Skip to content

[fix](kt): Fix barrier call to use cpu_group instead of device_group #27

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

[fix](kt): Fix barrier call to use cpu_group instead of device_group #27

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
18 changes: 7 additions & 11 deletions python/sglang/srt/layers/moe/kt_ep_wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -464,11 +464,7 @@ def _create_cpu_buffers(self):
else:
self.shm_unique_id = None
if dist.is_initialized():
unique_id_list = [self.shm_unique_id]
dist.broadcast_object_list(
unique_id_list, src=0, group=get_tp_group().cpu_group
)
self.shm_unique_id = unique_id_list[0]
self.shm_unique_id = get_tp_group().broadcast_object(self.shm_unique_id)

for name in self.weight_names:
gpu_tensor = getattr(self.gpu_layer, name)
Expand Down Expand Up @@ -499,14 +495,14 @@ def _create_cpu_buffers(self):
self.cpu_buffers[name] = cpu_buffer

if dist.is_initialized():
dist.barrier(group=get_tp_group().device_group)
get_tp_group().barrier()

self.all_rank_buffer_ptrs = self._collect_all_rank_buffer_pointers()

# Unlink shared memory after all ranks have collected pointers.
# The memory remains accessible as long as we hold references via mmap.
if dist.is_initialized():
dist.barrier(group=get_tp_group().device_group)
get_tp_group().barrier()
for shm in self.shm_handles.values():
shm.unlink()

Expand Down Expand Up @@ -706,7 +702,7 @@ def submit_write_expert(expert_id):

# Barrier to ensure all ranks see the written data
if dist.is_initialized():
dist.barrier(group=get_tp_group().device_group)
get_tp_group().barrier()

with torch.cuda.stream(copy_stream):
slot = e % 2 # Double buffering
Expand Down Expand Up @@ -874,7 +870,7 @@ def submit_write_expert(expert_id, slot):

# Barrier to ensure all ranks see the written data
if dist.is_initialized():
dist.barrier(group=get_tp_group().device_group)
get_tp_group().barrier()

with torch.cuda.stream(copy_stream):
for _, cpu_buf, gpu_t in weight_infos:
Expand Down Expand Up @@ -1046,7 +1042,7 @@ def submit_write_expert(expert_id, slot):

# Barrier to ensure all ranks see the written data
if dist.is_initialized():
dist.barrier(group=get_tp_group().device_group)
get_tp_group().barrier()

with torch.cuda.stream(copy_stream):
for _, cpu_buf, gpu_t in weight_infos:
Expand Down Expand Up @@ -1198,7 +1194,7 @@ def submit_write_expert(expert_id, slot):

# Barrier to ensure all ranks see the written data
if dist.is_initialized():
dist.barrier(group=get_tp_group().device_group)
get_tp_group().barrier()

with torch.cuda.stream(copy_stream):
for _, cpu_buf, gpu_t in weight_infos:
Expand Down
Loading