DeepSeek v4.1 Flash适配Dwarfstar量化方案
DeepSeek v4.1 Flash 1017 for Dwarfstar (ds4)
本地部署DeepSeek的同学必看，给出了具体的ds4量化配方和M5 Max性能数据，直接照做即可跑通Vision模型。
Hey y'all, just following up on my PR to get the new vision tower for 1017 working in ds4 with the quant I promised (the language model loads already, but vision needed my PR). I built to the exact same recipe antirez used for his original DS4.1 Q2: 2-bit routed experts with 8-bit or higher for everything else, combined with the original FP8 engram table. This worked well for the original DS4.1, and initial testing looks good here too. The imatrix is built from 1017, not the original checkpoint.
大家好，跟进一下我的 PR，旨在让 ds4 中的 1017 新视觉塔（vision tower）按照我承诺的量化方式运行（语言模型已经可以加载，但视觉部分需要我的 PR）。我使用了与 antirez 为其原始 DS4.1 Q2 构建完全相同的配方：2-bit 路由专家（routed experts），其余部分为 8-bit 或更高精度，并结合原始的 FP8 engram 表。这在原始 DS4.1 上效果良好，初步测试在这里看起来也不错。imatrix 是基于 1017 构建的，而非原始检查点（checkpoint）。
The wired backbone of the model is just over 150GB, so it will stream from SSD on 128GB systems with nearly all of the routed experts held in the hot cache, allowing for decode speeds up to 20tps on an M5 Max. The ~200GB engram table is paged in from SSD on-demand, not held in RAM.
模型的有线骨干网（wired backbone）大小略超过 150GB，因此在配备 128GB 内存的系统上可以从 SSD 流式传输，同时几乎所有路由专家都保留在热缓存中，从而在 M5 Max 上实现高达 20 tps 的解码速度。约 200GB 的 engram 表按需从 SSD 分页加载，不占用 RAM。
This model is significantly revamped, with ~40% of the routed expert tensors changed, 30% of engrams, and ~50% of the vision tower's expert tensors altered. By all appearances, it's the new leading open-weights model, so I had to give it a try. If you are curious too, let me know what you think!
该模型经过大幅改进，约 40% 的路由专家张量被更改，30% 的 engrams 被修改，以及约 50% 的视觉塔专家张量被调整。从各方面来看，它是新的领先开源权重模型，所以我必须尝试一下。如果你也好奇，请告诉我你的看法！
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力