Skip to content

fix: Fix placeholder IPC handles when a rollout engine spans multiple nodes #2520

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
leondada wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: Fix placeholder IPC handles when a rollout engine spans multiple nodes #2520

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 18 additions & 9 deletions slime/backends/megatron_utils/update_weight/update_weight_from_tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -406,32 +406,41 @@ def _send_to_colocated_engine(
long_live_tensors.append(flattened_tensor_data)
serialized_tensors.append(MultiprocessingSerializer.serialize(flattened_tensor_data, output_str=True))

# Each rank contributes its own empty placeholder. The gather source must not
# fabricate placeholders on behalf of peers: a serialized placeholder carries a
# CUDA IPC handle tagged with the *producing* rank's device, and an engine worker
# on another node cannot import a handle for a device it does not see
# (sglang raises "Invalid device_uuid=..."). This only shows up when a single
# engine spans more than one node (rollout_num_gpus_per_engine > num_gpus_per_node).
local_empty_tensor_data = _empty_flattened_tensor_data()
long_live_tensors.append(local_empty_tensor_data)
local_empty_serialized_tensor = MultiprocessingSerializer.serialize(local_empty_tensor_data, output_str=True)

serialized_named_tensors = (
[None] * dist.get_world_size(ipc_gather_group) if ipc_gather_src == dist.get_rank() else None
)
dist.gather_object(
serialized_tensors,
(serialized_tensors, local_empty_serialized_tensor),
object_gather_list=serialized_named_tensors,
dst=ipc_gather_src,
group=ipc_gather_group,
)

refs = []
if dist.get_rank() == ipc_gather_src:
num_buckets = max(len(tensors) for tensors in serialized_named_tensors)
empty_serialized_tensor = None
per_rank_tensors = [item[0] for item in serialized_named_tensors]
per_rank_empty_tensor = [item[1] for item in serialized_named_tensors]
num_buckets = max(len(tensors) for tensors in per_rank_tensors)
for i in range(num_buckets):
serialized_tensors_for_dtype = []
for tensors in serialized_named_tensors:
for rank_index, tensors in enumerate(per_rank_tensors):
if i < len(tensors):
serialized_tensors_for_dtype.append(tensors[i])
continue

if empty_serialized_tensor is None:
empty_tensor_data = _empty_flattened_tensor_data()
long_live_tensors.append(empty_tensor_data)
empty_serialized_tensor = MultiprocessingSerializer.serialize(empty_tensor_data, output_str=True)
serialized_tensors_for_dtype.append(empty_serialized_tensor)
# Use the placeholder produced by *that* rank so the handle always
# refers to a device visible to the engine worker reading this slot.
serialized_tensors_for_dtype.append(per_rank_empty_tensor[rank_index])

kwargs = {
"serialized_named_tensors": serialized_tensors_for_dtype,
Expand Down
Loading