VQVLA: Algorithm-Hardware Co-Design for Efficient VLA Inference
Researchers propose VQVLA, an algorithm-hardware co-design framework to accelerate Vision-Language-Action (VLA) model inference on GPUs. VLA models, used in embodied AI, suffer from high latency. VQVLA introduces MotionVQ, a motion-aware vector quantization scheme that adjusts precision based on robot execution state, reducing memory access while preserving task success. It also proposes a merged-centroid vectorized GEMM paradigm that eliminates redundant multiplications via spatial aggregation and temporal centroid reuse. The framework targets redundancy in memory and computation left unexploited by existing accelerators like Dadu-Corki.
Key facts
- VQVLA is an algorithm-hardware co-design framework for VLA inference acceleration.
- MotionVQ is a motion-aware vector quantization scheme.
- MotionVQ adjusts quantization precision based on robot execution state.
- The merged-centroid vectorized GEMM paradigm operates on codebook-index representation.
- It eliminates redundant multiplications through spatial aggregation and temporal centroid reuse.
- Existing accelerators like Dadu-Corki treat VLA models as full-precision workloads.
- VLA models have high inference latency on GPUs.
- The framework reduces memory access while preserving task success rate.
Entities
—