[Bugfix][SpecDecode] Scope MTP completeness checks outside bucketed updates (#49178)

Signed-off-by: aoshen02 <aoshen02@users.noreply.github.com>
Co-authored-by: aoshen02 <aoshen02@users.noreply.github.com>
This commit is contained in:
aoshen02
2026-07-21 18:56:39 -07:00
committed by GitHub
co-authored by aoshen02
parent 0500ca6a58
commit ba18929079
14 changed files with 131 additions and 37 deletions
@@ -274,7 +274,12 @@ class IPCWeightTransferEngine(
weight = rebuild_cuda_tensor(*list_args)
weights.append((name, weight))
self.model.load_weights(weights)
from vllm.model_executor.model_loader.mtp_validation import (
disable_mtp_completeness_check,
)
with disable_mtp_completeness_check():
self.model.load_weights(weights)
def shutdown(self) -> None:
pass
+31 -26
View File
@@ -168,36 +168,41 @@ class NCCLWeightTransferEngine(
"Call init_transfer_engine() first."
)
if update_info.packed:
# Build iterator of (name, (shape, dtype)) from update_info
def state_dict_info_iterator():
from vllm.model_executor.model_loader.mtp_validation import (
disable_mtp_completeness_check,
)
with disable_mtp_completeness_check():
if update_info.packed:
# Build iterator of (name, (shape, dtype)) from update_info
def state_dict_info_iterator():
for name, dtype_name, shape in zip(
update_info.names, update_info.dtype_names, update_info.shapes
):
dtype = getattr(torch, dtype_name)
yield (name, (shape, dtype))
packed_nccl_broadcast_consumer(
iterator=state_dict_info_iterator(),
group=self.model_update_group,
src=0,
post_unpack_func=self.model.load_weights,
buffer_size_bytes=update_info.packed_buffer_size_bytes,
num_buffers=update_info.packed_num_buffers,
device=self.device,
)
else:
# Use simple one-by-one broadcasting
for name, dtype_name, shape in zip(
update_info.names, update_info.dtype_names, update_info.shapes
):
dtype = getattr(torch, dtype_name)
yield (name, (shape, dtype))
packed_nccl_broadcast_consumer(
iterator=state_dict_info_iterator(),
group=self.model_update_group,
src=0,
post_unpack_func=self.model.load_weights,
buffer_size_bytes=update_info.packed_buffer_size_bytes,
num_buffers=update_info.packed_num_buffers,
device=self.device,
)
else:
# Use simple one-by-one broadcasting
for name, dtype_name, shape in zip(
update_info.names, update_info.dtype_names, update_info.shapes
):
dtype = getattr(torch, dtype_name)
weight = torch.empty(shape, dtype=dtype, device=self.device)
self.model_update_group.broadcast(
weight, src=0, stream=torch.cuda.current_stream()
)
self.model.load_weights([(name, weight)])
del weight
weight = torch.empty(shape, dtype=dtype, device=self.device)
self.model_update_group.broadcast(
weight, src=0, stream=torch.cuda.current_stream()
)
self.model.load_weights([(name, weight)])
del weight
def shutdown(self) -> None:
if self.model_update_group is not None: