Extended Utility Diagnostics (EUD) Plugin#

The EUD plugin launches NVIDIA EUD executables and integrates their results into a DCGM diagnostic response. It exposes two canonical test names:

eud: Run the MODS-based EUD for supported GPUs.
cpu_eud: Run EUD for a supported NVIDIA CPU.

These are deeper platform diagnostics than DCGM’s CUDA workload plugins. They are intended for administrator-led isolation after ordinary readiness, memory, interconnect, or stress tests indicate a possible hardware problem.

Availability and prerequisites#

The corresponding external executable must be present and advertised by the diagnostic plugin. GPU and CPU support is product-specific. Both tests require root execution; the diagnostic module has special handling to execute EUD as root when its normal service account is non-root.

This identity transition is specific to the diagnostic module. For the Linux credential and privilege model that controls executable access, see process credentials and administrative privileges.

Behavior#

EUD can pause other DCGM module activity while the external diagnostic runs. It can consume or temporarily take control of the target hardware and must not be run on GPUs or CPUs serving ordinary workloads.

DCGM reports setup failures separately from a failed hardware test. Setup and execution errors include an unavailable or incorrectly permissioned binary, non-root execution, spawn failure, timeout, a process that does not terminate, or a nonzero executable status. DCGM_FR_EUD_TEST_FAILED indicates that EUD itself reported failure.

Parameters#

EUD wrapper parameters#
Parameter	Type	Default	Description
`is_allowed`	Boolean	`true`	Allow the named EUD test to run. `false` produces a skipped result.
`full_profile`	Boolean	See description	Accepted by the DCGM 4.6 interface, but does not independently override the profile selected by the numbered diagnostic suite.
`parse_mle`	Boolean	`false`	Request parsing of supported external diagnostic log information. The information available is package-specific.
`passthrough_args`	String	None	Arguments interpreted by the installed EUD executable. Their names and meanings are not part of the DCGM interface.
`original_level`	Integer	Suite-provided	Carries the original numbered-suite selection to the wrapper. It is not intended as a replacement for `dcgmi diag --run`.

Repeated eud.passthrough_args or cpu_eud.passthrough_args assignments are concatenated; other diagnostic parameters may be defined only once. Use passthrough arguments only with documentation for the installed EUD package.

Examples#

Run GPU EUD on GPU 0:

$ dcgmi diag --run eud --entity-id gpu:0

Run CPU EUD on CPU 0:

$ dcgmi diag --run cpu_eud --entity-id cpu:0

An unavailable-test result means that the required plugin, executable, target product, or privilege was not available; it is not a passing or failing EUD hardware result.

Results and failure conditions#

EUD wrapper results#
Condition	Result or code	Interpretation
The external diagnostic reports passing results for the selected entities.	Pass	EUD completed the product-specific checks that it reported. Internal test coverage is defined by the installed EUD package.
EUD reports a failed entity or system result.	`DCGM_FR_EUD_TEST_FAILED`; fail	This is a completed EUD diagnostic finding rather than a wrapper setup failure.
The executable is absent.	Skip	EUD is unavailable on this installation; no hardware result was produced.
The executable fails its ownership, file-type, or permission checks.	Permission-specific `DCGM_FR_*` code; fail	Correct the executable installation rather than treating the result as a GPU or CPU defect.
The diagnostic process is not running with the required privilege.	`DCGM_FR_EUD_NON_ROOT_USER`; fail	EUD did not start.
DCGM telemetry cannot be paused before EUD or resumed afterward.	`DCGM_FR_PAUSE_RESUME_FAILED`; fail	A pause failure prevents a reliable run. After a resume failure, restore DCGM telemetry as directed by the error message.
The external process cannot be started or exits before establishing a usable diagnostic session.	`DCGM_FR_EUD_SPAWN_FAILURE`; fail	Inspect the EUD stdout, stderr, and diagnostic logs.
The external process stops providing progress and reaches its timeout.	`DCGM_FR_EUD_TIMEOUT`; fail	Confirm that no EUD process remains and retain its logs.
GPU EUD requests recovery followed by another run.	`DCGM_FR_RETEST_REQUESTED`	Follow the recovery guidance in the message and rerun EUD; do not interpret the current run as a clean hardware pass.
The external process exits with another nonzero status.	`DCGM_FR_EUD_NON_ZERO_EXIT_CODE`	Use the reported entity results together with the EUD logs. The status identifies abnormal execution, not a specific hardware defect.
The plugin is disabled.	`DCGM_FR_TEST_DISABLED`; skip	EUD was not run.