SOURCE-LINKED INTELLIGENCE

DaCe GPT

CORDIS · observation · Publication date unknown

d specialization. DaCe was shown to work at scales ranging from small computational kernels to applications with up to a million lines of code when guided by performance experts. Recent advances in large language models (LLMs) enable coding agents to excel at tasks like API usage and documentation-based pattern matching (“how do I use OpenSSL to do X in Python?”). However, they consistently fail at tuning performance-critical imperative code, as this requires understanding dataflow and how to correctly restructure it. Finally, the resulting changes span many lines of code, requiring more context than available to LLMs. In this project, we aim to democratize performance tuning by integrating LLMs with DaCe — not as direct code optimizers, but as reasoning interfaces that use a DaCe backend. We propose a novel approach in which users provide plain imperative code (Python or Fortran). Our system, DGPT, powered by a reasoning model, compiles the code into DaCe, analyzes it, applies performance-improving transformations, and verifies the resulting code correctness and performance. Thu

Read original source ↗ Open in workspace

recordType

award

status

SIGNED

region

EU

value

150000

unit

EUR