3 ms·JetStream: Throughput+memory optimized engine for LLM inference on XLA devices2 points by lnyan 2y ago