Qwen-Image text-to-image (H100)

Hardware: gpu_1x_h100. Save the files under Source files into a folder named qwen-image-text-to-image, then follow the steps below.

Runs Qwen/Qwen-Image with diffusers QwenImagePipeline to render a prompt containing in-image text to a PNG. The model is ungated and Apache-2.0 licensed: no Hugging Face token or API key is needed.

The ~20B-parameter transformer uses ~41 GB in bf16; together with the text encoder, VAE and inference workspace it fits an 80 GB H100 (gpu_1x_h100). Measured peak: 60246 MiB reserved (staging, 2026-10-08).

Files

supercomputer.json — pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel with the Hugging Face cache on the qwen-image-weights volume at /workspace/.cache/huggingface, persistent across super stop and resume.
run.sh — installs exact released dependency versions (diffusers 0.35.1), loads the bf16 pipeline on CUDA, and writes /workspace/out/qwen_image.png. Uses the pipeline's default image dimensions and inference steps, with the upstream recommended true CFG scale of 4 and a blank negative prompt.

Run

cd examples/qwen-image-text-to-image
super create --sku gpu_1x_h100       # → s-xxxx
super run . --script run.sh --on s-xxxx
super run . --script run.sh --on s-xxxx -- 'A bookstore with a sign reading "OPEN BOOKS"'

The default prompt depicts a cozy bookshop with a sign reading "OPEN BOOKS". The first run installs dependencies and downloads tens of GB of weights; subsequent runs reuse the mounted cache. A cold run on staging (1x H100, 2026-10-08) took 962 s end to end and peaked at about 59 GiB of GPU memory.

Get the results out

super download s-xxxx /workspace/out/qwen_image.png ./qwen_image.png
super stop s-xxxx

Notes

SUPERCOMPUTER_STRESS_DURATION_S retains the stress harness's soak behavior: generate extra images under /workspace/out/gen/ until the requested window ends, then write the final pass artifact. Unset/0 generates one image.
OUT can override the final artifact path.
The base image supplies PyTorch 2.6.0; run.sh prepends /opt/conda/bin so Python and pip use that CUDA environment.
supercomputer.json pins HTTPS certificate variables to the system CA bundle for runtime pip and model downloads.

Source files

`run.sh`

#!/usr/bin/env bash
# Run inside a supercomputer session (see README.md).
set -euo pipefail

# The base image's torch+CUDA environment lives in conda, outside super run's PATH.
export PATH="/opt/conda/bin:$PATH"
OUT="${OUT:-/workspace/out/qwen_image.png}"
# Preserve the stress harness's window, including dependency install and model load.
STRESS_END=$(( $(date +%s) + ${SUPERCOMPUTER_STRESS_DURATION_S:-0} ))
PROMPT="${1:-A cozy bookshop at dusk, with a clearly legible sign reading \"OPEN BOOKS\", warm lighting and detailed shelves.}"

python -m pip install \
    diffusers==0.35.1 transformers==4.55.4 accelerate==1.10.1 \
    safetensors==0.6.2 sentencepiece==0.2.1 Pillow==11.3.0

python - "$PROMPT" "$OUT" "$STRESS_END" <<'PYTHON'
import sys
import time
from pathlib import Path

import torch
from diffusers import QwenImagePipeline

prompt, output, stress_end = sys.argv[1:]
output = Path(output)
stress_end = int(stress_end)
pipe = QwenImagePipeline.from_pretrained(
    "Qwen/Qwen-Image", torch_dtype=torch.bfloat16
).to("cuda")
output.parent.mkdir(parents=True, exist_ok=True)

n = 0
while True:
    image = pipe(prompt=prompt, negative_prompt=" ", true_cfg_scale=4.0).images[0]
    if time.time() >= stress_end:
        image.save(output)
        break
    gen_dir = output.parent / "gen"
    gen_dir.mkdir(exist_ok=True)
    image.save(gen_dir / f"qwen_image_{n:04d}.png")
    n += 1
if n:
    print(f"soak: generated {n} extra images")
print(f"wrote {output}")
print(f"peak_vram_mib={torch.cuda.max_memory_reserved() // 2**20}")
PYTHON

`supercomputer.json`

{
    "sku": "gpu_1x_h100",
    "image": {
        "base": "pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel",
        "apt": [
            "git",
            "ca-certificates",
            "build-essential",
            "python3-dev"
        ],
        "env": {
            "PYTHONUNBUFFERED": "1",
            "HF_HOME": "/workspace/.cache/huggingface",
            "HF_HUB_CACHE": "/workspace/.cache/huggingface/hub",
            "SSL_CERT_FILE": "/etc/ssl/certs/ca-certificates.crt",
            "PIP_CERT": "/etc/ssl/certs/ca-certificates.crt",
            "REQUESTS_CA_BUNDLE": "/etc/ssl/certs/ca-certificates.crt"
        }
    },
    "volumes": [
        {
            "name": "qwen-image-weights",
            "mount": "/workspace/.cache/huggingface"
        }
    ]
}