Evaluation

CPU inference and quantization quality evaluation

CPU-only inference speed on four architectures with gemma.cpp (8-bit switched-floating-point quantization) using a PaliGemma 2 3B (224px2) checkpoint fine-tuned on COCOcap; quality comparison between Jax/f32 inference on TPU and quantized gemma.cpp inference on CPU for five fine-tuning datasets (chosen for task coverage)