In a typical AMD fashion, they released a nice thing, and they're making the utmost effort to make sure no one whatsoever uses it and that it's quickly forgotte

In a typical AMD fashion, they released a nice thing, and they're making the utmost effort to make sure no one whatsoever uses it and that it's quickly forgotten. For some reason they used a "custom" architecture for this LLM, but the only custom thing about it is that they added an extra QK norm. This is a weird choice if the point of this model is to show that "hey, AMD's hardware can train LLMs too", because it makes it so that no software supports this architecture out-of-box. Why the hell didn't they simply use the llama architecture that everyone and their dog supports? Sure, if you're aiming for a superior model then having a different/better architecture is a valid choice, but then why shoot yourself in the foot and not use a per-head QK norm (like Gemma3 or the upcoming Qwen3) which is better than the per-tensor one they used? And of course, the model was released three weeks ago, and there's absolutely no effort on AMD's part to add support for its custom architecture absolutely anywhere. There's no PR to HF's transformers from what I can see, nor a PR to llama.cpp (even though it'd be a trivial to do, because, again, the custom architecture is not that very different). And any potential third parties are of course completely uninterested in adding this support, because of the anemic context length (only a mere 4k, in 2025!), and a poopy licence (where there already exist superior models on a true open license like e.g. Qwen2), so why would anyone use this model for anything?