Thanks to visit codestin.com
Credit goes to github.com

Skip to content

[Bug] Fix a bug in fused rope #1750

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

Merged
crcrpar merged 1 commit into from
Nov 16, 2023
Merged

[Bug] Fix a bug in fused rope #1750

Show file tree
Hide file tree
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
5 changes: 2 additions & 3 deletions csrc/megatron/fused_rotary_positional_embedding.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,8 +52,7 @@ __global__ void fused_rope_forward(int sq, int b, int np, int hn, int hn2,
int offset_head = offset_block + head_id * hn;
#pragma unroll
for (int hn_id = hn2 + threadIdx.x; hn_id < hn; hn_id += blockDim.x) {
int offset_src_dst = offset_head + hn_id;
dst[offset_src_dst] = src[offset_src_dst];
dst[offset_head + hn_id] = src[offset_head + hn_id];
}
}
}
Expand Down Expand Up @@ -89,7 +88,7 @@ __global__ void fused_rope_backward(int sq, int b, int np, int hn, int hn2,
int offset_head = offset_block + head_id * hn;
#pragma unroll
for (int hn_id = hn2 + threadIdx.x; hn_id < hn; hn_id += blockDim.x) {
dst[offset_head + hn_id] = 1.0;
dst[offset_head + hn_id] = src[offset_head + hn_id];
}
}
}
Expand Down
6 changes: 4 additions & 2 deletions tests/L0/run_transformer/test_fused_rope.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,13 +84,15 @@ def test_forward_backward(self):

# unfused
output_unfused = apply_rotary_pos_emb(t, emb)
output_unfused.sum().backward()
loss_unfused = output_unfused.sum() * 2
loss_unfused.backward()
grad_unfused = t.grad.detach().clone()
t.grad = None

# fused
output_fused = fused_apply_rotary_pos_emb(t, emb)
output_fused.sum().backward()
loss_fused = output_fused.sum() * 2
loss_fused.backward()
grad_fused = t.grad.detach().clone()

self.assertEqual(
Expand Down