Expose memmap_dtype in the data configuration

allenai / OLMo

Modeling, training, eval, and inference code for OLMo

https://allenai.org/olmo

Apache License 2.0

4.48k stars 449 forks source link

Expose memmap_dtype in the data configuration #595

Closed leon-g-xu closed 3 months ago

leon-g-xu commented 4 months ago

🚀 The feature, motivation and pitch

Can we request to add this feature in the OLMo framework? We'd like to expose the memmap_dtype as a config in data configuration section. Currently the framework always assumes that the memmap_dtype is np.uint16. In our run we actually need to use np.uint32. Here's the PR to expose this config: https://github.com/allenai/OLMo/pull/594

Alternatives

No response

Additional context

No response

leon-g-xu commented 4 months ago

Is there a process on how I can contribute back to the code? How do I get my PR reviewed?

2015aroras commented 3 months ago

Fixed in #594