f7bd7fc

microsoft · Aug 28, 2024 · a64afdc · a64afdc
1 parent 0dc380c
commit a64afdc
Show file tree

Hide file tree

Showing 3 changed files with 3 additions and 0 deletions.
diff --git a/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu b/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
@@ -415,6 +415,7 @@ Status EfficientAttention(
   p.v_head_size = parameters.v_head_size;
   p.causal = parameters.is_unidirectional;
   p.scale = scale;
+  p.use_smooth_softmax = false;
 
   if (nullptr == data.mask_index) {
     p.seqlen_k_ptr = nullptr;

diff --git a/onnxruntime/contrib_ops/cuda/bert/packed_attention_impl.cu b/onnxruntime/contrib_ops/cuda/bert/packed_attention_impl.cu
@@ -515,6 +515,7 @@ Status FusedScaledDotProductAttentionCutlass(
   p.qk_head_size = parameters.head_size;
   p.v_head_size = parameters.v_head_size;
   p.causal = false;
+  p.use_smooth_softmax = false;
   p.scale = parameters.scale == 0.0f ? 1.f / sqrt(static_cast<float>(qk_head_size))
                                      : parameters.scale;
   p.seqlen_k_ptr = nullptr;

diff --git a/onnxruntime/contrib_ops/cuda/bert/packed_multihead_attention_impl.cu b/onnxruntime/contrib_ops/cuda/bert/packed_multihead_attention_impl.cu
@@ -693,6 +693,7 @@ Status FusedAttentionCutlass(
   p.qk_head_size = parameters.head_size;
   p.v_head_size = parameters.v_head_size;
   p.causal = false;
+  p.use_smooth_softmax = false;
   p.scale = parameters.scale == 0.0f ? 1.f / sqrt(static_cast<float>(qk_head_size))
                                      : parameters.scale;
   p.seqlen_k_ptr = nullptr;