Scaffold Surogate DSL blocks and HuggingFace weight mappings for new model architectures.
npx skills add https://github.com/invergent-ai/surogate --skill new-model
Please help me install this Agent Skill. Skill: new-model Source: https://github.com/invergent-ai/surogate/tree/main/.claude/skills/new-model Command: npx skills add https://github.com/invergent-ai/surogate --skill new-model
This capability enables teams to add support for new HuggingFace models by scaffolding the required DSL components: a transformer block, a model wrapper, and the corresponding HF weight mappings.
Provide the model name or HF repo and the DSL will scaffold the block, model, and HF mappings for a new architecture.
High-intent search queries and answers about installing and using this skill.
To add a new HuggingFace model architecture, you scaffold the required transformer block, model wrapper, and HF weight mappings. This ensures proper block structure and configuration alignment with HF config fields for production-ready training.
Mapping HuggingFace config fields to DSL parameters involves generating corresponding block definitions and weight mappings for the target architecture. This automates configuration alignment and integrates the new model with existing modules.
Yes, you can integrate custom HuggingFace backbones like Gemma2 or DeepSeek with minimal boilerplate. Providing the model name or HF repo scaffolds the block, model, and HF mappings for the new architecture.
Adding new model support does not require manual block definition and weight mapping. The process automates the creation of the DSL block, model wrapper, and HF weight mappings directly from the provided HuggingFace repo.
When integrating a new model architecture into a DSL, the scaffolded components include a transformer block, a model wrapper, and the corresponding HuggingFace weight mappings to ensure proper structural and configuration alignment.
gar-aiDetect and resolve tensor naming mismatches when porting HuggingFace models to Candle.
huidgeTrack ML experiments and manage model training metadata with Weights & Biases.
AlexiosBluffMaraTokenize text with Rust-based HuggingFace tokenizers for NLP preprocessing.
CyapstayeRemove refusal behaviors from open-weight LLMs via weight projection.
