The inference config¶
MLPerf Inference submissions, written as the flat field set the submission checker expects, with node hardware lifted to the top level.
mlperf-sysinfo init inference writes a commented starter config covering the
same fields.
A working config¶
profile: inference
output:
dir: results/h100_run1
system:
name: H100x8
category: datacenter
availability: available
accelerator: cuda
cooling: air
type_detail: "1-node 8x H100"
nodes:
include_local: true
submission:
submitter: MyOrg
contact: mlperf@myorg.example
division: closed
notes:
hardware: "8x H100 SXM, single node"
software: "TensorRT-LLM 0.12"
Required¶
check stops if any of these is missing or still holding starter text.
| Field | Why |
|---|---|
system.name |
Identifier for the system under test |
system.category |
Whether this is a datacenter or an edge system |
system.availability |
Availability at submission time, e.g. available, preview or rdi |
submission.submitter |
Organisation making the submission |
submission.contact |
Contact email for questions about this submission |
submission.division |
Either closed or open |
Worth filling in¶
Absence is a warning, not a stop. Starter text in one of them is a stop, wherever it turns up.
| Field | Why |
|---|---|
system.accelerator |
Without it, accelerators are not probed at all |
submission.notes.software |
Software detail no probe can report, such as versions, flags or patches |
submission.notes.hardware |
Hardware detail no probe can report, such as interconnect topology or firmware |
Options¶
system¶
| Key | Type | Default | Notes |
|---|---|---|---|
name |
string | — | Required. Identifier for the system under test |
category |
string | — | Required. datacenter or edge. Written as system_type |
availability |
string | — | Required. e.g. available, preview, rdi. Written as status |
accelerator |
enum | none |
Recommended. cuda | rocm | xpu | tpu | none. Left out it defaults to none, which skips accelerator probing entirely, so a GPU or TPU system left at the default captures no accelerator at all. See Accelerators for what each value probes |
cooling |
string | — | e.g. air, liquid, passive |
type_detail |
string | — | Free text, written as system_type_detail. Rack or chassis detail the field list has no column of its own for |
size |
string | computed | Overrides system_size. Rarely needed. See Sample outputs for what this profile computes |
nodes¶
| Key | Type | Default | Notes |
|---|---|---|---|
include_local |
bool | false |
Whether the machine running the command is part of the system |
ssh |
list | [] |
user@host or user@host:port |
ssh_key_preconfigured |
bool | false |
Key auth is already set up; skip the key-file lookup |
groups |
map | — | Function name → list of {match, count} |
Why include_local defaults to false
An orchestrator machine driving a cluster should not describe itself by
accident. A config with none of include_local: true, an ssh entry, or a
serving.node is rejected, because there would be nothing to collect.
power¶
Optional, opt-in. Enables Redfish capture from the BMC.
| Key | Type | Default | Notes |
|---|---|---|---|
redfish.endpoint |
string | — | BMC address |
redfish.username |
string | — | Use ${VAR} |
redfish.password |
string | — | Use ${VAR} |
submission¶
| Key | Type | Default | Notes |
|---|---|---|---|
submitter |
string | — | Required. Organisation making the submission |
contact |
string | — | Required. Contact email for questions about the submission. Written as submitter_contact |
division |
string | — | Required. Either closed or open |
notes.hardware |
string | — | Recommended. Written as hw_notes |
notes.software |
string | — | Recommended. Written as sw_notes |
notes.other_hardware |
string | — | Written as other_hardware |
What this profile ignores¶
There is no serving or run section here. An inference capture has nowhere
to put a URL, a run configuration, or submission.container_link, and those
values are not read.
serving.node is the exception
Every profile treats it as a machine to collect from, on the grounds that a
node only mentioned as where the server runs is still part of the system. So
a config carried over from endpoints will quietly capture that node too,
and number_of_nodes will say so. Delete the serving block, or move the
host to nodes.ssh where you can see it.