push dsv0.8.2 version

csrc/includes/softmax.h

+/*
+Copyright The Microsoft DeepSpeed Team
+*/
+
 #pragma once
 
 #include <cuda.h>

csrc/includes/strided_batch_gemm.h

+/*
+Copyright The Microsoft DeepSpeed Team
+*/
+
 #pragma once
 
 #include <cuda.h>

csrc/includes/type_shim.h

+/*
+Copyright The Microsoft DeepSpeed Team
+*/
+
 /* Taken from NVIDIA/apex commit 855808f3fc268e9715d613f3c2e56469d8c986d8 */
 #include <ATen/ATen.h>
 

csrc/lamb/fused_lamb_cuda_kernel.cu

@@ -7,12 +7,12 @@
 #include "ATen/TensorUtils.h"
 #include "ATen/cuda/CUDAContext.h"
 #include "ATen/cuda/detail/IndexUtils.cuh"
-//#include "ATen/Type.h"
+// #include "ATen/Type.h"
 #include "ATen/AccumulateType.h"
 
 #include <iostream>
 
-//#include <helper_functions.h>
+// #include <helper_functions.h>
 #if defined(__HIP_PLATFORM_HCC__) && HIP_VERSION > 305
 #include <hip/hip_cooperative_groups.h>
 #else
@@ -105,7 +105,7 @@ __device__ void reduce_block_in_shared_memory(T* s_a, T* s_b, T* g_a, T* g_b)
 
     cg::sync(cta);
 
-#if (__CUDA_ARCH__ >= 300)
+#if (__CUDA_ARCH__ >= 300) || (defined(__HIP_PLATFORM_HCC__) && HIP_VERSION >= 502)
     if (tid < 32) {
         cg::coalesced_group active = cg::coalesced_threads();
 

csrc/transformer/cublas_wrappers.cu

+/*
+Copyright The Microsoft DeepSpeed Team
+*/
+
 #include "cublas_wrappers.h"
 
 #ifdef __HIP_PLATFORM_HCC__