ai-for-science-ai4s-main

AI for Science Main Entry Skill

This Skill is only responsible for route determination and sub-skill routing, and does not expand on specific migration or tuning details. When the user only provides a broad AI for Science requirement, a model name, a TensorFlow/Keras project, or only says "help me migrate to Ascend/collect profiling", start here to determine which sub-skill to enter.

Three Main Routes

Direction	Entry Condition	Recommended Sub-Skill
Profiling Collection	The code can already train or infer, and only needs to collect traces, analyze hotspot operators, call stacks, memory, or bottlenecks	ai4s-profiling
Model Migration	The model name is known, or an AI4S model needs to be migrated from GPU/CUDA to Ascend NPU	ai4s-basic or model-specific skill
TF Framework	The original project is TensorFlow/Keras, and a decision is needed on whether to keep TensorFlow or rewrite to PyTorch	ascend-tf-community / tf-to-pytorch

Model Routing Table

Model or Task	Skill Entered	Description
Boltz2	boltz2	Protein structure prediction and end-to-end inference reproduction
BoltzGen	boltzgen	Generative protein design and inverse folding
DeepFRI, keep TensorFlow	deepfri-tf-npu	Keep the TF runtime and the original implementation
DeepFRI, migrate to PyTorch	deepfri	Perform TF to PyTorch rewriting and weight conversion
DiffSBDD	diffsbdd	Structured drug design and diffusion inference
GENERator	generator	DNA sequence generation model migration
OligoFormer	oligoformer	siRNA efficacy prediction and RNA-FM dependency adaptation
ProteinBERT	proteinbert	Protein language model weight conversion, embedding, and fine-tuning
New model not yet accumulated	ai4s-basic	First go through the general migration process, then accumulate a model-specific skill

Decision Rules

1. If the user can already run it and only wants to collect profiling or locate performance issues, directly enter ai4s-profiling。
2. If the user explicitly wants to keep the original TensorFlow/Keras implementation, enter ascend-tf-community。
3. If the user explicitly wants to migrate to PyTorch, or subsequently wants to integrate with the torch_npu ecosystem and unify the training and inference process, enter tf-to-pytorch。
4. If the user has already provided a specific model name, preferentially enter the corresponding model skill; only when there is no model-specific skill, enter ai4s-basic。
5. After this Skill completes routing, continue executing environment, adaptation, validation, and reference material reading in the corresponding sub-skill, and do not repeat the expansion here.