Skip to main content
Last Updated: October 24, 2025
Deploy from the SaladCloud Portal.

​
Overview

This recipe provisions a GPU-ready Kelpie worker with the entire RISC Zero toolchain already installed:
  • Stable Rust + cargo
  • rzup toolchains (Rust, cargo-risczero, r0vm)
  • CUDA 12.x for accelerated proving
  • A clone of the latest risc0 repo under /opt/risc0
  • Sample helpers in /opt/risc0/examples/
This container uses the Kelpie worker. You enqueue jobs (via the Kelpie API), the worker downloads your project from S3-compatible storage, runs your cargo command (e.g., cargo run -F cuda --release), and uploads /opt/results/ back to the same bucket. Use this recipe when you need to prove custom guest programs or benchmark RISC Zero locally on Salad GPUs without wiring up your own infrastructure. You can also use it as a starting point to build more complex proof pipelines of your own.

​
Prerequisites

  1. Salad organization + project where you can deploy recipes and create API keys.
  2. S3-compatible bucket (Cloudflare R2, AWS S3, MinIO, etc.) for inputs and outputs. You will provide the Access Key, Secret Key, Region, and (for R2) Endpoint URL when deploying the recipe.
  3. (Optional) Queued autoscaling invite the Kelpie service account to your Salad org if you plan to use Kelpie’s queue-aware scaler.

​
Deploy the Recipe

  1. Open the SaladCloud Portal, choose organization and project, then click Deploy a Container Group and search for RISC Zero ZKP.
  2. Fill in the form:
    • Container Group Name – unique name of the container group.
    • Storage Access Key / Secret Key – credentials for your S3-compatible bucket.
    • Storage Region – e.g., auto for Cloudflare R2 or us-east-1 for AWS.
    • Storage Endpoint URL – leave blank for AWS; for R2 provide the custom endpoint (https://<account>.r2.cloudflarestorage.com).
  3. Deploy. The container group comes online with 16 vCPUs, 8 GB RAM, and a 24 GB GPU. You can modify those settings by clicking Modify Configuration Each replica runs one Kelpie worker.

​
Kelpie Job Flow

Kelpie jobs are simple JSON documents with:
  • command + arguments – what to run (e.g., bash -lc "RUSTFLAGS=\"-C target-cpu=native\" cargo run -F cuda --release |& tee /opt/results/hello_world.log").
  • sync.before – download your project from S3.
  • sync.after – upload /opt/results/ once the job completes.
The repo ships two helper scripts (see salad-recipes/recipes/risc0/):
  • submit_kelpie_job.py – example job submitter
  • check_kelpie_job.py – poll job status
These are meant as starting points. Edit them to point at your prefix/bucket.

​
Sample Submission Script

The bundled script already knows how to run the hello-world demo on GPU. Update build_sync to download your project prefix under ""before”, then run: 
export SALAD_API_KEY=...
export SALAD_ORGANIZATION=your-org
export SALAD_PROJECT=your-project
export CONTAINER_GROUP_ID=<container-group-id>
export BUCKET_NAME=<your-bucket>
python recipes/risc0/submit_kelpie_job.py
What it does:
  1. Creates a job ID (job-xxxxxx).
  2. Builds a Kelpie job that:
    • sync.after uploads /opt/results/ to s3://<bucket>/hello-world/<job-id>/
    • Runs RUSTFLAGS="-C target-cpu=native" cargo run -F cuda --release in /opt/risc0/examples/hello-world while streaming output to /opt/results/hello_world.log
  3. Posts the job to Kelpie and (optionally) polls until it finishes.
  4. Prints Kelpie job id.
Replace the cargo run ... command with your own proof pipeline (e.g., running cargo run --release --bin my-host or python scripts/prove.py). Make sure any artifacts you care about land under /opt/results/ so Kelpie uploads them.

​
Sync Patterns

Typical sync for custom projects: 
{
  "before": [
    {
      "bucket": "risc0",
      "prefix": "projects/my-proof/",
      "local_path": "/opt/projects/",
      "direction": "download"
    }
  ],
  "during": [],
  "after": [
    {
      "bucket": "risc0",
      "prefix": "results/${JOB_ID}/",
      "local_path": "/opt/results/",
      "direction": "upload"
    }
  ]
}
Copy that pattern into your submission script. You can reference environment variables (e.g., ${JOB_ID}) inside prefixes to keep outputs separated by job.

​
Monitor Jobs

Use the helper or the raw API to check job status: 
JOB_ID=job_id_returned_in_step_above \
python recipes/risc0/check_kelpie_job.py
This prints the Kelpie job JSON, including status, and all the other metadata.

​
Customizing the Worker

  • Different Provers: Copy your host + guest crates into the bucket and point Kelpie at them via sync.before. The base image already has CUDA + Rust.
  • External Dependencies: you can download them inside your bash -lc script (e.g., pip install -r requirements.txt) or fork the recipe repo and modify it.
  • Autoscaling: Kelpie supports queue-aware autoscaling, including scale-to-zero when the queue is empty. Invite the Kelpie service account to your organization, then create a scaling rule with the Create Scaling Rule endpoint.

​
Resources