Recipes
Recipes are complete, reproducible training runs: the exact source revisions, configs, node layout and launch commands used to train a draft model that ships in SpecBundle. Where the guides explain each option, a recipe shows one proven combination end to end.
- deepseek-v4-flash-dspark-disaggregated.htmlRead the recipe →
- Kimi K3 DSpark DisaggregatedMove the four-node colocated Kimi K3 continual run to one TP8 capture node and one four-rank trainer node without changing the draft architecture or training schedule.Target moonshotai/Kimi-K3Read the recipe →
- Qwen3.8-27B DFlash2 DisaggregatedTrain the five-layer GQA DFlash2 draft for Qwen3.8-27B with patched SGLang capture servers and Mooncake, on one 8-GPU node or across two, using the server/trainer split and flow control measured on B300 and H200.Target Qwen/Qwen3.8-27BRead the recipe →
Contribute a recipe
Trained a draft model that others should be able to reproduce? Add a Markdown file to docs/recipes/ with a title, description, target, method and topology in its frontmatter and open a pull request. The card above and the sidebar entry are generated from the file, so nothing else needs to change. See the docs guide for the template.