Skip to content

Fix unpadding input with padding mask compute error #38

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
wwbitejotunn wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix unpadding input with padding mask compute error #38

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3450d89
var mask fix
Apr 15, 2024
c450e75
refine
Apr 15, 2024
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
18 changes: 6 additions & 12 deletions csrc/flash_attn/src/flash_fwd_kernel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,11 +146,6 @@ inline __device__ void compute_attn_1rowblock(const Params &params, const int bi
const BlockInfo</*Varlen=*/!Is_even_N> binfo(params, bidb);
if (m_block * kBlockM >= binfo.actual_seqlen_q || binfo.actual_seqlen_k == 0) return;

// umiswing: residue is for predication of additional mask gmem access.
// Additional mask for varlen qkv is supported, but a varlen mask is not supported.
const int m_residue = params.seqlen_q % kBlockM ? params.seqlen_q % kBlockM : kBlockM;
const int n_residue = params.seqlen_k % kBlockN ? params.seqlen_k % kBlockN : kBlockN;

const int m_block_max = cute::ceil_div(binfo.actual_seqlen_q, kBlockM);

int n_block_max = cute::ceil_div(binfo.actual_seqlen_k, kBlockN);
Expand All @@ -175,7 +170,6 @@ inline __device__ void compute_attn_1rowblock(const Params &params, const int bi
+ (n_block_max - 1) * kBlockN * params.v_row_stride + (bidh / params.h_h_k_ratio) * params.v_head_stride;
const index_t row_offset_p = ((bidb * params.h + bidh) * params.seqlen_q_rounded
+ m_block * kBlockM) * params.seqlen_k_rounded + (n_block_max - 1) * kBlockN;

const uint64_t row_offset_mask = (uint64_t)((bidb * params.mask_head_mod_size
+ (bidh % params.mask_head_mod_size)) * params.mask_seq_q_mod_size
+ (m_block * kBlockM % params.mask_seq_q_mod_size)) * params.seqlen_k
Expand Down Expand Up @@ -392,9 +386,9 @@ inline __device__ void compute_attn_1rowblock(const Params &params, const int bi
Tensor scores = make_tensor(acc_s.data(), flash::convert_layout_acc_rowcol(acc_s.layout()));

if (Is_attn_mask) {
flash::apply_attn_mask<Kernel_traits::TiledMma>(scores, tPgMask, tPcMask,
m_block == m_block_max - 1 ? m_residue : params.seqlen_q,
n_block == n_block_max - 1 ? n_residue : params.seqlen_k,
flash::apply_attn_mask<Kernel_traits::TiledMma>(scores, tPgMask, tPcMask,
params.seqlen_q,
params.seqlen_k,
params.unscale_softmax);
tPgMask.data() = tPgMask.data() + (-kBlockN);
}
Expand Down Expand Up @@ -519,9 +513,9 @@ inline __device__ void compute_attn_1rowblock(const Params &params, const int bi
Tensor scores = make_tensor(acc_s.data(), flash::convert_layout_acc_rowcol(acc_s.layout()));

if (Is_attn_mask) {
flash::apply_attn_mask<Kernel_traits::TiledMma>(scores, tPgMask, tPcMask,
m_block == m_block_max - 1 ? m_residue : params.seqlen_q,
n_block == n_block_max - 1 ? n_residue : params.seqlen_k,
flash::apply_attn_mask<Kernel_traits::TiledMma>(scores, tPgMask, tPcMask,
params.seqlen_q,
params.seqlen_k,
params.unscale_softmax);
tPgMask.data() = tPgMask.data() + (-kBlockN);
}
Expand Down