Haozhe Zhu

I have explained it in previous issues. PyTorch has modified its API in previous version.

Hi Sophia, thank you for using my code. Because I dumped the floating-point weights (rather than the quantized ones) in the checkpoint. The quantization runs on the fly, generating quantized...

I am not very similar with deployment tools. I guess there would be some much convenient ways to generate a really quantized model (I mean all weights are integers, but...

Great. Thank you, Sophia. I'm still struggling with my theis. Maybe I will consider this feature in a few weeks (and also fix some known compatibility bugs).

Actually, a newer version is maintained in [NeuralZip](https://github.com/zhutmost/neuralzip),and you can find more experiment results there. But it is developed with PyTorch-Lightning framework, rather than the raw PyTorch.

I have no idea. I cannot explain it without more details. In my code, the validation/test epoch (just like inference) is much faster than training epoch.

For research proposes, you can run it with cpu backend by just modifying the config yaml. I am not similar with deployment in production environment. Maybe you can use the...

Looks that it is very rarely seen. s_scale is a floating pointed number, it generally does not become 0 during training. I will try to add an eps, when I...

I think combining spinalHDL with Scala to construct hardware is great. It is not a must to do everything in native SpinalHDL. For your question, you can read the file...

I guess you need a text file in binary encoding instead of a binary file? A temporary solution is ```Scala Source.fromFile("abc.txt" ).foreach{ .. // convert String to bits } ```...