rrhjm53270

I enjoyed the smooth transition from minio to RustFS quite a lot.

My impression: the re-aristocratization of scientific research seems inevitable.

I tried kernel autoreasearch using DeepSeek-V4-Flash as well. It spent about 1-2 hours to complete the FlashAttention optimization job (https://github.com/fengwang/FA5090/tree/main/v7) and cost me only $0.2. I believe we are ready to offload a lot of this kind well-defined constrained optimization problems to AI Agent autoresearch.

I guess this one is for Qwen3.8-2.4T-A95B from the URL

Well, I think Chinese doesn't really have any grammar most of the time.

Good job. Thanks for sharing. I have a similar NAS server with 2 Intel(R) Xeon(R) CPU E5-2699 CPUs. I will test as well.

Would it possible to stack up to 16x32GB VRAM, and test the performance of a MOE model such as Deepseek-v4-flash?

Ironic

Putin and Trump have already made the world chaotic enough; don't let Xi get involved too.

It is a bit off-topic, but would it possible to provide a light mode for this blog? I used to work during the day time, and my pupils had to contract to read, making it a very poor reading experience.

As an individual, this paper suggests that I must move beyond "Is this news true?" to "Why does it look like everyone agrees with this?"

Why not let AI write the code and then have it reviewed by humans? If you use AI to review my code, then you can't stop me from using another AI to refute it: this only foreshadows the beginning of internal friction.

A trick that I used was

``` struct dummy{}; dummy d = ANYTHING_YOU_WANT_GO_GET_THE_TYPE; ```

Compile it with g++, and get the type info from compilation error -:)

It is very interesting that among top-20 models, all non-proprietary ones are from China.

Do you have any plan for the serialization and deserialization of your tensor and nn library?

I prefer write LaTex document, compile to pdf and then convert pdf to html using pdf2htmlEX.

Thank you for sharing such an interesting work.

A little comment: adding some more aggressive optimization optimization options to simd C++ code to see the performance difference.

On my side with a AMD Ryzen 9 7900X3D CPU, I have

- 0.0592569 ms for `-O3 -march=native` option, and - 1.7741e-05 ms for `-funsafe-math-optimizations -Ofast -flto=auto -pipe -march=native`

I did roughly the same thing in one of my hobby project https://github.com/fengwang/nnl. But in stead of using SSD, I load all the weights to the host memory, and while inferencing the model layer by layer, I asynchronously copy memory from global to shared memory in the hope of better performance. However, my approach is bounded by the PCI-E bandwidth.