FAQ
Getting started
Section titled “Getting started”Do I need to launch an instance to call a model? No. Token Factory is free and always on. Get an API key from Token Factory and start calling. It’s on the China site only. See Token Factory.
Can I run more than one instance? No. Each account can have one instance at a time. Destroy it to launch another.
How many GPUs can I get?
It depends on the sizes assigned to your account. Size in the launch form lists them, and rcc catalog list shows the same list.
How do I get credits on the China site? Exchange AMD AI Developer Program points for a coupon on 兑取算力, then redeem it under Credits & Usage → Redeem credits. 1 point buys 1 hour, up to 20 points a day. See AMD AI Developer Program.
Who do I contact? Email member-service@mail.developer.amd.com.cn. Troubleshooting lists what to include.
What spends credits? Running instances: 1 credit per GPU per hour. Token Factory doesn’t use credits. A model service you run yourself does, because it runs on an instance.
Am I charged while an instance is idle? Yes. An instance spends credits for as long as it exists, whether or not you’re using it. Destroy it when you’re done.
Will my instance be stopped for me? Only when your credits run out, or when the Auto-reclaim time shown on the launch form has passed. Don’t rely on either to save credits.
Are Token Factory usage and credits the same thing?
No. Token Factory has its own daily usage cap, counted in points (pts). See it on Credits & Usage, or with GET /v1/usage.
Storage
Section titled “Storage”Do my files survive?
Only files under /workspace, and only if you launched with Persistent /workspace. Otherwise everything is deleted with the instance. See Keep your data.
Can I add persistent storage to a running instance? No. Storage is chosen at launch. Destroy the instance and launch again with Persistent /workspace.
How much persistent storage do I get? 25 GB. Some accounts can choose 100 GB. The two are separate volumes, so files saved on one don’t show up on the other.
Models and compatibility
Section titled “Models and compatibility”Does my CUDA code work?
Usually, at the PyTorch level. ROCm builds of PyTorch support the torch.cuda.* API, so high-level PyTorch code usually runs unchanged. Hand-written CUDA kernels need porting with HIP, and libraries that ship only CUDA binaries won’t work.
Which models are in Token Factory?
The list changes. See the model catalog, or call GET /v1/models instead of hard-coding names.
Can I serve any model on my own instance? Any model vLLM supports that fits in the GPU memory of the size you chose. Weights are downloaded on first start, so large models take a few minutes before they answer. See Serve a model.
Does streaming work?
Yes. Set stream: true on chat completions.
Are embeddings available?
In Token Factory, only if the model catalog lists an embedding model. A model service on your instance serves whatever vLLM supports, which usually includes /v1/embeddings.
Access
Section titled “Access”Can I share my endpoint with a teammate? Not through the platform. To expose a service publicly, run it in the instance and use rc-tunnel, with your own authentication in front of it.
Can I use an API key to open JupyterLab or launch instances? No. API keys work for Token Factory only. JupyterLab and the console need a browser session, and scripts manage instances with Radeon Cloud CLI.
Why can’t I launch an instance? Usually one of these: no free node in the resource pool, not enough credits, an instance already running, or an account still under review. See Troubleshooting.
Operations
Section titled “Operations”Can I script launches? Yes, with Radeon Cloud CLI. See Scripting and automation. Always delete the instance at the end of the script, even when it fails, so a crash doesn’t leave it running.
Does the endpoint URL stay the same when I launch again? Don’t rely on it. The base URL and key can change, so copy them from the instance card after each launch and keep them in configuration, not in code.