Skip to content

v0.8.0

Choose a tag to compare

View all tags
@cirquit cirquit released this 26 Mar 14:54
· 5 commits to main since this release
v0.8.0
f3bf3a4
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's changed

  • Breaking: Trainer, evaluator, generator, validator, and task moved from fairseq2.recipe to fairseq2 package root. (#1417)
  • Breaking: LM recipes restructured: text_generate renamed to generate, SFT configs removed/renamed, recipe config classes changed. (#1431, #1432, #1433)
  • Breaking: RecipeModel is deprecated. Access the model directly via .module instead. (#1403)
  • Breaking: pq.ParquetDataset replaced with pyarrow.dataset interface. (#1490)
  • Breaking: resolve_optional renamed to maybe_resolve. (#1462)
  • Breaking: Revised ModelCheckpointLoader API. (#1475)
  • Breaking: Refactored tensor sharded modules (embedding, projection, FFN, attention). (#1476)
  • fsspec integration for remote filesystem support. Checkpoints can be saved to and loaded from S3 via --checkpoint-dir s3://bucket/path/. Requires s3fs. (#1126)
  • New GlobalFileSystem replaces LocalFileSystem as default, dispatching to the appropriate backend based on URI scheme. (#1126)
  • PyTorch 2.9.1 and 2.10 (forward compatibility) are now supported. PyTorch 2.9 introduced breaking changes to LR scheduler return types, which have been addressed. (#1477, #1491, #1456)
  • New context managers for procedural programming: GangContext, DeviceContext, DataTypeContext, current_dtype. Eliminates need to pass state through nested function calls. (#1474, #1473, #1464)
  • CheckpointManager, Optimizer, and LRScheduler now exposed in RecipeContext. (#1461)
  • Synchronous asset loading across ranks for models and tokenizers. Use when all ranks need identical assets loaded simultaneously. (#1429, #1426)
  • CheckpointManager.register_save_hook allows custom logic during checkpoint saves. (#1439)
  • Config files now support ${env:<NAME>} to interpolate environment variables. (#1435)
  • --no-rich CLI flag disables rich text output for log parsing. (#1421)
  • Hugging Face export now runs in isolated process with saved command line and logs for debugging. (#1459, #1458, #1437, #1434)
  • Improved support for gated Hugging Face models. (#1422)
  • get_family utility functions for detecting model families. (#1454)
  • Gemma3n model family (E2B/E4B) with text + audio inference and SFT training. (#1496)
  • Generic HuggingFace model integration: load, shard, and train any HuggingFace CausalLM model directly through HgCausalLMAdapter without requiring a native fairseq2 reimplementation. Inc
    ludes FSDP sharding, HF tokenizer integration, and SFT recipe support. (#1479)
  • AssetDownloadManager gains local_only parameter and custom download subpath support. (#1423, #1425)
  • Recipes now set Python random and numpy seeds for reproducibility. (#1419)
  • Wandb metric recorder now respects wandb environment variables. (#1440)
  • Improved share_parameters implementation. (#1484)
  • Fixed cross_entropy with reduction="mean" to properly exclude padding tokens from the denominator. (#1455)
  • Fixed Flash3SDPA to support the flash-attn-3 v3.0.0 package API (flash_attn_3._C / torch.ops.flash_attn_3) in addition to the legacy flash_attn_3_cuda module. (#1495)
  • Fixed data pipeline sampling bug when allow_repeats=False with many pipelines. (#1471)
  • Fixed DataParallelFacade weakref errors. (#1447, #1436)
  • Fixed WER calculation to use lists instead of tensors. (#1413)
Assets 2
Loading