[DCU] cpp test compile pass

tests/cpp/util/CMakeLists.txt

@@ -2,6 +2,7 @@
 #
 # See LICENSE for license information.
 
+if(USE_CUDA)
 add_executable(test_util
                test_nvrtc.cpp
                test_string.cpp
@@ -10,6 +11,16 @@ add_executable(test_util
 
 find_package(OpenMP REQUIRED)
 target_link_libraries(test_util PUBLIC CUDA::cudart GTest::gtest_main ${TE_LIB} CUDA::nvrtc CUDNN::cudnn  OpenMP::OpenMP_CXX)
+else()
+add_executable(test_util
+               test_nvrtc_hip.cpp
+               test_string.cpp
+               ../test_common.hip)
+
+
+find_package(OpenMP REQUIRED)
+target_link_libraries(test_util PUBLIC hip::host hip::device GTest::gtest_main ${TE_LIB} OpenMP::OpenMP_CXX)
+endif()
 target_compile_options(test_util PRIVATE -O2 -fopenmp)
 
 include(GoogleTest)

transformer_engine/pytorch/csrc/type_shim.h

@@ -6,6 +6,7 @@
 #pragma once
 
 #include <ATen/ATen.h>
+#include "common/utils.cuh"
 
 // Forward/backward compatiblity hack around
 // https://github.com/pytorch/pytorch/commit/3aeb78079bcd68282fe9117088e138b77318e288
@@ -267,8 +268,6 @@
       AT_ERROR(#NAME, " not implemented for '", toString(TYPEIN), "'");               \
   }
 
-constexpr uint32_t THREADS_PER_WARP = 32;
-
 template <typename T>
 __device__ __forceinline__ T
 reduce_block_into_lanes(T *x, T val, int lanes = 1,

transformer_engine/pytorch/module/batched_linear.py

@@ -91,6 +91,12 @@ class _BatchedLinear(torch.autograd.Function):
         # TODO Support Float8 Current Scaling  # pylint: disable=fixme
         if fp8 and FP8GlobalStateManager.get_fp8_recipe().float8_current_scaling():
             raise NotImplementedError("BatchedLinear does not yet support Float8 Current Scaling")
+        # TODO Support Float8 Delayed Scaling  # pylint: disable=fixme
+        if fp8 and FP8GlobalStateManager.get_fp8_recipe().delayed():
+            raise NotImplementedError("BatchedLinear does not yet support Float8 Delayed Scaling")
+        # TODO Support Float8 Per Tensor Scaling  # pylint: disable=fixme
+        if fp8 and FP8GlobalStateManager.get_fp8_recipe().float8_per_tensor_scaling():
+            raise NotImplementedError("BatchedLinear does not yet support Float8 Per Tensor Scaling")
 
         # Make sure input dimensions are compatible
         in_features = weights[0].shape[-1]