Skip to content

Use different P2P tag for acts/grads in pipeline #12

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
reyoung wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Use different P2P tag for acts/grads in pipeline #12

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
11 changes: 7 additions & 4 deletions varuna/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,9 @@
import os, sys
import time

P2P_TAG_ACTIONS = 1
P2P_TAG_GRADS = 2

class Pipeline:
""" Pipeline parallelism for Varuna """

Expand Down Expand Up @@ -131,7 +134,7 @@ def acts_receiver(self):
for d in self.fwd_inp_shape_changes:
fwd_inp_shape[d] = self.last_chunk_size
acts_tensor = torch.ones(fwd_inp_shape, dtype=dtype)
handle = dist.irecv(acts_tensor, src=self.receive_rank)
handle = dist.irecv(acts_tensor, src=self.receive_rank, tag=P2P_TAG_ACTIONS)
recv_handles.put((handle, acts_tensor))
if recv_handles.qsize()>4:
handle, tensor = recv_handles.get()
Expand All @@ -156,7 +159,7 @@ def grads_receiver(self):
for d in self.bwd_grad_shape_changes:
bwd_grad_shape[d] = self.last_chunk_size
grads_tensor = torch.ones(bwd_grad_shape, dtype=dtype)
handle = dist.irecv(grads_tensor, src=self.send_rank)
handle = dist.irecv(grads_tensor, src=self.send_rank, tag=P2P_TAG_GRADS)
recv_handles.put((handle, grads_tensor))
if recv_handles.qsize()>4:
handle, tensor = recv_handles.get()
Expand All @@ -176,7 +179,7 @@ def acts_sender(self):
send_handles = Queue()
while count > 0:
output_acts = self.acts_send_queue.get()
handle = dist.isend(output_acts, dst=self.send_rank)
handle = dist.isend(output_acts, dst=self.send_rank, tag=P2P_TAG_ACTIONS)
send_handles.put(handle)
if send_handles.qsize()>4:
handle = send_handles.get()
Expand All @@ -197,7 +200,7 @@ def grads_sender(self):
while count > 0:
input_grads = self.grads_send_queue.get()
# TODO: why is this contiguous needed ???
handle = dist.isend(input_grads.contiguous(), dst=self.receive_rank)
handle = dist.isend(input_grads.contiguous(), dst=self.receive_rank, tag=P2P_TAG_GRADS)
send_handles.put(handle)
if send_handles.qsize()>4:
handle = send_handles.get()
Expand Down