lite

build/Linux/Release/amdgpu/onnxruntime/core/providers/rocm/math/softmax_blockwise_impl.cuh

@@ -25,7 +25,7 @@ namespace onnxruntime {
 namespace rocm {
 
 constexpr int ALIGN_BYTES = 16;
-const int max_threads = 1024;
+const int max_threads = 256; //const int max_threads = 1024;
 
 dim3 SoftMax_getBlockSize(int ILP, uint64_t dim_size) {
   uint64_t block_size = 1;