xDiT inference benchmark configuration for Cluster Validation Suite (CVS)

xDiT inference benchmark configuration for Cluster Validation Suite (CVS)#

2026-09-24

7 min read time

Applies to Linux

CVS ships five xDiT suites under cvs/tests/inference/xdit/. Each suite reads a JSON file from cvs/input/config_file/inference/xdit/. There is no separate threshold file; expected_results live inside benchmark_params.

  • Single-node templates run one independent docker+torchrun job on every node in the cluster file.

  • Distributed templates run one coordinated torchrun job (nnodes >= 2). Replace every <changeme> (NCCL/network fields) before running.

See Run xDiT diffusion inference tests with CVS for more information on running these tests.

Note

  • {user-id} and {home} in path strings are resolved at runtime.

  • Models must already be staged on every participating node. Prefer an absolute path in model_repo; a Hugging Face repo id requires a pre-populated cache under hf_home.

  • FLUX.1-dev and FLUX.2-dev share pytorch_xdit_flux_dev_*; pick the matching JSON.

Configuration files#

Each xDiT suite maps to one or more JSON templates; select the file that matches your model and execution mode.

Config file

Use with suite

mi3xx_pytorch_xdit_flux1_dev_single.json

pytorch_xdit_flux_dev_single

mi3xx_pytorch_xdit_flux1_dev_distributed.json

pytorch_xdit_flux_dev_distributed

mi3xx_pytorch_xdit_flux2_dev_single.json

pytorch_xdit_flux_dev_single

mi3xx_pytorch_xdit_flux2_dev_distributed.json

pytorch_xdit_flux_dev_distributed

mi3xx_pytorch_xdit_wan22_14b_single.json

pytorch_xdit_wan22_14b_single

mi3xx_pytorch_xdit_wan22_14b_diffusers_single.json

pytorch_xdit_wan22_14b_diffusers_single

mi3xx_pytorch_xdit_wan22_14b_diffusers_distributed.json

pytorch_xdit_wan22_14b_diffusers_distributed

Copy a template:

cvs config list inference/xdit
cvs config copy inference/xdit/mi3xx_pytorch_xdit_flux1_dev_single.json \
  --output ~/cvs_workspace/inference/xdit/mi3xx_pytorch_xdit_flux1_dev_single.json

File structure#

Every template has two top-level keys:

Key

Description

config

Image, model path, output dir, optional distributed rendezvous/NCCL, container_config.

benchmark_params

flux1_dev_t2i (FLUX.1 and FLUX.2) or wan22_i2v_a14b (WAN native and Diffusers).

Example: FLUX.1-dev single-node#

mi3xx_pytorch_xdit_flux1_dev_single.json (abbreviated)
{
    "config": {
        "container_image": "amdsiloai/pytorch-xdit:v25.11.2",
        "container_name": "flux-benchmark",
        "hf_token_file": "{home}/.hf_token",
        "hf_home": "{home}/.cache/huggingface",
        "output_base_dir": "{home}/cvs_flux_output",
        "model_repo": "black-forest-labs/FLUX.1-dev",
        "model_rev": "",
        "container_config": {
            "device_list": ["/dev/dri", "/dev/kfd"],
            "volume_dict": {},
            "env_dict": {}
        }
    },
    "benchmark_params": {
        "flux1_dev_t2i": {
            "prompt": "A small cat",
            "num_inference_steps": 25,
            "num_repetitions": 25,
            "height": 1024,
            "width": 1024,
            "ulysses_degree": 8,
            "ring_degree": 1,
            "use_torch_compile": true,
            "torchrun_nproc": 8,
            "expected_results": {
                "auto": { "max_avg_pipe_time_s": 10.0 },
                "mi300x": { "max_avg_pipe_time_s": 3.0 },
                "mi350": { "max_avg_pipe_time_s": 2.0 },
                "mi355": { "max_avg_pipe_time_s": 7.0 }
            }
        }
    }
}

General config parameters#

The following parameters appear in the config block and are common across all xDiT templates.

Parameter

Example

Description

container_image

amdsiloai/pytorch-xdit:v25.11.2

Image with PyTorch xDiT. FLUX.2 and WAN Diffusers templates use rocm/ufb-private:….

container_name

flux-benchmark

Docker name (distributed ranks use {container_name}-rankN).

hf_token_file

{home}/.hf_token

Hugging Face token for gated models (FLUX.2 chat template).

hf_home

{home}/.cache/huggingface

Host HF cache (mounted at /hf_home). Must contain hub/ when model_repo is a repo id.

output_base_dir

{home}/cvs_flux_output

Host directory for flux_<target>_outputs or wan_22_<target>_outputs.

model_repo

HF id or /data/models/…

Repo id (offline cache) or absolute host path (preferred). WAN Diffusers requires an absolute path.

model_rev

snapshot hash or ""

HF snapshot id when using cache mode. WAN native template pins 206a9ee1….

container_config.device_list

["/dev/dri", "/dev/kfd"]

GPU device nodes passed into the container.

container_config.volume_dict

host → container map

Extra bind mounts. FLUX.2 mounts flux2_example.py; WAN Diffusers mounts wan_i2v_example.py.

container_config.env_dict

{"NCCL_PROTO": "Simple"}

Extra environment variables inside the container.

