composer
composer copied to clipboard
Supercharge Your Model Training
# What does this PR do? adds a property decorator for each loggerdestination, so a user can easily get a loggerdestination by just calling `mlflow_logger = trainer.logger.mlflow` for example
# What does this PR do? Fix broken gpu daily tests by creating the tmp_path directory # What issue(s) does this change relate to? Before: https://github.com/mosaicml/composer/actions/runs/6385289342/job/17329711777 **Pytorch 1.3** ``` FAILED...
Is FlashAttention really used while using HuggingFaceModel supported as one of ComposerModel types.
Given that from PyTorch 2.0 the dynamic dispatch to FlashAttention happens if the required conditions satisfy, I do not find a way to ensure whether FlashAttention is used by default....
# What does this PR do? Instead of padding till max_seq_len, use the maximum length of the batch. I have provided a simple and concise solution. # What issue(s) does...
## 🚀 Feature Request EarlyStopper and ThresholdStopper cannot deal with complex Metrics such as [MAP](https://docs.mosaicml.com/projects/composer/en/stable/api_reference/generated/composer.metrics.MAP.html#map). See: https://github.com/mosaicml/composer/blob/8d4dfaf017369bc0d27bc4f707e0fa709005c7b5/composer/callbacks/early_stopper.py#L122 https://github.com/mosaicml/composer/blob/8d4dfaf017369bc0d27bc4f707e0fa709005c7b5/composer/callbacks/threshold_stopper.py#L100 This line should only convert the metric to a tensor if it...
## 🚀 Feature Request Can I save latest checkpoint when crashed or press Ctrl+C? ## Motivation ## [Optional] Implementation ## Additional context
I have been experiencing llm-foundry/eval takes a lot of time compared to lm-evaluation-harness. After digging into the code, I found padding token is appended till the maximum length of the...
## 🚀 Feature Request Allow Time conversions ## Motivation For custom callbacks, it would be nice to expose `_convert_time` in the API. Hopefully, it is expose `_convert_time` in the API....
# What does this PR do? This PR attempts to support wall clock time as a interval metric. Users will be able to set wall clock time in training yamls...
# What does this PR do? Adds an API for extracting model state dict