Skip to content

[SYCLTLA] Fix FlashAttention FWD performance on PVC #2415

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
merged 3 commits into from
Dec 1, 2025
Merged

[SYCLTLA] Fix FlashAttention FWD performance on PVC #2415

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4477262
fix perf
LuFinch Nov 27, 2025
4c8e6b0
remove intel_gpu_bmg_g31 due to stock pytorch CI is 2025.1
LuFinch Nov 28, 2025
57c6114
refine
LuFinch Nov 28, 2025
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
4 changes: 2 additions & 2 deletions src/ATen/native/transformers/xpu/flash_attn/sycltla/kernel/xe_sdpa_fwd_bshd.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -191,7 +191,7 @@ class FMHAPrefill {
}

// Find the length of the longest non masked sequence within that subgroup
int calculate_longest_non_masked_length(
CUTLASS_DEVICE int calculate_longest_non_masked_length(
const int& seq_len_kv,
const int& seq_len_qo,
const int& last_seq_coord,
Expand Down Expand Up @@ -222,7 +222,7 @@ class FMHAPrefill {
}

template <class Tensor>
void handle_corner_cases(
CUTLASS_DEVICE void handle_corner_cases(
Tensor& tSr,
const int& thread_idx,
const int& SubgroupSize,
Expand Down
7 changes: 3 additions & 4 deletions src/ATen/native/transformers/xpu/flash_attn/sycltla/mha_bwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1461,18 +1461,17 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> flash_attention_backward_sycltla(
.get_info<
sycl::ext::oneapi::experimental::info::device::architecture>();
constexpr auto supported_architectures =
std::array<sycl::ext::oneapi::experimental::architecture, 4>{
std::array<sycl::ext::oneapi::experimental::architecture, 3>{
sycl::ext::oneapi::experimental::architecture::intel_gpu_pvc,
sycl::ext::oneapi::experimental::architecture::intel_gpu_pvc_vg,
sycl::ext::oneapi::experimental::architecture::intel_gpu_bmg_g21,
sycl::ext::oneapi::experimental::architecture::intel_gpu_bmg_g31};
sycl::ext::oneapi::experimental::architecture::intel_gpu_bmg_g21};
if (std::find(
supported_architectures.begin(),
supported_architectures.end(),
device_architecture) == supported_architectures.end()) {
TORCH_CHECK(
false,
"XPU device architecture does not support flash attention backward. Supported architectures are: intel_gpu_pvc, intel_gpu_pvc_vg, intel_gpu_bmg_g21, intel_gpu_bmg_g31.");
"XPU device architecture does not support flash attention backward. Supported architectures are: intel_gpu_pvc, intel_gpu_pvc_vg, intel_gpu_bmg_g21.");
}

auto grad_query = at::empty_like(query);
Expand Down
31 changes: 15 additions & 16 deletions src/ATen/native/transformers/xpu/flash_attn/sycltla/mha_fwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -333,19 +333,19 @@ void run_mha_fwd_(
TileShapeOutPut,
SubgroupLayout,
PipelineStages);
} else {
constexpr int PipelineStages = 2;
using TileShapeQK = Shape<_256, _32, _64>;
using TileShapePV = Shape<_256, _32, _32>;
using TileShapeOutPut = Shape<_256, _128, _32>;
using SubgroupLayout = Layout<Shape<_16, _1, _1>, Stride<_1, _1, _1>>;
run_mha_fwd_specialized(
TileShapeQK,
TileShapePV,
TileShapeOutPut,
SubgroupLayout,
PipelineStages);
}

constexpr int PipelineStages = 2;
using TileShapeQK = Shape<_256, _32, _64>;
using TileShapePV = Shape<_256, _32, _32>;
using TileShapeOutPut = Shape<_256, _128, _32>;
using SubgroupLayout = Layout<Shape<_16, _1, _1>, Stride<_1, _1, _1>>;
run_mha_fwd_specialized(
TileShapeQK,
TileShapePV,
TileShapeOutPut,
SubgroupLayout,
PipelineStages);
} else if (headdim == 192) {
constexpr int PipelineStages = 2;
using TileShapeQK = Shape<_256, _64, _64>;
Expand Down Expand Up @@ -537,18 +537,17 @@ flash_attention_forward_sycltla(
.get_info<
sycl::ext::oneapi::experimental::info::device::architecture>();
constexpr auto supported_architectures =
std::array<sycl::ext::oneapi::experimental::architecture, 4>{
std::array<sycl::ext::oneapi::experimental::architecture, 3>{
sycl::ext::oneapi::experimental::architecture::intel_gpu_pvc,
sycl::ext::oneapi::experimental::architecture::intel_gpu_pvc_vg,
sycl::ext::oneapi::experimental::architecture::intel_gpu_bmg_g21,
sycl::ext::oneapi::experimental::architecture::intel_gpu_bmg_g31};
sycl::ext::oneapi::experimental::architecture::intel_gpu_bmg_g21};
if (std::find(
supported_architectures.begin(),
supported_architectures.end(),
device_architecture) == supported_architectures.end()) {
TORCH_CHECK(
false,
"XPU device architecture does not support flash attention. Supported architectures are: intel_gpu_pvc, intel_gpu_pvc_vg, intel_gpu_bmg_g21, intel_gpu_bmg_g31.");
"XPU device architecture does not support flash attention. Supported architectures are: intel_gpu_pvc, intel_gpu_pvc_vg, intel_gpu_bmg_g21.");
}

auto problem_shape = ProblemShapeRegular(
Expand Down
Loading