Cluster Validation Suite (CVS) test configuration files reference#
2026-09-25
3 min read time
Each CVS test has a corresponding JSON configuration file. You must configure the JSON file for each test you want to run in CVS.
The test configuration files are in the cvs/input/config_file directory of the cloned repo. You can go to each directory and edit the parameters as necessary for your testing requirements.
Note
Ensure ROCm is installed correctly, and the GPU drivers are loaded.
Cluster file#
In addition to a per-test --config_file, every cvs run invocation needs a --cluster_file that declares the SSH credentials, the node list, and the execution backend (baremetal or container). See Cluster Validation Suite (CVS) cluster file: configuration and backend selection for the full schema, the container block reference, and which suites consume the orchestrator today.
The cluster file is shared across all test suites and must be configured once per cluster before running any CVS tests.
Test configuration files#
Configuration references are grouped by test category. Each page documents parameters and example JSON for suites under cvs/input/config_file/.
Burn-in / Diag#
Network#
InfiniBand (IB Perf) — IB bandwidth and latency benchmarks
RCCL — multi-node collective communication performance
MORI (RDMA Performance) — RDMA read/write bandwidth and latency
Training#
JAX MaxText — MaxText pre-training (single-node and distributed)
Megatron — Llama and DeepSeek, single-node and distributed
TorchTitan — TorchTitan pre-training (single-node and distributed)
Aorta (Distributed Training) — Aorta RCCL/training throughput benchmark
Inference#
ATOM (vLLM Benchmarking) — ATOM inference benchmarks
vLLM Inference — vLLM serving throughput and latency
SGLang Inference — single-node, distributed, and disaggregated LLM serving
xDiT Inference — FLUX.1/FLUX.2 text-to-image and WAN 2.2 image-to-video