README.md

---
license: apache-2.0
---

[Optimum Habana](https://github.com/huggingface/optimum-habana) is the interface between the Transformers library and Habana's Gaudi processor (HPU). It provides a set of tools enabling easy and fast model loading and fine-tuning on single- and multi-HPU settings for different downstream tasks.
Learn more about how to take advantage of the power of Habana HPUs to train Transformers models at [hf.co/hardware/habana](https://huggingface.co/hardware/habana).

## GPT2 model HPU configuration

This model contains just the `GaudiConfig` file for running the [GPT2](https://huggingface.co/gpt2) model on Habana's Gaudi processors (HPU).

**This model contains no model weights, only a GaudiConfig.**

This enables to specify:
- `use_habana_mixed_precision`: whether to use Habana Mixed Precision (HMP)
    - `hmp_opt_level`: optimization level for HMP, see [here](https://docs.habana.ai/en/latest/PyTorch/PyTorch_User_Guide/PT_Mixed_Precision.html#configuration-options) for a detailed explanation
    - `hmp_bf16_ops`: list of operators that should run in bf16
    - `hmp_fp32_ops`: list of operators that should run in fp32
    - `hmp_is_verbose`: verbosity
- `use_fused_adam`: whether to use Habana's custom AdamW implementation
- `use_fused_clip_norm`: whether to use Habana's fused gradient norm clipping operator

## Usage

The model is instantiated the same way as in the Transformers library.
The only difference is that the Gaudi configuration has to be loaded and provided to the trainer:

```
from optimum.habana import GaudiConfig, GaudiTrainer, GaudiTrainingArguments
from transformers import GPT2Tokenizer, GPT2Model

tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = GPT2Model.from_pretrained('gpt2')
gaudi_config = GaudiConfig.from_pretrained("Habana/gpt2")
args = GaudiTrainingArguments(
    output_dir="/tmp/output_dir",
    use_habana=True,
    use_lazy_mode=True,
)

trainer = GaudiTrainer(
    model=model,
    gaudi_config=gaudi_config,
    args=args,
    tokenizer=tokenizer,
)
trainer.train()
```