llama.cpp model loading is far slower over NAS than local NVMe
In local tests by 丘山录, reading a 475.25 GiB model set took 39.6 seconds from onboard NVMe, 114 seconds from a local SSD, and 1 hour 13 minutes 40 seconds from a NAS over a 1 Gb link. Changing only the read block size on the same disk reduced full-read time from 125.1 seconds at 1 MB to 39.7 seconds at 128 MB, while NAS throughput was largely constrained by the network link. Separate tests found that combining shards improved read performance, with the benefit more pronounced over SMB than on local NVMe.
The measurements provide a reproducible deployment workflow for choosing model storage, read block sizes, and shard layouts in llama.cpp, while remaining specific to the tested hardware and single-run results.
llama.cpp 完整教程,放错盘读一遍多等 1 小时
结论先行:把模型塞 NAS 省空间的人,省的是硬盘钱,付的是每次 73 分钟的等待——这笔账算不过来,瓶颈在链路不在容量
这句判断来自一次实测量出来的数:同一份 475.25 GiB 权重,躺在同一块盘上,什么都不换,只改每次读多少字节,125.1 秒变成 39.7 秒。
环境、命令、量出来的数都在下面,可以直接照着重跑。
模型下载完,多数人的下一步是直接敲 llama-cli -m 模型.gguf。
我今天把「下载完到跑起来」中间那一段拆开量了一遍。
#llama.cpp#NVIDIA DGX Spark#NVIDIA GeForce RTX 5090 D#Intel SSDPF2KX019T1M