Reference
Check supported models, platform limits and troubleshooting steps before configuring or recovering a workflow.
Use these pages to check compatibility, understand a limit or find the next step when something fails. For a walkthrough from data to a model, start with the quickstart.
Look up a requirement
| Reference | What you will find |
|---|---|
| Supported models | Model families, text and vision support, SFT and DPO restrictions, GPU requirements and Hugging Face access |
| Limits | Compute, plans, imports, images, training, serving, API requests and organization limits |
| Troubleshooting | Connection errors, imports, dataset readiness, GPU capacity, training recovery, serving and API failures |
Before a paid launch
- Check both the model's supported training objective and adapter method. Vision SFT uses LoRA; supported vision DPO can use LoRA or QLoRA. Image workflows need the vision capability included with Pro.
- Confirm GPU memory, provider quota, image compatibility and available capacity. A connection check is not a capacity reservation. Use the cloud provider guides for setup requirements.
- Use a training trial to check a small sample. Trials use paid compute even when they do not count toward the monthly training allowance.
- Review your organization's plan and compute limits in the console. Plans and billing explains what each allowance counts and when it resets.
When work stops
Read the error on the run, deployment or Activity entry, then follow the matching troubleshooting section. A finished or failed run can still have cleanup pending. Confirm release in the provider console, and recover any needed output before deleting retained resources.
For an API integration, use the chat completions contract and keys and usage. For an agent connection, see MCP access and tokens.
If the documented steps do not resolve the issue, contact support with the record ID and error message. Keep API keys, cloud credentials and private dataset content out of the message.
Need a hand? Visit troubleshooting