[Issue]: [hipblasLt]: hipblasLtCreate() is calling exit() if hipMalloc fails on machines with no devices

Problem Description

Summary:

hipDNN unit tests are failing prematurely on machines with no GPU devices. This is not a GTest assertion failure. The process exits with code 1 from inside hipblasLtCreate(), after printing:

Hip error: 'no ROCm-capable device is detected'(100) at /TheRock/rocm-libraries/projects/hipblaslt/library/src/amd_detail/hipblaslt.cpp:164

Backtrace shows:

   hipblasLtCreate()
   hipblaslt_plugin::HipblasltHandleFactory::createHipblasltHandle()
   hipdnnEnginePluginCreateImpl()
   EnginePlugin::createHandle()
   EnginePluginResourceManager::EnginePluginResourceManager()
   hipdnnHandle::hipdnnHandle()
   hipdnnCreate()
   expectAdditionalBackendApiForwardsToBackend()
   DirectBackendWrapperAdditionalApiForwardsToBackend_Test::TestBody()

The test is trying to validate wrapper API forwarding, but hipdnnCreate eagerly initializes engine plugin handles. Loading/initializing the hipBLASLt plugin calls hipblasLtCreate, which hard-exits when no ROCm-capable device is available or /dev/kfd is inaccessible.

The decision on whether or not to exit() should be left to the consumer of the library.

Suggested Workaround:

Skip hipblasLt handle creation in hipDNN by checking device count and return a gentler error rather than calling exit()

Operating System

Ubuntu 24.04.4 LTS

CPU

AMD EPYC 9654 96-Core Processor

GPU

None

ROCm Version

7.14

ROCm Component

hipBLASLt

Steps to Reproduce

Call hipblasLtCreate() on a machine with no GPU.

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

Start in library/src/amd_detail/hipblaslt.cpp at line 164 and reproduce hipblasLtCreate() on a machine without a GPU or accessible /dev/kfd. Trace how the no-device hipMalloc failure is handled, then verify that the library returns an error instead of exiting the process.