SingaporeRep
Member
Benchmark Addict
- Joined:
- Jul 2024
- Posts:
- 227
- From:
- Singapore, SG
Singapore to Singapore should be fine
I was hitting their public NAT64 from my home lab, not a Vultr instance. Different path entirely. Still died twice. Numbers dont lie.
fio, iperf, geekbench. results or gtfo.
garykwh
Member
Inference on a Budget
- Joined:
- Jun 2024
- Posts:
- 173
- From:
- San Jose, US
Back to the LLM angle: HuggingFace Hub now has proper AAAA records for most endpoints. I tested last month. The problem is the CDN they use for LFS downloads, thats still v4-only in some regions. So your metadata resolves but the actual weights time out. Infuriating.
For inference not training you can cache everything to a block storage volume with v4 then migrate to v6-only. I do this with RunPod actually, not RackNerd.
CUDA cores are my love language
ronwit
Member
Budget King
- Joined:
- Jun 2024
- Posts:
- 122
- From:
- Osaka, Japan
Actually I found my old guide!!! Its on my blog from last year. Basically:
- Install tayga, set prefix to 64:ff9b::/96
- Bind to your v4 interface on the jump host
- Point DNS to your own unbound with DNS64 enabled
- Profit
Actually this needs a v4 box somewhere so not free. But if you already have one for other stuff, might as well. My Minecraft server runs on this setup, players dont even know.
instant noodles, instant deploys