The authors frame GPU kernel optimization as generating complete optimization scripts rather than ranking candidates or emitting kernel code.
8B model learns to construct and refine GPU optimization scripts, beating larger LLMs
Little Dragon, an 8-billion-parameter model, achieves above 91% per-attempt correctness after supervised fine-tuning and reinforcement learning on 108,594 compiler search examples.
Big Tech
Annabelle Sujun Tang · Wei Tang · Songyao Jiang · Tianhui Li · Shengzhe Xu · Jiin Woo · +4 more
University of California, San Diego · Amazon Annapurna Labs
Research Digest··2 min read
Thread:Agent Harness Optimization
The authors introduce a training method that teaches an 8B language model to generate complete GPU optimization scripts and improve existing ones.
Why this paper
From Amazon Annapurna Labs and University of California, San Diego · Part of Agent Harness Optimization, now 143 papers
In one line
A training method teaches an 8B language model to generate correct GPU optimization scripts, outperforming larger frontier LLMs.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§