Comments (9)
dlwh
commented on August 15, 2024
1
Ugh yeah I'll disable/turn into opt in until we figure this out. Thanks for the report. I should have tested it more carefully. The loss curve was approximately the same in my quick test but apparently it's no good.
from levanter.
dlwh
commented on August 15, 2024
i bet i messed up mqa
from levanter.
dlwh
commented on August 15, 2024
pretty sure this will be fixed in #596 (though I'll make it default on after i test it a bit more)
from levanter.
dlwh
commented on August 15, 2024
@versae can you try latest main with:
--model.attn_backend splash
and (the default:)
--model.attn_backend jax_flash
and lmk if things seem good? pretty sure it's correct now
from levanter.
versae
commented on August 15, 2024
It seems fixed! Haven't run extensive tests, but initial losses match now.
Also, the Enum keys for the AttentionBackend() class seems to be all upper case, so I had to use SPLASH and JAX_FLASH.
from levanter.
versae
commented on August 15, 2024
Nice catch BTW
from levanter.
dlwh
commented on August 15, 2024
thanks. I should have caught it before, but glad it's fixed now
from levanter.
dlwh
commented on August 15, 2024
the lowercase thing can be fixed by upgrading the draccus dependency btw
from levanter.
dlwh
commented on August 15, 2024
declaring it fixed in #598
from levanter.
Related Issues (20)
- Instant Data loader resumes
- Random thoughts on the new docker stuff HOT 15
- Launch script multislice HOT 6
- Sharding error when trying to serialize PEFT models HOT 19
- Change launch.py to not rebuild docker container on TPU node failure (with retries)
- preemptible in config wasn't registered as capacity_type HOT 1
- Seg fault when run unit test HOT 1
- Running on AWS Sagemaker HOT 4
- Ray Cluster Error at tokenizing documents HOT 6
- No accelerator found. Please run on a TPU or GPU HOT 10
- Installing torch on custom Levanter fork HOT 1
- Add a dumb "run_progress" for wandb x axis
- ModuleNotFoundError: No module named 'jax.experimental.maps' after main-brunch update (#662) HOT 11
- Permission 'storage.objects.get' denied on resource HOT 2
- Jax on Levanter HOT 2
- Is there a way to initialize Mistral from a locally downloaded checkpoint? HOT 1
- Multi-slicing training jobs fail due to Ray timeout error
- Question about MFU calculation
- More robust and dynamic multi-slicing TPU training HOT 1
- Regression at pip install in docker build HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from levanter.