Compiling the Vision Encoder: Squeezing 3% More Throughput from Qwen3-VL on Hopper GPUs

· Dev.to