GitLab

tools/merge_datasets.py
- tool to merge multiple dataset files into a single dataset
- testing conducted and included in the megatron-testing repo https://gitlab-master.nvidia.com/ADLR/megatron-testing

tools/preprocess_data.py
- magic numbers changed to required command line arguments

megatron/data/indexed_dataset.py
- when merging, fix to properly update document index
- testing conducted and included in the megatron-testing repo (see above)
- fix follows this history https://github.com/bigscience-workshop/Megatron-DeepSpeed/pull/66

Do not cast data returned from indexed_dataset to int64, rely on caller to cast to appropriate type.

preprocess_data:
- Adds ability to not split sentences. This is used for gpt2 datasets.

- Adds ability to create multiple datasets from different json keys,
this is current untested.

indexed_dataset:
- Add new "get" function to get a portion of an entry.

Some cleanup of types, etc. in the helpers cpp code.

Also removed the fix_lua_indexing stuff from fairseq.