WASM / WebGPU Finetune (PoC)

Only F32 tensors get gradients. stories260K is fully F32, so the whole model trains.




Sample before training:
Sample after training:
Log: