Skip to content

gfx1250: fix MXFP8 scale_inv shape mismatch in C++ tests #691

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

Draft
matthiasdiener wants to merge 1 commit into
base: dev
Choose a base branch
from
+23 −6
Draft

gfx1250: fix MXFP8 scale_inv shape mismatch in C++ tests #691

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
29 changes: 23 additions & 6 deletions tests/cpp/test_common.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,14 +160,25 @@ std::pair<scale_inv_meta, scale_inv_meta> get_scales(const NVTEShape& shape,

scale_inv_meta ret_rowwise, ret_colwise;

size_t align_Y_rowwise = scale_tensor_alignment_Y_rowwise;
size_t align_X_rowwise = scale_tensor_alignment_X_rowwise;
size_t align_Y_colwise = scale_tensor_alignment_Y_colwise;
size_t align_X_colwise = scale_tensor_alignment_X_colwise;
#ifdef __HIP_PLATFORM_AMD__
// gfx1250 MX pre-swizzle requires MXFP8 scales padded to a multiple of 4 in both dims
if (getDeviceComputeCapability() == 125) {
align_Y_rowwise = align_X_rowwise = align_Y_colwise = align_X_colwise = 4;
}
#endif

const size_t block_size_X_rowwise = 32;
size_t scale_dim_Y_rowwise = DIVUP_TO_MULTIPLE(first_dim, scale_tensor_alignment_Y_rowwise);
size_t scale_dim_X_rowwise = DIVUP_TO_MULTIPLE(DIVUP(last_dim, block_size_X_rowwise), scale_tensor_alignment_X_rowwise);
size_t scale_dim_Y_rowwise = DIVUP_TO_MULTIPLE(first_dim, align_Y_rowwise);
size_t scale_dim_X_rowwise = DIVUP_TO_MULTIPLE(DIVUP(last_dim, block_size_X_rowwise), align_X_rowwise);
ret_rowwise.shape = {scale_dim_Y_rowwise, scale_dim_X_rowwise};

const size_t block_size_Y_colwise = 32;
size_t scale_dim_Y_colwise = DIVUP_TO_MULTIPLE(DIVUP(first_dim, block_size_Y_colwise), scale_tensor_alignment_Y_colwise);
size_t scale_dim_X_colwise = DIVUP_TO_MULTIPLE(last_dim, scale_tensor_alignment_X_colwise);
size_t scale_dim_Y_colwise = DIVUP_TO_MULTIPLE(DIVUP(first_dim, block_size_Y_colwise), align_Y_colwise);
size_t scale_dim_X_colwise = DIVUP_TO_MULTIPLE(last_dim, align_X_colwise);
ret_colwise.shape = {scale_dim_Y_colwise, scale_dim_X_colwise};

ret_rowwise.type = DType::kFloat8E8M0;
Expand Down Expand Up @@ -1322,8 +1333,14 @@ std::array<size_t, 4> get_scale_tensor_dims(const size_t rows,
alignment_X = is_rowwise ? nvfp4_scale_tensor_alignment_X_rowwise
: nvfp4_scale_tensor_alignment_X_colwise;
} else {
alignment_Y = 1;
alignment_X = 1;
// MXFP8: gfx1250 MX pre-swizzle requires scales padded to a multiple of 4 in both dims
if (getDeviceComputeCapability() == 125) {
alignment_Y = 4;
alignment_X = 4;
} else {
alignment_Y = 1;
alignment_X = 1;
}
}
#else
const size_t alignment_Y = is_rowwise
Expand Down
Loading