Milo Cress

Mid Epoch Resumption of Streaming Datasets

2

Mid Epoch Resumption of Streaming Datasets Uses a Feistel network to reproducibly and efficiently shuffle streaming datasets while constraining the maximum number of shards that must be present on the...

# What does this PR do? Allows `load_fsdp_monolith_rank0_only` and `use_orig_params` to both be set to `True` - this was initially disabled due to a pytorch issue I think, but has...

Hourly token and pool volume return $0

2

This issue is partially mentioned in https://github.com/Uniswap/v3-subgraph/issues/79 Pool and token volume return 0 even when the volume is clearly not 0. Consider the query: ```graphql { tokenHourDatas(first: 2, orderBy: periodStartUnix,...

log eval dataset misconfiguration

1

# Make Eval Dataset Misconfiguration Errors Visible through Mosaic Logger Wraps eval dataset creation with a mosaic logger try/catch. ## In Train context `mpt-125m-chinchilla-regression-6OSoWo` (the log trace indicates that the...

https://github.com/mosaicml/composer/blob/da18d9e4513bc1629849158f5a16108350532adf/composer/trainer/trainer.py#L1042C1-L1043C1 Uses the default eval interval from composer rather than forcing eval_interval to be a required argument

Catches tokenization failures on custom HF datasets with missing/extra/mistyped columns.