[Questions]: How to implement NF4/NF2 matmul kernel function?

[llCurious]

hi @TimDettmers .

The paper shows that you quantize the weights to 2/4 bits using NF format. I wonder how to you handle the input activations (denoted as x). Is x also quantized to 2/4 bits?

If not, do you dequantize the quantized weights to float format to perform the matmul using float kernels? This seems to slow down the inference efficiency despite lower memory footprint.

[neoweasley]

The paper "QLORA: Efficient Finetuning of Quantized LLMs" says "In practice, this means whenever a QLORA weight tensor is used, we dequantize the tensor to BFloat16, and then perform a matrix multiplication in 16-bit." As far as i'm concerned, NF16 and Double Quantization may slow down the training and inference efficiency, which means this work sacrifices time and performance for memory. However, this paper reports "24hours of finetuning on a single GPU", and I wonder how.