Optimizing Machine Translation Deployment with Document-Chunking and Quantization

Combining W4A8 or W8A8 quantization with a document-chunking strategy improves the latency-throughput Pareto-curve for EuroLLM and Hy-MT2 models on A100/H100 GPUs. The study evaluates performance across model sizes from 1.7B to 22B parameters under realistic server workloads.