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
+18 −13
Merged

hipblaslt Fallback hotfix #652

Show file tree
Hide file tree
Changes from all 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
29 changes: 17 additions & 12 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 @@ -1999,18 +1999,23 @@ 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) {
if (is_mxfp8) {
NVTE_CHECK(inputBias->data.dptr == nullptr,
"hipBLASLt MXFP8 GEMM does not support bias");
}
#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
2 changes: 1 addition & 1 deletion transformer_engine/jax/cpp_extensions/gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -653,7 +653,7 @@ def _dims_are_consecutive(dims):
n = reduce(operator.mul, rhs_non_contracting_shape)
k = lhs_contracting_size
layout = ("T" if lhs_is_transposed else "N") + ("T" if rhs_is_transposed else "N")
workspace_size = _hipkittens_workspace_bytes(m, n, k, layout)
workspace_size = max(_hipkittens_workspace_bytes(m, n, k, layout), get_cublas_workspace_size_bytes())
else:
# Declare cuBLAS workspace
workspace_size = get_cublas_workspace_size_bytes()
Expand Down
Loading