Skip to content

Fix causal+sliding window mask alignment for autoregressive decode #185

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
juliantorr-es wants to merge 1 commit into
base: tridao
Choose a base branch
from
Open

Fix causal+sliding window mask alignment for autoregressive decode #185

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
4 changes: 2 additions & 2 deletions csrc/flash_attn_ck/mha_bwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,15 +279,15 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q x num

mask_info mask;
if (is_causal) {
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + "0";
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + "0";
mask = mask_info::decode(mask_identify, seqlen_q, seqlen_k); // casual
}
else if (window_size_left == -1 && window_size_right == -1) {
mask = mask_info::decode("0", seqlen_q, seqlen_k); // no mask
}
else {
// Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
mask = mask_info::decode(mask_identify, seqlen_q, seqlen_k); // local
}

Expand Down
4 changes: 2 additions & 2 deletions csrc/flash_attn_ck/mha_fwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,15 +219,15 @@ mha_fwd(at::Tensor &q, // batch_size x seqlen_q x num
if (is_causal) {
// Causal is the special case where window_size_right == 0 and window_size_left < 0.
window_size_right = 0;
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + "0";
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + "0";
mask = mask_info::decode(mask_identify, seqlen_q, seqlen_k); // casual
}
else if (window_size_left == -1 && window_size_right == -1) {
mask = mask_info::decode("0", seqlen_q, seqlen_k); // no mask
}
else {
// Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
mask = mask_info::decode(mask_identify, seqlen_q, seqlen_k); // local
}

Expand Down
6 changes: 3 additions & 3 deletions csrc/flash_attn_ck/mha_varlen_bwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -292,15 +292,15 @@ mha_varlen_bwd(const at::Tensor &dout, // total_q x num_heads

mask_info mask;
if (is_causal) {
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + "0";
mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // casual
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + "0";
mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // causal
}
else if (window_size_left == -1 && window_size_right == -1) {
mask = mask_info::decode("0", max_seqlen_q, max_seqlen_k); // no mask
}
else {
// Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // local
}

Expand Down
6 changes: 3 additions & 3 deletions csrc/flash_attn_ck/mha_varlen_fwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -407,15 +407,15 @@ mha_varlen_fwd(at::Tensor &q, // total_q x num_heads x head_si
if (is_causal) {
// Causal is the special case where window_size_right == 0 and window_size_left < 0.
window_size_right = 0;
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + "0";
mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // casual
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + "0";
mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // causal
}
else if (window_size_left == -1 && window_size_right == -1) {
mask = mask_info::decode("0", max_seqlen_q, max_seqlen_k); // no mask
}
else {
// Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
std::string mask_identify = "b:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
std::string mask_identify = "t:" + std::to_string(window_size_left) + "," + std::to_string(window_size_right);
mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // local
}

Expand Down