Tag: 分布式系统 (2 articles)

Hugging Face's TRL library introduces delta weight sync, transmitting only the ~1-2% of weights that change between RL steps, reducing sync overhead by two orders of magnitude and making trillion-parameter async RL training dramatically cheaper.

By integrating Mooncake's distributed KV cache store, vLLM overcomes the efficiency bottleneck of recomputing long-context prefixes in AI Agent workloads, achieving a 3.8x throughput increase and 46x lower time-to-first-token.