Multi-level Machine Learning-Guided Autotuning for Efficient Code Generation on a Deep(…)
summary
ML2Tuner implements a validity-prediction model to skip configurations that would crash and a second performance model that uses hidden compiler features collected by a hidden-feature extractor to refine predictions. Implemented for an extended VTA accelerator with XGBoost and evaluated on ResNet-18 and similar layers, it finds equivalent best code using 12.3% of samples compared to a TVM-like tuner, cuts invalid profiling attempts by ~60.8%, and lowers prediction loss by about 10% per layer.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:54Autotuning and VTA accelerator
- 5:24Simple scheduling on VTA
- 6:09ML-guided autotuning process
- 8:17Invalid configurations with model V
- 9:00Feature extraction and model A
- 12:11Experimental results ResNet-18 EfficientNet-V3
- 14:36Summary
- 15:20Q&A
related talks
10mAuto-tuning vllm
19mCausalTuner: Feature-Aware Causal Guidance for Compiler Auto-tuning
16mmTuner: Accelerating Parameter-Efficient Fine-Tuning on Multi-GPU Servers with...
01h 22mAutomated Machine Learning & Tuning with FLAML
26mVLLM-Helion: SOTA LLM Performance by Advanced Autotuning and Fine-grained Dispatching
32mTaming the Machine: Basics of ML Models Training and Inference Optimization
34mPerformance tuning Twitter services with Graal and Machine Learning
56mPerformance tuning Twitter services with Graal and Machine Learning