●builderYou can potentially reduce TTFT in multi-turn multimodal applications by utilizing idle compute during the decoding phase.
●researcherYou can use this framework to study the trade-offs between prefill latency and context completeness in MLLMs.