Skip to content

Fix: add missing end_sync barrier in fused allreduce+rmsnorm kernel #4346

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
yuzho-amd wants to merge 1 commit into
base: main
Choose a base branch
from
+7 −0
Open

Fix: add missing end_sync barrier in fused allreduce+rmsnorm kernel #4346

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
7 changes: 7 additions & 0 deletions csrc/include/custom_all_reduce.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2044,6 +2044,11 @@ __global__ void __launch_bounds__(1024, 1)
*reinterpret_cast<OP*>(output + out_idx) = zero_pack;
}
}
// Pair start_sync with a final barrier before any rank can reuse the
// registered input buffer or this block's signal slot for the next fused
// collective. Without it, a faster rank may advance while a peer is still
// reading the previous invocation, causing silent cross-request corruption.
end_sync<ngpus, true>(sg, self_sg, rank);
}

// Per-group quant variant of the 1-stage fused allreduce+rmsnorm kernel.
Expand Down Expand Up @@ -2127,6 +2132,7 @@ __global__ void __launch_bounds__(1024, 1)
acc, weight_p, hidden_dim, eps, idx, tidx, padded_block_size,
group_size, output, scale_out, active, bf16_output, token_num);
}
end_sync<ngpus, true>(sg, self_sg, rank);
}

template <typename T, typename OutT, int NGPUS>
Expand Down Expand Up @@ -2236,6 +2242,7 @@ __global__ void __launch_bounds__(1024, 1)
acc, weight_p, hidden_dim, eps, idx, tidx, padded_block_size,
output, scale_out, active, bf16_output);
}
end_sync<ngpus, true>(sg, self_sg, rank);
}

template <typename T, int NGPUS>
Expand Down
Loading