riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-31 10:12:23 +06:00
Code Issues Actions 23 Packages Projects Releases Wiki Activity

fixed typo in fp16 training section for perf_train_gpu_one (#19736)

Browse Source
This commit is contained in:
Dhruv Singal 2022-10-24 07:04:28 -07:00 committed by GitHub
parent 8db92dbe26
commit 5cbf1fa8ca
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
docs/source/en/perf_train_gpu_one.mdx
View File

@ -311,7 +311,7 @@ We can see that this saved some more memory but at the same time training became
## Floating Data Types
The idea of mixed precision training is that no all variables need to be stored in full (32-bit) floating point precision. If we can reduce the precision the variables and their computations are faster. Here are the commonly used floating point data types choice of which impacts both memory usage and throughput:
The idea of mixed precision training is that not all variables need to be stored in full (32-bit) floating point precision. If we can reduce the precision the variables and their computations are faster. Here are the commonly used floating point data types choice of which impacts both memory usage and throughput:
- fp32 (`float32`)
- fp16 (`float16`)
@ -328,7 +328,7 @@ While fp16 and fp32 have been around for quite some time, bf16 and tf32 are only
### FP16 Training
The idea of mixed precision training is that no all variables need to be stored in full (32-bit) floating point precision. If we can reduce the precision the variales and their computations are faster. The main advantage comes from saving the activations in half (16-bit) precision. Although the gradients are also computed in half precision they are converted back to full precision for the optimization step so no memory is saved here. Since the model is present on the GPU in both 16-bit and 32-bit precision this can use more GPU memory (1.5x the original model is on the GPU), especially for small batch sizes. Since some computations are performed in full and some in half precision this approach is also called mixed precision training. Enabling mixed precision training is also just a matter of setting the `fp16` flag to `True`:
The idea of mixed precision training is that not all variables need to be stored in full (32-bit) floating point precision. If we can reduce the precision the variales and their computations are faster. The main advantage comes from saving the activations in half (16-bit) precision. Although the gradients are also computed in half precision they are converted back to full precision for the optimization step so no memory is saved here. Since the model is present on the GPU in both 16-bit and 32-bit precision this can use more GPU memory (1.5x the original model is on the GPU), especially for small batch sizes. Since some computations are performed in full and some in half precision this approach is also called mixed precision training. Enabling mixed precision training is also just a matter of setting the `fp16` flag to `True`:
```py
training_args = TrainingArguments(per_device_train_batch_size=4, fp16=True, **default_args)