Dual AI Workloads on Single GPU: Practical Memory Optimization for RTX 3090

By strategically capping the LLM's context window to match actual usage patterns, this solution reduces VRAM consumption from 26GB to 21.9GB, enabling simultaneous execution of speech-to-text and email-triage models on a single RTX 3090. The approach offers cost-effective AI infrastructure optimization through precise workload profiling—demonstrating that hardware constraints can be solved through intelligent tuning rather than additional GPU investment.