Context Create Plugin#
The context_create plugin detects conditions that prevent the diagnostic
process from creating a CUDA context on a selected physical GPU. A failure
narrows the problem to GPU visibility, access, or CUDA initialization rather
than computation under load; it does not distinguish a software configuration
problem from an underlying hardware cause.
Availability and prerequisites#
context_create is available only when its plugin is loaded and the target
is a supported physical GPU. It is not part of the numbered suites; request it
by canonical name. As with every explicitly named GPU test, the built-in
software deployment checks run first.
Behavior#
For every selected GPU, the plugin resolves the device from its PCI bus ID and attempts CUDA context creation. Failure can identify CUDA initialization or device-resolution problems, inaccessible GPUs, and configuration that prevents context creation.
Exclusive-process and prohibited compute modes require special
interpretation. When context creation fails and the GPU is in either mode,
the default behavior is to skip rather than claim a hardware failure. The
ignore_exclusive parameter changes the handling of an otherwise skippable
CUDA context result; it does not change the GPU’s compute mode. Ordinary
command-line workflows should preserve the default and inspect the mode
separately.
Parameters#
Parameter |
Type |
Default |
Description |
|---|---|---|---|
|
Boolean |
|
Allow the plugin to run. |
|
Boolean |
|
Treat the CUDA result that this plugin normally regards as an exclusive/prohibited-mode skip as a failure instead. |
Examples#
Test context creation on GPUs 0 and 1:
$ dcgmi diag --run context_create --entity-id gpu:0,gpu:1
If the result is skipped because of compute mode, inspect the current mode before changing it:
$ dcgmi config --get --verbose
Results and failure conditions#
Condition |
Result or code |
Interpretation |
|---|---|---|
A CUDA context is created and destroyed on every selected GPU. |
Pass |
Basic CUDA initialization and context creation succeeded. |
Context creation fails and the GPU mode does not explain the failure. |
|
Inspect the CUDA error in the diagnostic message. This establishes a context-creation problem, not its underlying hardware or software cause. |
CUDA initialization or PCI-bus device resolution fails. |
|
The plugin could not reach the context-creation step. |
DCGM cannot obtain the selected GPU’s attributes. |
|
This is a DCGM communication or data-retrieval failure. |
Context creation is blocked by exclusive-process or prohibited compute mode under the default handling. |
Skip |
The plugin deliberately makes no health claim. Inspect the compute mode and active workload. |
|
|
The plugin was disabled by configuration. |
The run is interrupted. |
|
The test did not complete. |
No GPU was selected, or an unexpected initialization error occurs. |
|
Correct the selection or inspect the diagnostic log before drawing a conclusion about GPU health. |
See also#
dcgmi diag, dcgmi config, Diagnostics, Software Deployment Checks