LLM Fine-Tuning for Cloud Workflow Resource Prediction
Researchers have introduced LASER, a framework that fine-tunes large language models on serialized workflow job configurations to predict resource consumption and runtime for cloud computing tasks. The approach addresses the challenge of semi-structured job configurations—including shell commands, parameters, dependency graphs, and metadata—which traditional machine learning methods struggle to process without extensive feature engineering. LASER employs scientific notation output encoding for multi-target regression across orders of magnitude and constrained decoding with prefix filling, reducing inference latency by over 30% while ensuring output validity. Full-attention fine-tuning further improves accuracy. The work is detailed in a paper on arXiv (ID 2512.19701).
Key facts
- LASER fine-tunes LLMs on serialized workflow job configurations.
- It predicts resource consumption and runtime for cloud workflow jobs.
- Semi-structured job configurations include shell commands, parameters, dependency graphs, and metadata.
- Traditional ML requires brittle feature engineering to flatten this information.
- Scientific notation output encoding handles targets spanning multiple orders of magnitude.
- Constrained decoding with prefix filling reduces inference latency by over 30%.
- Full-attention fine-tuning improves accuracy.
- The paper is available on arXiv with ID 2512.19701.
Entities
Institutions
- arXiv