This recipe is outdated, we recommend using symmetric quantization. You can remove --asym from the command.
A sample command to generate an INT4 model.
auto-round \
--model meta-llama/Llama-2-7b-chat-hf \
--device 0 \
--group_size 128 \
--bits 4 \
--iters 1000 \
--nsamples 512 \
--asym \
--format 'auto_gptq,auto_round' \
--output_dir "./tmp_autoround"Due to licensing restrictions, we are unable to release the model.
Install lm-eval-harness from source, and the git id 96d185fa6232a5ab685ba7c43e45d1dbb3bb906d.
Since we encountered an issue evaluating this model with lm-eval, we opted to evaluate the qdq model instead. In our assessment, we found that its accuracy closely matches that of the real quantized model in most cases except for some small models like opt-125m.
| Metric | FP16 | int4 qdq |
|---|---|---|
| Avg. | 0.5901 | 0.5897 |
| mmlu | 0.4640 | 0.4545 |
| lambada_openai | 0.7105 | 0.7037 |
| hellaswag | 0.5780 | 0.5706 |
| winogrande | 0.6638 | 0.6614 |
| piqa | 0.7639 | 0.7633 |
| truthfulqa_mc1 | 0.3023 | 0.3035 |
| openbookqa | 0.3340 | 0.3260 |
| boolq | 0.7976 | 0.8064 |
| rte | 0.6968 | 0.7292 |
| arc_easy | 0.7382 | 0.7336 |
| arc_challenge | 0.4420 | 0.4352 |