set --gpu-max-threads-per-block=1024

cmake/CMakeLists.txt

@@ -1345,6 +1345,7 @@ function(onnxruntime_set_compile_flags target_name)
       # some #pragma unroll will fail, do not treat them as error
       # #warning must not be treated as error
       list(APPEND ORT_HIP_WARNING_FLAGS -Wno-error=pass-failed "-Wno-error=#warnings")
+      list(APPEND ORT_HIP_WARNING_FLAGS --gpu-max-threads-per-block=1024)
 
       foreach(FLAG ${ORT_HIP_WARNING_FLAGS})
         target_compile_options(${target_name} PRIVATE "$<$<COMPILE_LANGUAGE:HIP>:SHELL:${FLAG}>")