Current Setup

Model Configuration

• Base Model: KernelLLM (Llama-3.1-8B-Instruct)
• Fine-tuned: SFT on GPUmode data
• Task: Converting PyTorch code to Triton code
• Training Method: RL post-training
• Hardware: Single H100 node

Experiment Type

• Mode: Single-turn code conversion
• Evaluation Metrics:
  • Compilation success
  • Correctness
  • Performance

Prompt Data

Mix up KernelBench Level 1 - Level 4: 270 cases

RL Framework Integration

• RL Framework: Slime
• Evaluation Sandbox: KernelBench
• Integration Status: ✅ Successfully integrated and working

Experiment Results

Key Findings

• Reward Improvement: ~5% increase observed
• Reward Behavior: Some fluctuation observed, but overall upward trend