That makes very little sense when you can just use WebGPU and have fine-grained control...

for something like LLM inference, having this kind of abstraction seems like quite a hindrance

Okay. I will give a raw WebGPU implementation a try and see if it is smaller and more efficient.