Choose the right Cluster Validation Suite (CVS) config template for your workload

Choose the right Cluster Validation Suite (CVS) config template for your workload#

2026-09-25

5 min read time

Applies to Linux

For the complete field-by-field schema for each test suite’s config file, see Cluster Validation Suite (CVS) test configuration files reference. Use cvs config copy <path> --output <dest> to copy a template, or cvs config list <path> to browse templates in a directory.

Platform, health, RCCL, and other diagnostic/network configs use fixed filenames (see Burn-in / Diag and Network below). Training and inference workloads use the naming patterns in their respective sections.

Threshold pairs#

Some training and inference suites ship a matching threshold file for each config: the same basename with _threshold inserted before .json (for example …_single_threshold.json). Copy both files and keep them in the same directory. The config references its threshold file via a threshold_json field.

Burn-in / Diag#

The following suites are available for burn-in and diagnostic workloads:

Suite

Config location

List / README

Platform

input/config_file/platform/host_config.json

cvs config list platform

Health

input/config_file/health/mi300_health_config.json

cvs config list health

Preflight

input/config_file/preflight/preflight_config.json

cvs config list preflight

README_preflight_config.md

Network#

The following suites are available for network testing:

Suite

Config location

List / README

IB Perf

input/config_file/ibperf/ibperf_config.json

cvs config list ibperf

RCCL

input/config_file/rccl/rccl_config.json

cvs config list rccl

MORI

input/config_file/mori/mi35x_mori_config.json

cvs config list mori

Training#

Training workload templates use:

{gpu}_{framework}_{model}_{mode}.json

Segment

Meaning

{gpu}

Target GPU architecture — for example mi300x, mi325x, mi355x, or mi3xx.

{framework}

Training stack — for example jaxmaxtext, megatron, torchtitan.

{model}

Model identifier — for example llama3_1_70b, llama-3.1-8b, deepseek-v2-lite.

{mode}

Topology — single (one node) or distributed (multi-node).

Example filenames#

input/config_file/training/jaxmaxtext/
├── mi3xx_jaxmaxtext_llama-3.3-70b_single.json
└── mi3xx_jaxmaxtext_llama-3.3-70b_distributed.json

mi3xx_jaxmaxtext_llama-3.3-70b_distributed.json → mi3xx · jaxmaxtext · llama-3.3-70b · distributed.

Available suites#

Suite

Config location

List / README

JAX MaxText

input/config_file/training/jaxmaxtext/ — mi{gpu}_jaxmaxtext_{model}_{single|distributed}.json + mi{gpu}_jaxmaxtext_{model}_{single|distributed}_threshold.json

cvs config list training/jaxmaxtext

Config README

Megatron

input/config_file/training/megatron/ — MI300X/MI325X share mi3xx_megatron_{model}_{single|distributed}.json plus SKU mi300x_* / mi325x_* threshold files; MI355X ships Llama 3.1 8B and Llama 3.3 70B *_single.json plus matching mi355x_* thresholds. NNODES is not in JSON; container.env NIC values include examples plus <changeme>.

cvs config list training/megatron

docs/reference/configuration-files/training/megatron.rst

TorchTitan

input/config_file/training/torchtitan/ — mi{gpu}_torchtitan_{model}_{single|distributed}.json + mi{gpu}_torchtitan_{model}_{single|distributed}_threshold.json

cvs config list training/torchtitan

Schema reference

Aorta

input/config_file/aorta/aorta_benchmark.yaml

cvs config list aorta

Inference#

Inference templates use one of two filename patterns:

{gpu}_{framework}_{model}_{mode}.json
{gpu}_{framework}_{model}_{precision}_{mode}.json

Use the second form when precision (fp8, mxfp4, bf16, and similar) is a separate token after the model name.

Segment

Meaning

{gpu}

Target GPU architecture — for example mi300x, mi30x, mi355x.

{framework}

Inference stack — for example atom, vllm, sglang, pytorch_xdit.

{model}

Model identifier — for example llama_70b, deepseek-r1, flux1_dev, gpt-oss-120b.

{precision}

Optional quantization or dtype token — for example fp8, mxfp4. Omitted when precision is part of {model} or not applicable.

{mode}

Topology or workload shape — single, distributed, or disaggregated (SGLang).

Example filenames#

Without {precision}:

input/config_file/inference/xdit/
├── mi3xx_pytorch_xdit_flux1_dev_single.json
└── mi3xx_pytorch_xdit_wan22_14b_single.json

mi3xx_pytorch_xdit_flux1_dev_single.json → mi3xx · pytorch_xdit · flux1_dev · single.

With {precision}:

input/config_file/inference/vllm/
├── mi3xx_vllm_llama33-70b_fp8_single.json
└── mi3xx_vllm_llama33-70b_fp8_distributed.json

mi3xx_vllm_llama33-70b_fp8_single.json → mi3xx · vllm · llama33-70b · fp8 · single.

Available suites#

Suite

Config location

List / README

ATOM

input/config_file/inference/atom/ — mi{gpu}_atom_{model}_{precision}_{mode}.json + mi{gpu}_atom_{model}_{precision}_{mode}_threshold.json

cvs config list inference/atom

How to run · Config reference

vLLM

input/config_file/inference/vllm/ — mi3xx_vllm_{model}_{precision}_{single|distributed}.json plus sibling mi325x_*_threshold.json files

cvs config list inference/vllm

How to run · Config reference

SGLang

input/config_file/inference/sglang/ — mi{gpu}_sglang_{model}_{single|distributed|disaggregated}.json

cvs config list inference/sglang

xDiT

input/config_file/inference/xdit/ — mi{gpu}_pytorch_xdit_{model}_{single|distributed}.json

cvs config list inference/xdit
Contents