The authors built KernelOPT to treat compiled models as structured programs rather than collections of isolated kernels.
Dispatch-aware agents safely optimize compiler-generated GPU kernels within full models
KernelOPT profiles compiled PyTorch programs, rewrites selected Triton sub-kernels, and verifies each candidate after reintegration into the model.
Industry
Aheli Poddar · Sanskar Prasad · Arindam Samanta · Subha Chakraborty · Vishal Goyal · Rohit Singh Rathaur
Red Hat
Research Digest··2 min read
Thread:Multi-Agent Coordination
Poddar et al.
Why this paper
From Red Hat · Released code · Part of Multi-Agent Coordination, now 24 papers
In one line
KernelOPT uses LLM agents to optimize GPU kernels in compiled PyTorch models, achieving geometric mean speedups up to 1.40x over torch.compile.
What it released
Code
What we could check
- ✓Code link in the paper (github.com)
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§