Is this library much slower than AutoGPTQ?

[CHNtentes]

I tried quantizing the same model with same quantize config, with AutoGPTQ I can finish in 15~20 minutes, but with this library I need over 2 hours...

[Qubitium]

Please use and increase batch_size value when calling model.quantize(). gptqmodel is faster than auto-gptq if used correctly.

auto-gptq has broken batch support for calibration. gptqmodel has batching support but you need to set to a proper value accordingly to your gpu capability and vram size.

[Qubitium]

@CHNtentes Was the speed issue resolved by batch_size?

[CHNtentes]

I tried 16 and it was faster, but still took around 30 min. Tried 64 but the speed only increased a little bit, and the vram usage was much higher...

[Qubitium]

Please post you full quantization code including path to HF hosted base model and/or datasets so we can verify this slowness as normal or regression.

[Qubitium]

@CHNtentes Memory usage will increase as batch_size increases but as you have found out, more is not better. Find the large batch_size that both target your gpu vram and also bring about faster quant speed. Having huge batch doesn't help when the gpu cores is not able to run such a large batch in parallel.

[Qubitium]

@CHNtentes In addition to batch_size=16, please test using our latest v1.5.3 release. Make sure your import gptqmodel in your python code before all torch/transformer package to get the memory saving. v1.5.3 has been tested to reduce memory by 25-30% on Nvidia gpu and now uses about 15% memory less than AutoGPTQ using same batch, model (QwQ 32B), same dataset.

Quantization time is a bit slower than AutoGPTQ at the moment but we hope to fix that in our next release.

More importanly, GPTQModel's error_loss is consistently lower for almost all layers vs AutoGPTQ apples to apples comparision using QwQ 32B model as test with same batch and calibration dataset. Lower error_loss will often result in better post-quantized models.

https://github.com/ModelCloud/GPTQModel/releases/tag/v1.5.3

[Qubitium]

@CHNtentes Memory usage issue has been fixed in https://github.com/ModelCloud/GPTQModel/releases/tag/v1.6.0

We are now 35% lower vram usage than before and 15% lower than AutoGPTQ and our test quants per layer, using same calibration data, with QwQ 32B test GPTQModel has consistent lower error_loss which is critical for quantization. Speed is about 7.5% slower than AutoGPTQ for QwQ 32B but with lower vram usage and higher quality quants, it is worth the one-off cost. We will try to improve the speed in our next release.

[CHNtentes]

Thanks for your reply. Will increasing batch_size result in higher loss? I tried batch_size 64 and it seems that the loss is slightly higher.

[Qubitium]

Do not use high batch size if it doesn't result in faster quant. Use as low as possible while maintaining max quant speed. It may or may not cause higher error_loss. This is very calibration data and model dependent.

[CHNtentes]

I did some more experiments and found out actually batch_size=1 is fastest... Maybe last time there was other processes using gpu.

[Qubitium]

@CHNtentes Do you see faster quant speed, lower vram usage in our latest 1.6.0 release?

[CHNtentes]

Well, I did not do direct comparison, but it seems to be a bit faster.

[Qubitium]

We are now 90% less memory usage than AutoGPTQ on quantization of large models and should also be much faster with multi-gpu acceleration during quantization.