Skip to main content
Dedicated container inference is available through the Together sales team. Contact sales to get access.
Using a coding agent? Install the together-ai skill to let your agent write correct dedicated container code automatically. See agent skills for details.
Dedicated container inference lets you run your own Dockerized inference workloads on Together’s managed GPU infrastructure. You bring the container, and Together handles compute provisioning, autoscaling, networking, and observability.
  • Fast setup: Wrap and deploy your model in 20 minutes.
  • Priority queueing: The managed job queue runs higher-priority jobs first, so you can serve your most valuable traffic ahead of the rest.
  • Scale with demand: Together adds replicas as traffic grows, up to thousands of GPUs, and can scale down to zero when idle.
You deploy with the Jig CLI in one of two ways: Build a custom worker: Use the Sprocket SDK in your container to connect your inference code to Together’s managed job queue, then build, push, and deploy the Docker image with Jig. Deploy a prebuilt image: To serve an off-the-shelf image like stock vLLM, deploy it as-is with Jig. You don’t write Sprocket code or build an image.
Dedicated container inference architecture

​
Quickstart

To deploy a custom worker with the Jig CLI, run three commands:

Deploy your first container

Deploy a sample inference worker, submit a job, and get the result.

​
Concepts

Architecture

Learn the deployment lifecycle and autoscaling, and troubleshoot common issues.

Jig CLI

Build and deploy containers, and manage secrets and volumes.

Sprocket SDK

Write inference workers with setup() and predict().

Queue API

Submit asynchronous jobs with priorities and track their progress.

​
Tutorials

Image generation

Generate images with FLUX.2 on a single GPU.

Video generation

Generate video with Wan 2.1 across multiple GPUs using torchrun.

OpenAI-compatible endpoint

Serve FLUX.2 behind the OpenAI Images API.

Prebuilt vLLM image

Serve Qwen2.5-VL from the stock vLLM image, with no Sprocket code and no Docker build.

​
Reference

Jig CLI

Look up CLI commands and pyproject.toml configuration.

Sprocket SDK

Look up base classes, file handling, and errors.

REST API

Look up the deployments, secrets, storage, and queue endpoints.