diff --git a/.azure-pipelines/templates/ut.yml b/.azure-pipelines/templates/ut.yml
index 743c66e6e..81a74cf85 100644
--- a/.azure-pipelines/templates/ut.yml
+++ b/.azure-pipelines/templates/ut.yml
@@ -40,8 +40,8 @@ steps:
     name: PyTests
     displayName: Run pytests
     remoteScript: |
-      mpirun --allow-run-as-root -tag-output -x MSCCLPP_HOME=/root/mscclpp -x GPU_MAX_HW_QUEUES=8 -np 8 python3 -m pytest ./python/test/test_mscclpp.py -x
-      mpirun --allow-run-as-root -tag-output -x MSCCLPP_HOME=/root/mscclpp -x GPU_MAX_HW_QUEUES=8 -np 8 python3 -m pytest ./python/test/test_fp8_accum.py -x
+      mpirun --allow-run-as-root -tag-output -x MSCCLPP_HOME=/root/mscclpp -x HSA_NO_SCRATCH_RECLAIM=1 -x GPU_MAX_HW_QUEUES=8 -np 8 python3 -m pytest ./python/test/test_mscclpp.py -x
+      mpirun --allow-run-as-root -tag-output -x MSCCLPP_HOME=/root/mscclpp -x HSA_NO_SCRATCH_RECLAIM=1 -x GPU_MAX_HW_QUEUES=8 -np 8 python3 -m pytest ./python/test/test_fp8_accum.py -x
 
 - template: stop.yml
   parameters:
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7b8316fb6..c3fb8fba5 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -112,11 +112,20 @@ if(MSCCLPP_ENABLE_COVERAGE)
     if(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
         message(STATUS "Code coverage enabled")
         
-        # Add coverage flags to C++ targets only (not CUDA)
-        add_compile_options($<$<COMPILE_LANGUAGE:CXX>:--coverage>)
+        # Add coverage flags to C++ targets only (not CUDA). For ROCm, keep
+        # coverage on the host compilation so HIP device linking does not look
+        # for gcov runtime symbols.
+        if(MSCCLPP_USE_ROCM)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-Xarch_host>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:--coverage>)
+            add_link_options($<$<LINK_LANGUAGE:CXX>:-Xarch_host>)
+            add_link_options($<$<LINK_LANGUAGE:CXX>:--coverage>)
+        else()
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:--coverage>)
+            add_link_options($<$<LINK_LANGUAGE:CXX>:--coverage>)
+        endif()
         add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-O0>)
         add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-g>)
-        add_link_options($<$<LINK_LANGUAGE:CXX>:--coverage>)
         
         # Find lcov
         find_program(LCOV_PATH lcov)
diff --git a/src/core/gpu_ipc_mem.cc b/src/core/gpu_ipc_mem.cc
index 49dfcf822..b885a1ac4 100644
--- a/src/core/gpu_ipc_mem.cc
+++ b/src/core/gpu_ipc_mem.cc
@@ -186,6 +186,8 @@ UniqueGpuIpcMemHandle GpuIpcMemHandle::create(const CUdeviceptr ptr) {
   if (err == cudaSuccess) {
     handle->typeFlags |= GpuIpcMemHandle::Type::RuntimeIpc;
   } else {
+    // The VMM fallback below handles this failure, so do not leak it as CUDA's last error.
+    (void)cudaGetLastError();
     WARN(GPU, "Failed to create runtime CUDA IPC handle for pointer ", (void*)basePtr,
          ": error=", static_cast<int>(err), " (", std::string(cudaGetErrorString(err)), ")");
   }