ConceptMoRI量化 / MoRI-IO

MoRI

MoRI is a communication optimization technique for distributed machine learning training that reduces inter-node communication bandwidth requirements by applying quantization compression to data transmitted during All-to-All collective communications. Its core approach reduces the volume of communicated data while maintaining acceptable model training precision, thereby alleviating communication bottlenecks in large-scale parallel training and improving overall training efficiency.

Timeline (last 90 days)

Oct 11

AMD's solution combines the MoRI technology stack with the vLLM inference framework for MoE model serving.

Unverified50%

All Facts (2)

Unverified

MoRI（Mixture of Routed Inference）量化通信策略在 dispatch 阶段使用 FP4 精度，combine 阶段使用 FP8 精度，实现了 2.56 倍的带宽压缩

90%Unverified

AMD's solution combines the MoRI technology stack with the vLLM inference framework for MoE model serving.

50%

Source Articles

AMD MoRI + vLLM: Wide Expert Parallelism and RDMA KV Cache Transfer for MoEOct 11 AMD MI355X Beats B200: Full-Stack Optimization Breakdown for 5% Lower TCO on DeepSeek-R1 InferenceMay 30