GALA: A High Performance Graph Neural Network Acceleration LAnguage and Compiler
summary
GALA is a domain-specific language and compiler that composes intra-operator scheduling commands with automatic inter-operator rewrites via two novel IRs, a compute IR and data IR, to capture global program and data relations. It implements training-aware rewrites (e.g., loop-invariant code motion moving non-learned work out of the training loop), tiling and sampling transforms, and generates GPU kernels evaluated on H100/A100. Compared to existing GNN and sparse-tensor systems (e.g., DGL, TACO), GALA reports 2.55× inference / 2.52× training speedups and exposes an explicit GALA DSL for tradeoff exploration.
AI-written summary · about
chapters (11)
- 0:00Gala GNN DSL intro
- 1:29Graph neural networks & sparse ops
- 2:55Intra vs inter operator optimizations
- 3:36Combined optimizations synergy
- 5:03Gala architecture & components
- 6:08Compute & data IR design
- 7:35IR transformations separation
- 8:38Training-aware code motion
- 10:02Evaluation setup & GNN models
- 11:32Gala performance summary
- 11:52Q&A
related talks
19mOptimizations and abstractions for sparse machine learning
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
25mOpen Source Model Performance Optimization With SGLang
23mBlazing Fast GenAI Inference With Torch.compile
20mLegion: Automatically Pushing the Envelope of Multi-GPU System for Billion-Scale...