Paper-to-Skill

Compile ML research papers into optimized Triton kernels for agentskills.io.

Features

• Constraint Extraction: Parses PDFs/markdown to extract math formulations and tensor requirements.
• AST Safety Sandbox: Validates generated code via AST inspection to block unsafe operations (os, subprocess, etc.).
• V2 Multi-File Synthesis: Separates kernels from PyTorch modules (triton_kernel.py vs nn_module.py).
• Hardware-Aware Routing: Targets specific GPU architectures (sm_80, sm_86, sm_90).
• Hermes Integration: Generates compliant SKILL.md with YAML frontmatter for automatic indexing.

Security

By default, code execution is restricted. To enable local bare-metal execution (e.g., on cloud GPU instances):

export ALLOW_DANGER_RUN_BARE_METAL=true

Warning

This flag bypasses host isolation. Always verify generated code before execution.

Setup

pip install -e .

Environment

Create a .env in the project root:

OPENAI_API_BASE="http://localhost:8000/v1"
OPENAI_API_KEY="your-api-key"

Usage

1. Compile Paper to Skill

python -m paper_to_skill compile --pdf paper.txt --target sm_86 --out ./skills/

2. Install to Hermes Registry

python -m paper_to_skill install --dir ./skills/sageattention-2/

Verification

Once installed, invoke the skill directly in Hermes:

hermes chat -q "/sageattention-2 'Run attention forward pass'"

Comparison

Feature	Baseline LLM	Paper-to-Skill
Math Accuracy	Hallucinated/Placeholders	Extracted from Source
Quantization	Generic Fallback	INT8/FP8 per paper
Performance	Suboptimal/Non-runnable	Optimized Triton Kernel