The size of .npz weight file is twice larger than .pth file?

[Lance729]

I use convert.py to convert llama's pth file into npz file, but the size goes from 13GB to 26GB.

I don't know if it is correct.

[shayanjl]

A 2x increase from 13 GB to 26 GB usually means the converted weights were saved in a larger dtype, most likely float32 instead of float16.

So this may still be correct.

The main things to check are:

the --dtype used during conversion
whether you converted to a non-quantized MLX model
whether the original .pth and output .npz are being stored in different formats

If the output was saved as float32, then the size doubling is expected. If you want a smaller file, try converting with --dtype float16 or using quantization.