fairseq2 Files

[0.6.0] - Oct 7th, 2025

• fairseq2.sharder is deprecated. fairseq2 now expects parallelism strategies to be applied within model factories. This gives model authors full control over how parallelism is applied to their models. More info
• Gangs can now be used as a context manager, along with a new maybe_get_current_gangs() helper function. This feature is particularly useful in procedural programming, as it eliminates the need to pass a Gangs instance through every function call. More info
• An experimental implementation of LLaMA 4 Scout model is now available.
• The recipe command line interface now accepts a new --no-exit-on-error flag to allow post-mortem debugging of recipe processes. More info
• The optimizer and learning rate scheduler recipe configurations now support multiple parameter groups. This is in particular convenient for models that require more than one learning rate to train (e.g. GAN models). More info
• The regime.save_model_only recipe option now accepts 'all' and 'all_but_last' as alternatives to a boolean value. Setting the option to 'all' is equivalent to True and means that only the model state is saved during checkpointing. This is beneficial for short-lived training jobs where the user does not expect to resume the job but requires frequent snapshots of the model for evaluation purposes. In this mode, checkpointing is faster and disk space is saved by avoiding the storage of trainer, optimizer, and data reader states. The 'all_but_last' option is similar to 'all', except that the full state is saved only for the last checkpoint while all previous checkpoints will store only the model state, as in the 'all' mode. This is helpful to avoid unnecessary disk space use if the user does not plan to branch off the training from a previous checkpoint.
• The default resume mode for Weights & Biases metric recorder changed from 'allow' to None to avoid noisy, safe-to-ignore warnings when resuming a preempted job.