Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Update eager.md #7710

Merged
merged 2 commits into from
Jul 19, 2024
Merged

Update eager.md #7710

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9773b39
Update eager.md
JackCaoG Jul 18, 2024
0265915
Update eager.md
JackCaoG Jul 18, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions docs/eager.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ The implementation of the `torch_xla.experimental.compile` is actually pretty st
```python
torch_xla.experimental.eager_mode(True)

compiled_model = torch_xla.compile(model, backend="openxla")
compiled_model = torch.compile(model, backend="openxla")
```
It is recommened to use the `torch.compile` instead of `torch_xla.experimental.compile` for inference to reduce the tracing overhad.

Expand All @@ -79,9 +79,7 @@ step_fn = torch_xla.experimental.compile(step_fn)
```
In training we asked user to refactor the `step_fn` out because it is usually better to compile the model's forward, backward and optimizer together. The long term goal is to also use `torch.compile` for training but right now we recommend user to use `torch_xla.experimental.compile`(for perfomrance reason).

## Performance

# Benchmark
## Benchmark

I run a 2 layer decoder only model training(it is pretty much just a llama2) with fake data on a single chip of v4-8 for 300 steps. Below is the number I observed.

Expand Down
Loading