July 2026 Summaries

At Team fal, significant improvements were achieved in the throughput of the Qwen 3.6 model for Ideogram V4's prompt expander, aiming to enhance user interactivity in text-to-image generation. By leveraging DSpark on SGLang, they managed to increase throughput by 16 times, primarily through the adoption of a 35B MoE model that balanced inference speed and image performance. This model was fine-tuned using PEFT for efficiency and stability. Despite initial challenges with speculative decoding and acceptance rates, the team incorporated DFlash, a diffusion-based model, and later integrated DSpark to further boost performance. The DSpark architecture involved using diffusion-based block predictors combined with Markovian heads to improve acceptance rates, which required modifications to both vLLM and SGLang for efficient model serving. Through extensive experimentation, they achieved ~1000 tokens per second with DSpark and ultimately settled on a configuration that provided 830 tokens per second, maintaining interactivity and efficiency while reducing the prompt expansion time to under two seconds.