Dstack Prototyping

dstack Prototyping

Use /dstack for CLI commands, YAML fields, apply/attach behavior, service URLs, and other dstack syntax. This skill explains how to use dstack runs while the model-serving configuration is still unknown.

Goal

Find a working dstack service configuration for the requested model.

Before submitting a service, use a task on real hardware to test the serving image, install/runtime assumptions, model download, cache path, command, port, launch flags, resources, env vars, backend/fleet choice, and local model request. Then submit the same configuration as a service and verify the model through the dstack service URL.

Choose Where To Run

Pick the offer whose hardware best fits the goal at hand. Only when several offers fit comparably, choose a VM-based backend, an SSH fleet, or a Kubernetes fleet: they support idle instances and/or instance volumes, so later runs reuse the provisioned/idle instance or instance volumes for caching model weights (and possibly other writes), while container-based backends start clean on every run.

Fetch https://dstack.ai/docs/concepts/backends.md and classify backends from the fetched document, not from memory.

If the intention is to use PD disaggregation, the fleet must use placement: cluster. Since PD disaggregation implies running a router, unlike workers that must run on GPUs, the router normally should run on a CPU instance. Use dstack fleet to see existing fleets and dstack fleet get <fleet name> --json to inspect a specific fleet.

Check Serving Sources

Check serving-framework sources early enough to choose the image, command, launch flags, resources, cache paths, request format, and expected model behavior.

For vLLM and SGLang, use these as credible sources:

• vLLM recipes and model index: https://recipes.vllm.ai/ and https://recipes.vllm.ai/models.json
• SGLang docs: https://docs.sglang.io/ (fetch /llms.txt for the page index)
• SGLang model recipes: https://docs.sglang.io/cookbook/autoregressive/intro
• Release notes: https://github.com/vllm-project/vllm/releases and https://github.com/sgl-project/sglang/releases
• Performance-loop methodology (profiling, benchmark contracts): https://www.lmsys.org/blog/2026-07-02-agent-assisted-sglang-development

Use A Task Before Service

Before submitting a service, start a long-lived task:

commands:
  - sleep infinity

or an equivalent idle command.

Submit the task detached, attach or SSH into it when available, and run commands inside the live environment. Test the image, installs, model download and cache path, serving command, port, launch flags, local model request, and expected model behavior.

When starting a long-running command in the background from a non-interactive SSH command, use nohup, redirect stdin from /dev/null, and redirect stdout/stderr to a log file so the SSH command returns while the process keeps running. For example (the command can be any long-running command):

nohup vllm serve ... </dev/null > /tmp/vllm.log 2>&1 &

If the image, hardware choice, or major install path changes, submit another task so the changed setup is tested before service verification.

Do not move to a service after checking only GPU visibility, imports, logs, or a health endpoint. Start the server inside the task and send a request that uses the requested model. For a chat or reasoning model, check the response behavior the endpoint is expected to support, such as reasoning output when that model is supposed to expose it.

Follow /dstack structured status guidance when polling task or service status. After requesting a task or service stop before another submission, wait until that run reaches a terminal status. This allows dstack to reuse its instance or instance volumes when available.

Verify As A Service

Submit the service after the task has verified the configuration: image, command, port, resources, env vars, cache mounts if used, backend/fleet choice, and model request.

Use the service as a duplicate check of the same configuration under dstack service runtime. The model request that worked locally in the task must also work through the dstack service URL.

If service verification fails because the image, install, model download, command, resources, cache, or model behavior needs to change, go back to a task. If the tested serving setup is still right and only the dstack service configuration is wrong, fix the configuration and submit the service again.

PD disaggregation

If the intention is to use PD disaggregation:

• Use node groups for the task and replica groups for the service: tasks' node groups are the equivalent of services' replica groups.
• In both cases, you run a router and prefill/decode workers separately, and you need to use a fleet with an interconnect (placement: cluster).
• With tasks, still use sleep infinity even when using groups (set it in each group's commands; top-level commands is not allowed with groups), and run the actual commands on each node interactively over SSH.
• When testing inference, call the router endpoint, not the workers directly (unless you want to test if they are alive).
• Look for "Node groups" and "PD disaggregation" in https://dstack.ai/docs/concepts/tasks.md and "Replica groups" and "PD disaggregation" in https://dstack.ai/docs/concepts/services.md.