黑客基础设施
1 分•作者: yjtpesesu2•5 个月前
潜水已久，注册过多个账号，一个一个来，绝无恶意。大家好。昨天关于在 Hugging Face 上针对流行的开源权重模型进行层复制和调整的讨论，促成了这次提交。
自从 GPT ~3.5 出现以来，很明显计算机已经可以模拟人类，至少在计算机看来是这样。 “互联网已死”理论实际上起源于 2012 年左右，但我一直难以找到验证，包括在 archive.org 上搜索。
所有这些动荡使得本地离线部署变得如此重要。
以下是当前黑客的基础设施：
1. 选择一个平台，然后确定 Linux 内核。
2. qemu/kvm，或者 incus，或者 docker，或者 lxc/lxd，关键是：第二步要与第一步选择的平台隔离。
3. 选择使用 kvm，因为它拥有最大的开发社区。相信我：你不会想被 bhyve 或 jails 相关的依赖项，或者 WSL2（或 1）所困住。
4. 主流，主线，主要，全天候。
5. 使用 libvirt，或者，也许直接选择简单的选项并使用 qemu 模拟。
6. 如果与宿主内核的交互经过定制，宿主操作系统可以以很小的开销或引入延迟来模拟虚拟机的内核。
7. 在宿主机上整理好一切（使用 incus？），将一些裸金属资源传递给你的“宠物”，一些高性能核心和几十 GB 的内存，也许还有一两块 6000 系列 GPU，一些网卡，这并不重要，所有宿主硬件都被传递了。
8. 除了 hypervisor。是 xen 吗？不可能，不理想，正在失去市场份额。很可能 qemu/kvm = LINUX。
9. HYPERVISOR 达到最佳状态。测量局部和全局的极小极大值，并达到最佳全局顶点...
10. 盈利？？？
查看原文
Long time lurker, many accounts, one at a time, no abuse. Hi. Yesterday's recount about layer duplication and adjustment for popular open weight models on huggingface, led to this submission.
Since GPT ~3.5 it has been apparent that computers can simulate human, as far as a computer is concerned. The dead-internet theory actually originated circa 2012, but I've had difficulty finding verification, including searching the archive.org .
All this turmoil makes offline on prem so important.
Here is the current hacker infra:
1. choose a platform, then decide on the linux kernel.
2. qemu/kvm maybe incus, or docker or lxc/lxd, point being: step two is isolate from the platform chosen in step one
3. choose to use kvm, because it has the largest development community. believe me: you don't want to get bogged down with the associated dependencies of bhyve or jails, or, for that matter, WSL2 (or 1).
4. mainstream, mainline, main, all day.
5. either libvirt or, maybe just choose the easy option and q emulate.
6. the host os can emulate the virtual machine's kernel with very little overhead or introduced latency if the interactions with the host kernel are tailored.
7. after sorting everything out on the host (use incus?) pass through some bare metal to your pet, some performance cores and a couple dozen gigs of ram, maybe a 6000 gpu or two, some nics, doesn't matter, all the host hardware is passed through
8. except the hypervisor. is it xen>? Can't be, not ideal, losing market share. Likellllly qemu/kvm = LINUX.
9. HYPERVISOR OPTIMUM ACHieved. Local and global minmax measured and optimal global apex...
10. profit???