Skip to content

FlashAttention: use subgroup size 32 on Intel Xe2 and Xe3 #33808

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
Jiajia Qin (qjia7) merged 1 commit into from
Nov 11, 2026
Merged

FlashAttention: use subgroup size 32 on Intel Xe2 and Xe3 #33808

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
16 changes: 14 additions & 2 deletions onnxruntime/contrib_ops/webgpu/bert/flash_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -633,10 +633,11 @@ Status ApplyFlashAttentionPrefill(onnxruntime::webgpu::ComputeContext& context,
bool has_head_sink,
const Tensor* head_sink) {
const auto& adapter_info = context.AdapterInfo();
const bool is_intel_xe2 = adapter_info.vendor == std::string_view{"intel"} &&
const bool is_intel = adapter_info.vendor == std::string_view{"intel"};
const bool is_intel_xe2 = is_intel &&
(adapter_info.architecture == std::string_view{"xe-2lpg"} ||
adapter_info.architecture == std::string_view{"xe-2hpg"});
const bool is_intel_xe3 = adapter_info.vendor == std::string_view{"intel"} &&
const bool is_intel_xe3 = is_intel &&
(adapter_info.architecture == std::string_view{"xe-3lpg"} ||
adapter_info.architecture == std::string_view{"xe-3lpg-xs"});
// Default: 64 threads, 16 KV rows per step.
Expand Down Expand Up @@ -700,6 +701,17 @@ Status ApplyFlashAttentionPrefill(onnxruntime::webgpu::ComputeContext& context,
program.AddInputs({{head_sink, ProgramTensorMetadataDependency::Type}});
}
program.AddOutputs({{attn_output, ProgramTensorMetadataDependency::TypeAndRank, 4}});

// An explicit subgroup size lets Tint use a constant lane clamp bound for subgroupShuffle, which keeps IGC from
Comment thread
jchen10 marked this conversation as resolved.
Resolved
// falling back to indirect register addressing.
constexpr uint32_t kIntelSubgroupSize = 32;
if ((is_intel_xe2 || is_intel_xe3) && context.HasFeature(wgpu::FeatureName::Subgroups) &&
context.HasFeature(wgpu::FeatureName::SubgroupSizeControl) &&
adapter_info.subgroupMinSize <= kIntelSubgroupSize && kIntelSubgroupSize <= adapter_info.subgroupMaxSize &&
workgroup_size % kIntelSubgroupSize == 0) {
program.SetSubgroupSize(kIntelSubgroupSize);
}

program.SetDispatchGroupSize(dispatch_size)
.SetWorkgroupSize(workgroup_size)
.CacheHint(has_attention_bias, parameters.head_size_, parameters.num_heads_, parameters.is_unidirectional_,
Expand Down
Loading