Fix rocm build issue

Author: Binyang2014

apps/nccl/src/allreduce.hpp

@@ -1237,7 +1237,6 @@ class AllreduceNvlsPacket : public mscclpp::AlgorithmBuilder {
 
   size_t scratchBufferSize_;
   std::shared_ptr<char> scratchBuffer_;
-  const int nSegmentsForScratchBuffer_ = 2;
   const size_t nvlsBufferSize_ = (1 << 30);
 
   std::shared_ptr<uint32_t> deviceFlag_;

src/include/execution_kernel.hpp

@@ -1177,7 +1177,7 @@ class ExecutionKernel {
       case DataType::FP8_E4M3:
         executionKernel<__fp8_e4m3, PacketType, ReuseScratch><<<nthreadblocks, nthreads, sharedMemSize, stream>>>(
             rank, (__fp8_e4m3*)src, (__fp8_e4m3*)dst, (__fp8_e4m3*)scratch, scratchOffset, scratchChunkSize, plan,
-            semaphores, flag
+            semaphores, localMemoryIdBegin, flag
 #if defined(ENABLE_NPKIT)
             ,
             NpKit::GetGpuEventCollectContexts(), NpKit::GetCpuTimestamp());
@@ -1188,7 +1188,7 @@ class ExecutionKernel {
       case DataType::FP8_E5M2:
         executionKernel<__fp8_e5m2, PacketType, ReuseScratch><<<nthreadblocks, nthreads, sharedMemSize, stream>>>(
             rank, (__fp8_e5m2*)src, (__fp8_e5m2*)dst, (__fp8_e5m2*)scratch, scratchOffset, scratchChunkSize, plan,
-            semaphores, flag
+            semaphores, localMemoryIdBegin, flag
 #if defined(ENABLE_NPKIT)
             ,
             NpKit::GetGpuEventCollectContexts(), NpKit::GetCpuTimestamp());