| language: | |
| - ja | |
| license: mit | |
| tags: | |
| - ja | |
| - japanese | |
| - gpt_neox | |
| - gpt | |
| - text-generation | |
| - lm | |
| - nlp | |
| - int8 | |
| - neural-compressor | |
| - Intel® Neural Compressor | |
| - PostTrainingStatic | |
| datasets: | |
| - oscar | |
| model-index: | |
| - name: gpt-neox-japanese-2.7b-int8 | |
| results: | |
| - task: | |
| name: Text Generation | |
| type: text-generation | |
| dataset: | |
| name: oscar | |
| type: oscar | |
| args: unshuffled_original_ast | |
| metrics: | |
| - name: Acurracy | |
| type: loss | |
| value: 4.9920 | |
| # INT8 gpt-neox-japanese-2.7b-int8 | |
| ## Post-training static quantization | |
| ### PyTorch | |
| This is an INT8 PyTorch model quantized with [Intel® Neural Compressor](https://github.com/intel/neural-compressor). | |
| The original fp32 model comes from the fine-tuned model [abeja/gpt-neox-japanese-2.7b](https://huggingface.co/abeja/gpt-neox-japanese-2.7b). | |
| The calibration dataloader is the train dataloader. The default calibration sampling size 100 isn't divisible exactly by batch size 8, so the real sampling size is 104. | |
| #### Test result | |
| | |INT8|FP32| | |
| |---|:---:|:---:| | |
| | **Accuracy (eval-loss)** |4.9920|3.5219| | |
| | **Model size (MB)** |2570|5360| | |
| #### Load with Intel® Neural Compressor: | |
| ```python | |
| from optimum.intel import INCModelForCausalLM | |
| model_id = "Intel/gpt-neox-japanese-2.7b-int8" | |
| int8_model = INCModelForCausalLM.from_pretrained(model_id) | |
| ``` | |