Petit MegaMoE: Fast Multi-GPU MoE on AMD GPUs
Petit MegaMoE accelerates multi-GPU MoE inference on AMD Instinct MI350X and MI355X GPUs with direct xGMI dispatch, fine-grained synchronization, and persistent kernels.
VeriLocc: Using LLMs to Allocate GPU Registers
VeriLocc uses an LLM to assign GPU registers across NVIDIA and AMD hardware, then verifies each result.
Optimizing FP4 Mixed-Precision Inference on AMD GPUs
Learn how we developed Petit, a collection of optimized FP16/BF16 x FP4 mixed-precision GPU kernels for AMD GPUs, achieving 1.74x faster inference and up to 3.7x performance improvements.