Optimize a VLA Model for Embodied Intelligence: OpenVINO Toolkit (Intel) Project #21

VLA models like unifolm-vla combine a vision-language backbone with a flow matching action head. On Intel iGPU, these models face two major deployment hurdles. First, structural blockers in the model code prevent successful OpenVINO conversion. Second, the iterative denoising loop requires kernel-level optimization to reach the frequencies needed for real-time robotic control. I will resolve four specific export blockers related to Python loops, random number generation, and precision handling. After achieving a successful export, I will implement kernel fusions for AdaLayerNorm and optimize Scaled Dot Product Attention to reduce latency on Arrow Lake and Panther Lake hardware. Project Deliverables: - A working OpenVINO export pipeline for unifolm-vla with all conversion blockers resolved. - Kernel-level optimizations including AdaLayerNorm fusion and SDPA pattern fixes. - A reference project featuring export scripts, quantization configurations, and latency benchmarks. - A detailed profiling report identifying bottlenecks and primitive selection via VTune.