Issue codes#

Every machine-readable diagnostic code HPX can emit, generated from the registry in helia_profiler.results.issues.

Run-validity issues#

Emitted by run evaluation into summary.json (issues[]) and result_manifest.json. Severity error makes the run invalid; warning makes it degraded. Two codes are mode-dependent: whether the broken number is the measurement of record decides fatal-vs-warn.

Code	Severity	Description
`pmu.counter_overflow`	`error`	One or more PMU counters overflowed during capture.
`pmu.missing`	`error`	The run has no PMU result.
`power.gate_below_minimum`	`error`	Measured power gate is shorter than the minimum accepted window.
`power.gate_duration_mismatch`	`warning`	Measured power-gate duration does not agree with the expected window.
`power.gate_duration_unverifiable`	`warning`	Power-gate duration cannot be verified because clean inference timing is invalid.
`power.gate_edges_missing`	`error`	GPIO-gated power capture is missing a gate edge.
`power.gate_not_lowered`	`error`	Power firmware did not confirm GATE low.
`power.observation_degraded`	`warning`	Power observation is diagnostic and not valid for efficiency metrics.
`power.observation_invalid`	`error`	Power observation integrity is invalid.
`power.observation_missing`	`error`	Power observation is missing.
`power.on_device_count_mismatch`	`error`	On-device measurement count differs from completed work.
`power.on_device_measurement_missing`	`error`	Internal power mode has no on-device measurement.
`power.on_device_overflow`	`error` (internal) / `warning` (external)	On-device power monitor reported accumulator overflow.
`power.plan_count_mismatch`	`error`	Power firmware requested count differs from the host plan.
`power.terminal_error`	`error`	Power firmware reported an error.
`power.terminal_incomplete`	`error`	Power firmware completed a different inference count than requested.
`power.terminal_missing`	`error`	Dedicated power firmware did not publish terminal status.
`power.window_clock_exceeds_host_time`	`warning`	Firmware-reported window is longer than the host wall time that contained it.
`power.window_clock_frozen`	`error` (internal) / `warning` (external)	Power firmware reported zero elapsed time for completed inferences.
`power.window_clock_mismatch`	`warning`	Firmware-reported window duration does not agree with the independently measured window.
`power.window_observer_mismatch`	`error`	The instrument-timed gate and the firmware's own window clock disagree about the same physical window.
`profile.clean_window_check_inoperative`	`warning`	The clean window's partial-stall check could not run; absence of stalls is not evidence of a healthy window.
`profile.clean_window_clock_rate_low`	`warning`	The clean window's cycle counter ran far below its expected rate, measured against an independent clock.
`profile.clean_window_frozen`	`warning`	The clean window completed inferences in zero elapsed time; the clock timing it never advanced.
`profile.clean_window_stalled`	`warning`	The clean-inference window's cycle counter stalled; derived timings understate the true per-inference time.

Comparability issues#

Emitted by hpx compare when assessing whether two runs may be compared. Severity governs scope: blocking stops the whole comparison, layer_blocking omits per-layer deltas, metric_blocking omits one metric group, and informative annotates without blocking anything.

Code	Severity	Description
`identity.model_mismatch`	`blocking`	Model SHA-256 differs; run-level performance deltas are not comparable.
`metric.power_integrity_invalid`	`metric_blocking`	Power metrics omitted because a power result's integrity is not valid.
`result.degraded`	`informative`	A result is degraded; affected metrics should be interpreted cautiously.
`result.incomplete`	`blocking`	A result bundle is not complete; no comparison is possible.
`result.invalid`	`blocking`	A result is invalid and cannot be compared.
`result.invalid_pmu_overflow`	`blocking`	A legacy result without a manifest has PMU counter overflow.
`topology.layer_count_mismatch`	`layer_blocking`	Per-layer deltas omitted because layer counts differ.
`topology.operation_sequence_mismatch`	`layer_blocking`	Per-layer deltas omitted because operation sequences differ.

Parameterized families#

These codes embed a comparison dimension name. The metric.power_… family doubles the power prefix because the dimension names themselves start with power_; that is the shipped wire format, pinned until a deliberate wire-format change renames it.

metric.power_<dimension>_mismatch (metric_blocking) — Power metrics omitted because a power comparison dimension differs between the runs. Dimensions: power_scope, power_mode, power_firmware, power_monitor, power_lockstep, power_clean_window_probe, power_firmware_fingerprint
dimension.<dimension>_differs (informative) — A comparison dimension differs between the runs; deltas remain comparable but should be read in that light. Dimensions: hpx_version, engine, board, soc, cpu_clock, toolchain, compiler_version, system_clock_hz, run_summary_schema_version, run_metadata_schema_version, transport, arena_location, weights_location