It took me all day to figure this out. It turns out that while HQQ will go ahead and fill 180GB of memory to do this - there's absolutely no reason for it! I did this from a slow**, 200 GB swap partition. On the off chance someone at Mobius see this - please don't ask transformers to load a 45B param model on to the CPU if you're not actually going to... call the model at all? It took ten minutes at SATA 2 speeds - and that was because it was padded to FP32 (CPU mode, right?).
45 Gigaweights * 2 Bytes per weight * fp32/bf16 = 180 GB of system memory allocated.
I wish I had that.
**May have been zswap's fault. I'm pretty sure 200MB/s and an idle CPU isn't the best you can hope for when you're doing sequential reads from a 4.0x4 NVME device? My GPU fell asleep between optimization passes. It even has a Gamer LED on it. I'll fix my sysctl next time.
- Try
$ python -i untitled.py
having saved that script from the mobius hf repo because you'll be spending a while in IDLE figuring out
>>> model.save_quantized("/absolute/path/noromaid")
at the end and trust me, quantizing something chunky and then watching python shred it because the save directory is somehow a recursive lambda function and not a string is heartbreaking. I don't know if it was supposed to emit more than the model.pt and the config.json but I'm taking what I can get.
If anyone's looking to donate I could do with an Epyc Rome and perhaps another pair of H100s? I've embedded my XMR address in attention tensors with help from a realy horny embedding so when it starts generating gibberish right before the good stuff just paste that in to feather and send me all your money. Thanks! :)
i'm joking. that's a joke. I didn't do that.