Distributed config fields#

Present on *_distributed.json templates (FLUX.1, FLUX.2, WAN Diffusers):

Parameter

Description

nnodes

Participating node count (must be >= 2). Optional server_node_list can subset the cluster.

master_addr, master_port

torchrun rendezvous (port default 29500). master_addr is <changeme>.

nccl_ib_hca, nccl_ib_gid_index

NCCL InfiniBand/RoCE devices and GID index.

nccl_socket_ifname, gloo_socket_ifname

Ethernet interfaces for socket/Gloo fallback.

nccl_debug

NCCL log level (templates use INFO).

benchmark_params.flux1_dev_t2i#

Used by all four FLUX templates. FLUX.2 sets model_type: flux2.

Parameter

Example

Description

model_type

flux2

FLUX.2 only. Selects flux2_example.py instead of run_usp.py.

prompt, seed

A small cat, 42

Generation prompt and RNG seed.

guidance_scale

4.0

FLUX.2 guidance (FLUX.1 templates omit this).

num_inference_steps

25 / 50

Denoising steps (FLUX.1 25, FLUX.2 50).

max_sequence_length

256 / 512

Text encoder sequence length.

no_use_resolution_binning

true

Disable resolution binning.

warmup_steps, warmup_calls, num_repetitions

1, 5, 25

Warmup then measured repetitions.

height, width

1024

Output image size.

ulysses_degree, ring_degree

8, 1 (single) / 8, 2 (2-node)

Sequence-parallel layout. Product with pipefusion/TP/DP must equal nnodes × torchrun_nproc.

use_torch_compile

true

Enable torch.compile.

torchrun_nproc

8

Processes (GPUs) per node.

expected_results

mi300x.max_avg_pipe_time_s

Pass/fail on average pipe_time from results/timing.json. Keys: auto, mi300x, mi350, mi355.

benchmark_params.wan22_i2v_a14b#

This block configures the WAN 2.2 image-to-video benchmark and supports both native WAN and Diffusers execution paths.

Native WAN (mi3xx_pytorch_xdit_wan22_14b_single.json)#

Runs /app/Wan2.2/run.py. Threshold metric is max_avg_total_time_s.

Parameter

Example

Description

prompt

(long I2V prompt)

Image-to-video prompt.

size

720*1280

Frame size.

frame_num

81

Number of video frames.

num_benchmark_steps

5

Measured steps after compile/warmup.

compile

true

Enable compile on the native launcher.

torchrun_nproc

8

GPUs per node.

expected_results

mi300x.max_avg_total_time_s

Average total_time from rank0_step*.json. Requires video.mp4.

Diffusers xFuser WAN#

mi3xx_pytorch_xdit_wan22_14b_diffusers_*.json additionally set:

Parameter

Description

model_format

diffusers.

wan_diffusers_launcher

xfuser_example.

wan_diffusers_run_script

In-container path to wan_i2v_example.py (default /benchmark/wan_i2v_example.py).

wan_xfuser_auto_input_image

Generate an in-container input image when true.

wan_xfuser_install_video_deps

Install video encode deps inside the container when true.

wan_xfuser_output_type

pil.

wan_diffusers_save_video_path

/outputs/results/video_i2v.mp4.

wan_diffusers_timing_json_path

results/timing.json.

require_video_artifact

Fail parse if video_i2v.mp4 is missing.

num_inference_steps, warmup_steps

Denoising and warmup (Diffusers templates: 40 and 1).

ulysses_size, ring_size

Parallel layout. Distributed: product must equal nnodes × torchrun_nproc.

expected_results

max_avg_pipe_time_s (auto, mi325 in the shipped templates).

Volume mounts#

FLUX.1 / WAN native templates ship an empty volume_dict. Bind-mount models via model_repo as an absolute path, or rely on hf_home.

FLUX.2 mounts the in-tree example when the image lacks it:

{
    "volume_dict": {
        "/home/{user-id}/cvs/cvs/lib/inference/xdit/scripts/flux2_example.py": "/benchmark/flux2_example.py"
    }
}

WAN Diffusers mounts the xFuser example:

{
    "volume_dict": {
        "/home/{user-id}/cvs/cvs/lib/inference/xdit/scripts/wan_i2v_example.py": "/benchmark/wan_i2v_example.py"
    }
}

Adjust the host path to your CVS checkout.

Performance metrics#

GPU type is detected from rocm-smi. Lookup order: exact key → auto.

  • FLUX — average pipe_time vs max_avg_pipe_time_s; artifacts results/timing.json and flux_*.png.

  • WAN native — average total_time vs max_avg_total_time_s; rank0_step*.json and video.mp4.

  • WAN Diffusers — average pipe/epoch time vs max_avg_pipe_time_s; results/timing.json and results/video_i2v.mp4.

Shipped numbers are starting points; tune expected_results for your stack before production gating.

Troubleshooting#

``/dev/kfd not found``

Run on GPU compute nodes, not login nodes.

Container image not found locally

docker pull the configured container_image on every execution node.

Local model path not found

Stage weights on every participating node. Diffusers WAN requires model_repo as an absolute path.

Parallel degree product != world_size

Align ulysses / ring (and FLUX pipefusion/TP/DP) with nnodes × torchrun_nproc.

Missing ``timing.json`` / ``video.mp4``

The benchmark docker exit code was non-zero or artifacts were written elsewhere; inspect the log tail on the failing node.

Contents