Skip to content

sycl: Update gate logic for Alchemist GPUs regarding OneDNN features. #26635

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
ggerganov merged 6 commits into from
Aug 21, 2026
Merged

sycl: Update gate logic for Alchemist GPUs regarding OneDNN features. #26635

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
161626d
feat: updated gating logic of fattn-onednn.cpp
HumerousGorgon Aug 5, 2026
17421e2
verified device types
HumerousGorgon Aug 5, 2026
b291b92
Update ggml/src/ggml-sycl/fattn-onednn.cpp
HumerousGorgon Aug 6, 2026
d388f31
Improved SPDA gate, added documentation.
HumerousGorgon Aug 6, 2026
044fb7b
Added arch var to reworked gate, fixing build errors.
HumerousGorgon Aug 9, 2026
8e55842
Fix trailing whitespaces.
HumerousGorgon Aug 12, 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
19 changes: 11 additions & 8 deletions ggml/src/ggml-sycl/fattn-onednn.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,14 +21,6 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
if (!g_ggml_sycl_fa_onednn) {
return false;
}
// Battlemage (Xe2) only, for now. On other Intel archs oneDNN's fused SDPA returns wrong results
// for some shapes (e.g. head_dim=64 on Arc / xe_hpg) -- an oneDNN bug tracked upstream at
// https://github.com/uxlfoundation/oneDNN/issues/5510. Remove this hardware limitation once that
// is fixed; until then non-BMG archs fall back to the existing FA kernel.
const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
if (arch != gpu_arch::intel_gpu_bmg_g21 && arch != gpu_arch::intel_gpu_bmg_g31) {
return false;
}
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
Expand Down Expand Up @@ -60,6 +52,17 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
}
}
}
// This is the improved SPDA gate. Rather than gating Alchemist GPUs from all SPDA features, we instead target only the failing shapes.
// If the GPU being assessed isn't in the grouping below, it has full access to all SPDA shapes. Otherwise, if it's an Alchemist GPU, we block only the shapes with head sizes that fail.
// It is much easier to compare the device to a small list of failing cases than to define all the passing ones.
const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
bool support_spda = !(arch == gpu_arch::intel_gpu_dg2_g10 ||
arch == gpu_arch::intel_gpu_dg2_g11 ||
arch == gpu_arch::intel_gpu_dg2_g12);

if (!support_spda && K->ne[0] == 64) {
return false;
}
// Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
// very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
// some stacks; past the cap we fall back to the native FA kernel instead.
Expand Down
Loading