Skip to content

hipblaslt Fallback hotfix #652

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
alextmagro merged 4 commits into from
Jun 30, 2026
Merged

hipblaslt Fallback hotfix #652

Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a66e797
hipblaslt Fallback hotfix
alextmagro Jun 26, 2026
0477d1c
Merge remote-tracking branch 'origin/dev' into hipkittens_fallback_ho…
alextmagro Jun 29, 2026
db36890
Remove hipkittens_gemm_complete var
alextmagro Jun 29, 2026
c84068c
Fix JAX workspace and bias check
alextmagro Jun 29, 2026
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
27 changes: 14 additions & 13 deletions transformer_engine/common/gemm/rocm_gemm.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1985,7 +1985,7 @@ void cublas_gemm(const Tensor *inputA, const Tensor *inputB, Tensor *outputD,

#ifdef USE_HIPKITTENS_GEMM

bool use_hipkittens = false;
bool use_hipkittens = false, hipkittens_gemm_complete = false;

@ipanfilo ipanfilo Jun 29, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is unused variable

@alextmagro alextmagro Jun 29, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry, missed that. Should be fixed now.

if (is_mxfp8) {
bool is_gfx950 = (cuda::sm_arch() == 95);
bool force_hipblaslt = false;
Expand All @@ -1999,18 +1999,19 @@ void cublas_gemm(const Tensor *inputA, const Tensor *inputB, Tensor *outputD,
if (use_hipkittens) {
auto param = CanonicalizeGemmInput(*inputA, transa, *inputB, transb, m, n, k);

kittens_mxfp8_gemm(param.A, param.B, outputD->data.dptr,
param.A_scale_inv, param.B_scale_inv,
m, n, k, is_transa, is_transb,
static_cast<int>(param.Atype),
static_cast<int>(param.Btype),
inputBias->data.dptr,
static_cast<int>(inputBias->data.dtype),
outputPreGelu->data.dptr,
static_cast<int>(outputD->data.dtype),
static_cast<int>(outputPreGelu->data.dtype),
workspace, workspaceSize, gemm_stream);
} else {
use_hipkittens = kittens_mxfp8_gemm(param.A, param.B, outputD->data.dptr,
param.A_scale_inv, param.B_scale_inv,
m, n, k, is_transa, is_transb,
static_cast<int>(param.Atype),
static_cast<int>(param.Btype),
inputBias->data.dptr,
static_cast<int>(inputBias->data.dtype),
outputPreGelu->data.dptr,
static_cast<int>(outputD->data.dtype),
static_cast<int>(outputPreGelu->data.dtype),
workspace, workspaceSize, gemm_stream);
}
if (!use_hipkittens) {
#endif
// FIXME(https://amd-hub.atlassian.net/browse/ROCM-26110): Remove this workaround once hipBLASLt supports NN/NT
// layouts for MXFP8 on gfx1250.
Expand Down
Loading