add --gpu-max-threads-per-block=1024 to fix rms_norm_kernel accuracy

e983c804 · zhuwenwen · 0dc4c8e9 · e983c804
Commit e983c804 authored Mar 19, 2025 by zhuwenwen
Hide whitespace changes
Inline Side-by-side

Showing with 4 additions and 3 deletions

cmake/utils.cmake cmake/utils.cmake +4 -3

No files found.
--- a/cmake/utils.cmake
+++ b/cmake/utils.cmake
@@ -119,10 +119,11 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)

    list(APPEND GPU_FLAGS
      "-DUSE_ROCM"
-      #"-DENABLE_FP8"
+      # "-DENABLE_FP8"
      "-U__HIP_NO_HALF_CONVERSIONS__"
      "-U__HIP_NO_HALF_OPERATORS__"
-      "-fno-gpu-rdc")
+      "-fno-gpu-rdc"
+      "--gpu-max-threads-per-block=1024")

  endif()
  set(${OUT_GPU_FLAGS} ${GPU_FLAGS} PARENT_SCOPE)
@@ -451,4 +452,4 @@ function (define_gpu_extension_target GPU_MOD_NAME)
  endif()

  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION} COMPONENT ${GPU_MOD_NAME})
-endfunction()
+endfunction()
\ No newline at end of file