Skip to content

Use int64_t instead of int32_t in decode attention kernel to avoid index calc overflow #5095

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

Closed
wants to merge 1 commit into from
Closed

Use int64_t instead of int32_t in decode attention kernel to avoid index calc overflow #5095

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 ...pu/experimental/gen_ai/src/attention/cuda/cutlass_blackwell_fmha/77_blackwell_fmha_gen.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -335,7 +335,7 @@ struct ExampleRunner {

using StrideQ = Stride<_0, _1, Stride<Stride<int, int>, int>>;
using StrideNewK = Stride<_0, _1, Stride<Stride<_0, int>, int>>;
using StrideCacheK = Stride<int, _1, Stride<Stride<_0, int>, int>>;
using StrideCacheK = Stride<int, _1, Stride<Stride<_0, int64_t>, int64_t>>;
using StrideNewV = StrideNewK;
using StrideCacheV = StrideCacheK;
using StrideO = StrideQ;
Expand Down Expand Up @@ -474,7 +474,7 @@ struct ExampleRunner {

stride_q = make_stride(_0{}, _1{}, make_stride(make_stride(options.d, options.d * size<3,0,0>(result)), options.d * size<3,0>(result)));
stride_new_k = make_stride(_0{}, _1{}, make_stride(make_stride(_0{}, options.d), options.d * size<3,0,1>(result)));
stride_cache_k = make_stride(options.d * size<3,0,1>(result), _1{}, make_stride(make_stride(_0{}, options.d), options.d * size<3,0,1>(result) * get<1>(result)));
stride_cache_k = make_stride(options.d * size<3,0,1>(result), _1{}, make_stride(make_stride(_0{}, static_cast<int64_t>(options.d)), static_cast<int64_t>(options.d) * size<3,0,1>(result) * get<1>(result)));

stride_new_v = stride_new_k;
stride_cache_v = stride_cache_k;
Expand Down
10 changes: 5 additions & 5 deletions ...m_gpu/experimental/gen_ai/src/attention/cuda/cutlass_blackwell_fmha/blackwell_gen_impl.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,9 +93,9 @@ struct GenRunner {

using StrideQ =
Stride<_0, _1, Stride<Stride<int, int>, int>>; // Q D ((H, Hr), B)
using StrideNewK = Stride<_0, _1, Stride<Stride<_0, int>, int>>;
using StrideNewK = Stride<_0, _1, Stride<Stride<_0, int64_t>, int64_t>>;
using StrideCacheK =
Stride<int, _1, Stride<Stride<_0, int>, int>>; // K D ((H, Hr), B)
Stride<int, _1, Stride<Stride<_0, int64_t>, int64_t>>; // K D ((H, Hr), B)
using StrideNewV = StrideNewK;
using StrideCacheV = StrideCacheK;
using StrideO = StrideQ;
Expand Down Expand Up @@ -198,13 +198,13 @@ struct GenRunner {
stride_new_k = make_stride(
_0{},
_1{},
make_stride(make_stride(_0{}, options.d), options.d * options.h_k));
make_stride(make_stride(_0{}, static_cast<int64_t>(options.d)), static_cast<int64_t>(options.d * options.h_k)));
stride_cache_k = make_stride(
options.d * options.h_k,
_1{},
make_stride(
make_stride(_0{}, options.d),
options.d * options.h_k * options.sk));
make_stride(_0{}, static_cast<int64_t>(options.d)),
static_cast<int64_t>(options.d * options.h_k * options.sk)));

stride_new_v = stride_new_k;
stride_cache_v = stride_cache_k;
Expand Down
Loading