🫙 Updated readme on archiving.

Author: rentruewang

README.md

@@ -1,3 +1,16 @@
+> <h2 id="archive"> 🫙 Archive </h1>
+
+> Project `koila` has 2 main components:
+
+>   1. Metadata tracking for `torch.Tensor`s.
+>   2. Decouple some symbolic info (batch size) to run a reduced graph, with gradient accumulation to prevent OOM.
+
+> For 1.: Now `PyTorch` officially has <a href="https://docs.pytorch.org/docs/stable/torch.compiler_fake_tensor.html"> FakeTensor</a> (koila predates it). It has great compatibility and support of torch operators, something `koila` never was able to do.
+
+> For 2.: `Koila` only tracks symbolic info partially, on the batch dimension. I have now something a lot better, a compiler for deep learning `aioway` <a href="https://github.com/rentruewang/aioway">link</a> that handles all these info, which by the time of writing (2025/09/14) is not yet open source, but will be soon.
+
+> As `FakeTensor` is now available, and `aioway` in development, I think to keep `koila` as it is, a POC that I did for fun.
+
 # 🐨 Koila
 
 > Koila solves `CUDA error: out of memory error` painlessly.

docs/_coverpage.md

@@ -4,3 +4,4 @@ No more `CUDA error: out of memory error`.
 
 [Github](https://github.com/rentruewang/koila)
 [Website](https://koila.rentruewang.com)
+[Archive](#archive)