Skip to content
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

Avoid duplicate .cpu() call #37

Merged
merged 9 commits into from
Nov 6, 2023
Merged

Avoid duplicate .cpu() call #37

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
dab00be
logs
mvpatel2000 Nov 6, 2023
b6237d8
logs
mvpatel2000 Nov 6, 2023
38cb36d
logs
mvpatel2000 Nov 6, 2023
f399552
logs
mvpatel2000 Nov 6, 2023
4391127
cast to cpu
mvpatel2000 Nov 6, 2023
1d91b99
simplify
mvpatel2000 Nov 6, 2023
673f645
move comment
mvpatel2000 Nov 6, 2023
1a766e1
updated commit
mvpatel2000 Nov 6, 2023
2066a3e
fix comment
mvpatel2000 Nov 6, 2023
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
9 changes: 8 additions & 1 deletion megablocks/layers/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -299,7 +299,8 @@ def parallel_forward_once(self, x, expert_weights, top_experts):
# TODO(tgale): It might be faster to do this on the GPU and
# then communicate the results back to the host.
send_counts = repeated_tokens_per_expert.cpu().sum(dim=-1)
recv_counts = parallel_tokens_per_expert.cpu().sum(dim=-1)
parallel_tokens_per_expert_cpu = parallel_tokens_per_expert.cpu()
recv_counts = parallel_tokens_per_expert_cpu.sum(dim=-1)

# Convert the send/recv counts to lists.
send_counts = send_counts.tolist()
Expand Down Expand Up @@ -374,6 +375,12 @@ def parallel_forward_once(self, x, expert_weights, top_experts):

# Locally permute the tokens and perform the expert computation.
# Block to make sure that the cross-device permutation is complete.
if isinstance(self.mlp, mlp.GroupedMLP):
# GroupedMLP requires counts on CPU. We can use the tensor already
# moved to CPU for the prior all_to_all, which avoids an extra
# device synchronization.
parallel_tokens_per_expert = parallel_tokens_per_expert_cpu.sum(
dim=0, dtype=torch.int)
parallel_x_handle.wait()
parallel_x = self.permute_and_compute(
parallel_x,
Expand Down