开发者发布适配64GB Jetson AGX Orin的Strata分支
一名Reddit开发者发布了Strata v0.1.45的分支版本，面向64GB Jetson AGX Orin，主要用于运行Qwen3.6-35B-A3B。该开发者称，llama.cpp启用MTP-3时解码速度约为21 tok/s，此前Strata运行约为30 tok/s，而采用采样MTP-2的受控短提示测试达到约40至41 tok/s。作者表示，这些结果来自其特定环境，并非llama.cpp与Strata之间的匹配基准。该分支还加入了统一INT8 KV缓存、前缀检查点和可选KV预分配。开发者披露代码大量由AI辅助完成，并表示已在其Orin设备上测试，同时邀请其他用户反馈。
Hey guys, hope all is well.
I've spent the last few days working on a fork of Strata v0.1.45 for the Jetson AGX Orin 64GB, mainly running Qwen3.6-35B-A3B.
I was getting around 21 tok/s decode with MTP-3 in llama.cpp, versus around 30 tok/s in my earlier Strata runs. With sampled MTP-2 enabled, my latest controlled short-prompt tests reached roughly 40–41 tok/s. These weren't matched llama.cpp-versus-Strata benchmarks, so treat them as results from my setup.
The fork also includes unified INT8 KV caching, prefix checkpoints and optional KV preallocation.
Not sure how many other Orin owners are here, but figured I'd share in case it's useful: https://github.com/moontato/Strata-Orin
Full disclosure: heavily AI-assisted/vibe-coded. I've tested it on my Orin, but would welcome feedback from others. Credit to the original Strata project for the foundation.
#Strata v0.1.45#Jetson AGX Orin 64GB#Qwen3.6-35B-A3B#llama.cpp