SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Just re-reading the whole doc every couple of months 😬 (#18489)

Browse Source 
* Delete valohai.yaml

* NLP => ML

* typo

* website supports https

* datasets

* 60k + modalities

* unrelated link fixing for accelerate

* Ok those links were actually broken

* Fix link

* Make `AutoTokenizer` auto-link

* wording tweak

* add at least one non-nlp task
This commit is contained in:
Julien Chaumond
2022-08-06 09:38:55 +02:00
committed by GitHub
parent b8c247b6d0
commit 8d1f9039d0
20 changed files with 39 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/en/perf_train_gpu_one.mdx
View File

@@ -609,7 +609,7 @@ for step, batch in enumerate(dataloader, start=1):
optimizer.zero_grad()
```
First we wrap the dataset in a [`DataLoader`](https://pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader). Then we can enable gradient checkpointing by calling the model's [`~PreTrainedModel.gradient_checkpointing_enable`] method. When we initialize the [`Accelerator`](https://huggingface.co/docs/accelerate/accelerator.html#accelerate.Accelerator) we can specifiy if we want to use mixed precision training and it will take care of it for us in the [`prepare`] call. During the [`prepare`](https://huggingface.co/docs/accelerate/accelerator.html#accelerate.Accelerator.prepare) call the dataloader will also be distributed across workers should we use multiple GPUs. We use the same 8-bit optimizer from the earlier experiments.
First we wrap the dataset in a [`DataLoader`](https://pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader). Then we can enable gradient checkpointing by calling the model's [`~PreTrainedModel.gradient_checkpointing_enable`] method. When we initialize the [`Accelerator`](https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator) we can specifiy if we want to use mixed precision training and it will take care of it for us in the [`prepare`] call. During the [`prepare`](https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator.prepare) call the dataloader will also be distributed across workers should we use multiple GPUs. We use the same 8-bit optimizer from the earlier experiments.
Finally, we can write the main training loop. Note that the `backward` call is handled by 🤗 Accelerate. We can also see how gradient accumulation works: we normalize the loss so we get the average at the end of accumulation and once we have enough steps we run the optimization. Now the question is: does this use the same amount of memory as the previous steps? Let's check: