December 2025: Delivered end-to-end on-device sampling for dual QPC vision-language models in quic/efficient-transformers, enabling on-device sampling in the language-decoder path and reducing host overhead while boosting inference throughput for multimodal workflows. Implemented a robust fix for gumbel noise to accurately simulate multinomial sampling during random draws. The changes unify the VLM sampling flow with the existing QEffForCausalLM path, enabling seamless usage via QEFFAutoModelForImageTextToText with qaic_config, including include_sampler, return_pdfs, and max_top_k_ids. The work supports deployment for models like Qwen/Qwen2.5-VL-3B-Instruct and demonstrates explicit usage patterns through the commit 58fd3a7228d9a7d35bab79c597666c09fe06a380. This contributes to lower host overhead, higher throughput, and smoother edge deployments across multimodal inference tasks.