add more limitations to use subgroup

microsoft · Jan 2, 2025 · 8328e14 · 8328e14
1 parent dff1961
commit 8328e14
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc b/onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc
@@ -569,7 +569,7 @@ Status MatMulNBits::ComputeInternal(onnxruntime::webgpu::ComputeContext& context
   // TODO: Support output_number > 1. Some cases are failed when output_number > 1.
   constexpr uint32_t output_number = 1;
   const uint32_t tile_m = M > kMinMForTileOptimization ? 4 : 1;
-  const bool use_subgroup = context.Device().HasFeature(wgpu::FeatureName::Subgroups) && context.AdapterInfo().vendor == std::string_view{"intel"};
+  const bool use_subgroup = context.Device().HasFeature(wgpu::FeatureName::Subgroups) && context.AdapterInfo().vendor == std::string_view{"intel"} && components_a == 4 && block_size == 32;
   MatMulNBitsProgram program{output_number, block_size, tile_m, gsl::narrow<int>(components_b), has_zero_points, use_subgroup};
   if (M > kMinMForTileOptimization && block_size == 32) {
     components = 1;