ -170,6 +174,12 @@ class IPCWeightTransferEngine(
)
finalize_layerwise_reload(self.model, self.model_config)
# Every reduce_tensor call is a fresh export with its own refcount
# slot, so releasing once per update always balances this update's
# export and lets the trainer reclaim its staging buffer. Callers
# that skip finish are still covered by the replace-on-next-export
# path inside the importer.
self._packed_importer.close()
def receive_weights(self, update_info: IPCWeightTransferUpdateInfo) -> None:
"""
@@ -202,6 +212,7 @@ class IPCWeightTransferEngine(
dtype_names=update_info.dtype_names,
tensor_sizes=update_info.tensor_sizes,
device_index=device_index,
importer=self._packed_importer,
)
else:
assert isinstance(update_info.ipc_handles, list)
@@ -236,7 +247,7 @@ class IPCWeightTransferEngine(
self.model.load_weights(weights)
def shutdown(self) -> None:
pass
self._packed_importer.close()
@staticmethod
def trainer_send_weights(*args: Any, **kwargs: Any) -> None: