Merge branch 'develop_v2.7' into release_v2.7

transformer_engine/common/transformer_engine.cpp

@@ -350,6 +350,12 @@ NVTEDType nvte_tensor_type(const NVTETensor tensor) {
   return static_cast<NVTEDType>(t->dtype());
 }
 
+// Because of a HIP compiler bug, we need to disable optimizations here
+// when compiling for AMD GPUs while test_float8blockwisetensor.py.
+// Todo: remove this once the HIP compiler bug is fixed.
+#ifdef __HIP_PLATFORM_AMD__
+#pragma clang optimize off
+#endif
 NVTEShape nvte_make_shape(const size_t *data, size_t ndim) {
   NVTEShape ret;
   if (ndim == 0) {
@@ -363,6 +369,12 @@ NVTEShape nvte_make_shape(const size_t *data, size_t ndim) {
   ret.ndim = ndim;
   return ret;
 }
+// Because of a HIP compiler bug, we need to disable optimizations here
+// when compiling for AMD GPUs while test_float8blockwisetensor.py.
+// Todo: remove this once the HIP compiler bug is fixed.
+#ifdef __HIP_PLATFORM_AMD__
+#pragma clang optimize on
+#endif
 
 NVTEShape nvte_tensor_shape(const NVTETensor tensor) {
   auto *t = transformer_engine::convertNVTETensor(tensor);