Mocking a ROCm Environment

AIPerf collects AMD GPU telemetry through the amdsmi collector (GPUTelemetryCollectorType.AMDSMI), a local collector that calls the amdsmi Python bindings shipped with ROCm. Those bindings are not pip-installable and require a real AMD GPU plus driver, so the AMD telemetry path — the amd_* metrics and the console/CSV/plot plumbing — cannot otherwise be exercised on developer machines, in CI, or on NVIDIA hardware.

The aiperf-mock-amdsmi package (under tests/aiperf_mock_amdsmi/) installs a fake top-level amdsmi module that implements exactly the API surface the collector consumes and returns static, per-model readings. AIPerf imports it transparently — nothing in src/aiperf changes — so you can benchmark “as if on an AMD server.”

Unlike the NVIDIA DCGM path (an HTTP Prometheus endpoint faked by the mock server’s /dcgm{N}/metrics), the AMD collector is local-library-based and cannot be pointed at an HTTP endpoint. Faking the bindings is therefore the way to drive the real collector off AMD hardware.

Install

$
make install-mock-amdsmi
$
# or:
$
uv pip install -e tests/aiperf_mock_amdsmi

It is intentionally not part of the default make install.

Do not install on real AMD hardware. Installing this package places a top-level amdsmi module in your virtualenv that shadows the real ROCm bindings at the import level. Even with the dormancy gate active (i.e. AIPERF_MOCK_AMDSMI unset), import amdsmi will raise OSError instead of loading the real driver bindings, silently suppressing real AMD telemetry.

Dormancy gate

Even when installed, import amdsmi raises OSError unless AIPERF_MOCK_AMDSMI is truthy. This mirrors the real “ROCm wheel installed without a working libamd_smi.so” failure that the collector already tolerates (except (ImportError, OSError)), so an installed-but-unactivated fake degrades to “amdsmi unavailable” rather than silently claiming AMD GPUs exist.

Configuration

All settings are read from the environment at amdsmi_init() time:

Readings are static per model — they do not vary with benchmark traffic. The one exception is the energy accumulator, which advances monotonically per read so the cumulative amd_energy_consumption delta metric (baseline -> final) is non-zero.

Example

$
# Terminal 1: start the mock inference server
$
aiperf-mock-server
$
$
# Terminal 2: benchmark against it, collecting fake AMD telemetry
$
AIPERF_MOCK_AMDSMI=1 AIPERF_MOCK_AMDSMI_MODEL=mi300x AIPERF_MOCK_AMDSMI_NUM_GPUS=8 \
>
  aiperf profile -m <model> --url http://localhost:<port> \
>
  --gpu-telemetry amdsmi ...

The console summary and JSON/CSV exports will show populated amd_* metrics and platform: "amd".