File size: 2,311 Bytes
398d5c5 d46c9e3 ea3256c d46c9e3 e831476 d46c9e3 e831476 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 |
---
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).
## Bert Base model HPU configuration
This model contains just the `GaudiConfig` file for running the [bert-base-uncased](https://huggingface.co/bert-base-uncased) 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 there are a few new training arguments specific to HPUs.
[Here](https://github.com/huggingface/optimum-habana/blob/main/examples/question-answering/run_qa.py) is a question-answering example script to fine-tune a model on SQuAD. You can run it with BERT with the following command:
```bash
python run_qa.py \
--model_name_or_path bert-base-uncased \
--gaudi_config_name Habana/bert-base-uncased \
--dataset_name squad \
--do_train \
--do_eval \
--per_device_train_batch_size 24 \
--per_device_eval_batch_size 8 \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--output_dir /tmp/squad/ \
--use_habana \
--use_lazy_mode \
--throughput_warmup_steps 2
```
Check the [documentation](https://huggingface.co/docs/optimum/habana/index) out for more advanced usage and examples.
|