[Skill] model-onboard: New Model Architecture Onboarding

[sunway513]

Skill: model-onboard

Priority: P1 — New model onboarding frequency is increasing rapidly

Motivation

Adding support for a new model architecture in ATOM requires touching multiple files with specific patterns: model implementation, model runner registration, weight loading, quantization config, and testing. The process is well-defined but has many steps where small mistakes cause hard-to-debug failures. Encoding this workflow as a skill reduces onboarding time and error rate.

What This Skill Should Do

Given a HuggingFace model name or architecture description:

1. Analyze the model architecture

   • Read HuggingFace config.json and modeling files
   • Identify: attention type (MHA/MQA/GQA/MLA), FFN type (dense/MoE/SwiGLU), normalization, position encoding
   • Map to closest existing ATOM model as a template

2. Generate model implementation skeleton

   • Create atom/models/<model_name>.py based on the closest template
   • Implement attention, FFN, and layer classes
   • Handle quantization hooks (FP8/FP4 weight loading)
   • Include proper weight name mapping for load_weights()

3. Register the model

   • Add entry to model_runner.py model registry
   • Add model to supported models documentation
   • Create basic test configuration

4. Validate the implementation

   • Load model weights successfully
   • Run a single forward pass
   • Compare output against HuggingFace transformers reference
   • Check output quality (cosine similarity > 0.999 for FP16, > 0.99 for FP8)

Acceptance Criteria

• Generate working model skeleton from HuggingFace model card
• Automatic model runner registration
• Weight loading validation
• Forward pass correctness check
• Support for common patterns: MoE, GQA, RoPE, sliding window, attention sinks