Update blog

Author: cmdr2

content/blog/2025-10-10-1760088945.md

@@ -84,4 +84,9 @@ At fp32:
 - IREE (CUDA): 5.4 ms / it
 - IREE (Vulkan): 12.8 ms / it
 ```
-{{< /details >}}
+{{< /details >}}
+
+### Footnote: What about "glue code"?
+Glue code is the code connecting the various sub-models with each other (Text Encoder, VAE, Unet, Transformer etc), as well as the scheduling/sampling code (Euler, Heun etc). This code won't be included in the ONNX export (because it isn't part of the model), and therefore won't be compiled by ML compilers. This glue code uses torch for modifying tensors and generating intermediate tensors (random, zeros etc).
+
+Which means this glue code will have to be ported to C++ (in order to avoid python and torch's installation size overhead). That's a decent amount of work. Projects like [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp) implement them in C++ but are those lines are fairly entangled with the rest of the codebase.