Skip to content

ggml-hrx: no V transpose for MLA strided V (GLM-4.7-Flash prompts >= 512 tokens: PPL 315,664 -> 5.92) #70

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
bong-water-water-bong merged 1 commit into from
Oct 2, 2026
Merged

ggml-hrx: no V transpose for MLA strided V (GLM-4.7-Flash prompts >= 512 tokens: PPL 315,664 -> 5.92) #70

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
6 changes: 5 additions & 1 deletion ggml/src/ggml-hrx/dispatch_registration/common/dispatch-flash-attention.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -436,7 +436,11 @@ static DispatchBinding prepare_flash_attention_value(const DispatchMatchContext
DispatchMatch & dispatch_match,
KernelSpecialization & attention) {
const int64_t columns = match.key_value_head_count * match.value_head_size;
if (match.query_token_count < 256 || match.key_value_token_count < 512 ||
// ggml_copy_transpose_f16 reads `columns`-wide contiguous rows. MLA (deepseek2/GLM) stores V with the QK
// head-size stride (the matcher accepts that layout, #95), so its rows are wider than `columns`: transposing
// would read misaligned rows. Such a V keeps the row-major layout, whose kernel takes the real row stride.
const bool contiguous_rows = match.value->nb[1] == static_cast<size_t>(columns) * sizeof(ggml_fp16_t);
if (!contiguous_rows || match.query_token_count < 256 || match.key_value_token_count < 512 ||
match.key_value_token_count % 32 != 0 || match.key_value_token_count > kCopyTransposeF16MaxExtent ||
columns % 32 != 0 || columns > kCopyTransposeF16MaxExtent) {
return { match.value->id, 0, match.value->byte_count };
Expand Down
Loading