记: 第三周，今天跑起了多机RL，在给miles/sglang加distributed lora training的feature，这张卡当train engine, 另一边2X5060Ti当rollout engine，家用交换机互联，已经跑起来了，十分开心 pr:

🤖 sglang ⭐ 28,406 stars Stop wasting GPU cycles. Serve LLMs and multimodal models at warp speed with a high-performance serving engine built for maximum throughput. 🔗 #AI #MachineLearning

This change to SGLang allows peers to recover after returning too many errors by probing them with Mooncake send_probe. The permanent denylist was causing a low volume of errors because the pods…

Landed changes to SGLang to address two use-after-free bugs in decode side KV cache. This was causing garbled output with Kimi K2.6 under heavy concurrency when using decode KV cache offload.

componentになっているかは少し古い場合はその様なコメント(