François Chollet (@fchollet) “Keras 3.1 introduced in-place int8 quantization for Dense/EinsumDense (and thus” — TopicDigg
Keras 3.1 introduced in-place int8 quantization for Dense/EinsumDense (and thus all Transformers). It delivers ~20% speedups over float16 on JAX & TF with recent GPUs.
