Model Support Mixture of Experts support Features fMHA support for chunked attention and paged kv cache Baichuan FP8 quantization support Memory optimization Reduced host memory when buildin...