[joshnguy-amd]

Address the review comments on #14041:

• amdsmi_get_gpu_compute_process_gpus returns AMDSMI_STATUS_INVAL for a NULL list, as amd-smi does, instead of reporting a count. RDC's own NULL-count call then fails here as it does on real hardware.
• amdsmi_init accepts AMDSMI_INIT_AMD_CPUS and finds no CPUs, as amd-smi does without the HSMP driver, so rdcd takes its production init path rather than the disable-CPUs retry.
• The GPU-less daemon, harness and embedded launches set RDC_DISABLE_ROCP, so the profiler module is not loaded into the hermetic tests.

Motivation

#14041 runs rdcd and rdci in CI against a fake amd-smi. Review found three places where the fake behaves differently from real amd-smi. Each difference hides RDC behavior that this suite is meant to exercise: a NULL-count call that fails on hardware, the CPU-init fallback path taken by every rdcd start, and the profiler module loading into a hermetic test. This PR makes the fake match amd-smi on those calls so the suite tests what runs in production.

Technical Details

Verified each point against the real sources before changing the fake:

• amd_smi.cc (amdsmi_get_gpu_compute_process_gpus) returns AMDSMI_STATUS_INVAL when dv_indices or num_devices is NULL. There is no count-only query. The fake now matches. This also means RdcSmiDiagnosticImpl.cc always takes its failure branch on real hardware, which is a separate RDC bug to file.
• amd_smi_system.cc (populate_amd_cpus) returns AMDSMI_STATUS_SUCCESS when esmi_init() fails, so a CPU init never fails without the HSMP driver. The fake accepts AMDSMI_INIT_AMD_CPUS and reports no CPUs. rdcd's "disable CPUs and retry" fallback in RdcEmbeddedHandler.cc is no longer taken.
• RDC_DISABLE_ROCP=yes is set in the rdcd_server fixture in conftest.py, in the daemon harness environment in test_fake_gpu.py, and in the embedded test, as test_gtest_runner.py already does.
• Two tests in test_fake_gpu_fake.py and the README now expect INVAL for a NULL list.

This PR targets users/adam360x/rdc-fake-gpu-tests and does not merge develop. #14041 conflicts with develop in rdc-ci.yml (the rdc_test runner label and timeout), which the #14041 author needs to resolve.

Issue Tracking

GitHub issue: #14041
JIRA ID:

Test Plan

• Compile fake_amdsmi.cc and run tests/ci/test_fake_gpu_fake.py, which drives the fake through ctypes.
• Run test_fake_gpu.py and tests/ci in CI (rdc_test) to check that the stricter amd-smi behavior doesn't expose new rdcd/rdci failures.

Test Result

• test_fake_gpu_fake.py: 24 passed locally against the compiled fake.
• test_fake_gpu.py and tests/ci: not run locally, because they need a built RDC. They run in CI on this PR.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.