3 ms·
Does it? I loaded Guanaco-65B on a 48GB graphics card using the new flag and it loaded just fine. I don't think the unquantized model fits on a 48GB card. So I'
by amilios 3y ago
Does it? I loaded Guanaco-65B on a 48GB graphics card using the new flag and it loaded just fine. I don't think the unquantized model fits on a 48GB card. So I'm not quite sure if this is true.
- MacsHeadroom 3y agoRAM, not VRAM. The model is fully loaded into (regular, system) RAM then quantized while being loaded into VRAM. So you still need enough system RAM (or RAM+Swap) to load the unquantized model.
- dTal 3y agoWhy would it need to be loaded fully into system RAM at once? Couldn't it be quantized on the fly, as it's being read from disk?
- MacsHeadroom 3y agoI'm not sure it could be for this complex quantization scheme. It's not a simple round to nearest quantization. Even if it is possible, it would require someone to develop that functionality. Currently it does load the entire unquantized model into RAM first.
- emikulic 3y agoLast time I looked into this, the answer was "because huggingface transformers and torch.load are written to do it this way" You could absolutely do something streaming, or mmap the weights instead of loading them into system RAM. Just the default interfaces don't.