Summary:
With >1 GPU, but only 1 node, all reduces fail when inputs are not bf16. This uses a modified copy of torch's grad norm to avoid failures
Test Plan:
- Run unit tests:
- Run single gpu training: `python -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100`
- Run 1 node, multi-gpu training `torchrun --nproc-per-node 8 -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100`
* allow loading of the entropy model directly
* remove unused argument
* remove spammy warning
* allow patch_batch_size to be adjusted in the forward() method
* revert to original patcher style, fix warning
* allow grads when calculating entropies
* fix grad flow
* return preds from calculate_entropies()
* remove legacy arg
* fix an error with monotonicity and small sequence lengths
* ensure patcher is serializable
* revert patcher to original
* remove unused import
Summary:
- Refactor local model configs to be separate and clearer
- Add attention arguments and correct which attention is used in local models
- Preparation for being able to have an entropy train script
- Fix failing unit tests
Test Plan:
Summary:
For compatibility with either local/nfs or S3 datasets, swap to fsspec.
Add a tool to compare local and remote filesystems
Test Plan:
- Ran regular train script
- Ran with config with data in S3
