FNsi

It's bf16 and I can't run it in my device too.

"bf16 which is only available on Ampere and later, I would expect some performance degradation if running it in fp16 instead" 😅

Looks better than unlimitformers! And that let the real time training possible? The model can be dynamic no more static! Also that means, other classic algorithms may work fine with...

> If I understand correctly, implementing infinite context length with NBCE would require infinite computing power, but perhaps I misunderstood? Large ctx leads large memory use, and I think it's...

> - Add program and argument descriptions the `ArgumentParser` > > - Remove the default argument values and set them as required instead (for the model, you could also default...

No clue, but I assume they would like to open that script while they are working on it🤦

I don't think that will work fine though. Many copys from devices will simply reduce the speed.

> It was posted 6 months ago 😱. i guess we have no hope. 😭 8 months 🙋

Those figures prove 1T data 7B is equal to 0.5T data 13B, then consider the ratio, at some point, like 3T data, they might have the same performance.

Set to 0.125 with 16k context Lora, did have a good result... Sorry, that won't happen in openblas. I rebuild with Blas and run shit..