1 Release published by 1 person

• v0.4.0

  published Feb 8, 2025

47 Pull requests merged by 5 people

• Do not close gangs in case of an error

  #1024 merged Feb 15, 2025

• Keep having model_key in checkpoints

  #1023 merged Feb 15, 2025

• Report data read time in trainer

  #1022 merged Feb 14, 2025

• Introduce support for manual garbage collection

  #1021 merged Feb 13, 2025

• Introduce support for generic memory stats

  #1020 merged Feb 13, 2025

• Improve checkpoint conversion handling

  #1019 merged Feb 13, 2025

• Improve checkpoint conversion handling

  #1018 merged Feb 13, 2025

• Doc fix

  #1017 merged Feb 13, 2025

• extras example with keep_jsonl_keys option

  #1009 merged Feb 13, 2025

• Clean up dataset and tokenizer loaders

  #1015 merged Feb 13, 2025

• Improve error handling in extension setup

  #1014 merged Feb 13, 2025

• Improve model checkpointing

  #1013 merged Feb 12, 2025

• Filter torch warnings by regex

  #1012 merged Feb 12, 2025

• Fix stale documentation

  #1011 merged Feb 11, 2025

• Update README.md

  #1010 merged Feb 11, 2025

• Introduce support for explicit BLEU tokenizer in MT recipes

  #1008 merged Feb 10, 2025

• Ensure that the model is in train mode

  #1007 merged Feb 10, 2025

• Re-introduce determine_default_device

  #1006 merged Feb 10, 2025

• Expose LLaMA RoPE scaling function as a public API

  #1005 merged Feb 10, 2025

• Resolve pytest warnings

  #1004 merged Feb 8, 2025

• Bump to v0.5.0.dev0

  #1003 merged Feb 8, 2025

• Bump to v0.4.0

  #1001 merged Feb 8, 2025

• Fix PT2.6 linting issues

  #1002 merged Feb 8, 2025

• Support PyTorch 2.6

  #1000 merged Feb 8, 2025

• Add dataset extras option to recipes

  #999 merged Feb 7, 2025

• doc update sprint 4

  #993 merged Feb 7, 2025

• Last refactoring bundle for 0.4

  #998 merged Feb 7, 2025

• Simplify vLLM doc

  #996 merged Feb 3, 2025

• doc update sprint 3

  #977 merged Jan 31, 2025

• using dtype and device in hub.load

  #995 merged Jan 30, 2025

• Fix CheckpointManager bugs

  #994 merged Jan 27, 2025

• Update doc on vLLM support

  #981 merged Jan 26, 2025

• Generate Hugging Face config.json

  #991 merged Jan 25, 2025

• Improve IO error handling

  #990 merged Jan 25, 2025

• Improve best checkpoint handling

  #989 merged Jan 25, 2025

• Introduce abstract ASR model and revise eval recipe

  #988 merged Jan 24, 2025

• Use logprob scores in sampling generator

  #987 merged Jan 24, 2025

• Introduce to_gangs helper

  #986 merged Jan 24, 2025

• Fix LLaMA test

  #985 merged Jan 24, 2025

• Fix intermediate_size calculation in Llama config convert function

  #982 merged Jan 24, 2025

• Refactors first party recipes

  #984 merged Jan 24, 2025

• Support runtime context in recipe loaders

  #979 merged Jan 18, 2025

• Allow training checkpoints to contain rich objects

  #978 merged Jan 16, 2025

• Enforce right package import paths

  #974 merged Jan 16, 2025

• Fix LLaMA checkpoint

  #973 merged Jan 15, 2025

• Nit updates

  #972 merged Jan 15, 2025

• Move batching strategy to DataReadConfig

  #968 merged Jan 15, 2025

1 Pull request opened by 1 person

• Non-deterministic map operator option in data loading

  #980 opened Jan 20, 2025

8 Issues closed by 6 people

• `fairseq2 lm generate` cannot load a checkpoint produced by `fairseq2 lm instruction_finetune`

  #1016 closed Feb 14, 2025

• Additional option in fairseq2 BLEU computation

  #997 closed Feb 10, 2025

• [doc] add e2e dpo tutorial

  #992 closed Feb 10, 2025

• `from_generator` function for BuilderDataPipeline

  #353 closed Jan 30, 2025

• fairseq2.assets.download_manager.AssetDownloadError: The download of the etox dataset has failed. See nested exception for details.

  #983 closed Jan 24, 2025

• --dump-config errors

  #960 closed Jan 16, 2025

• llama convert_checkpoint fails due to malformed json

  #970 closed Jan 15, 2025

• Incorrect model config for llama3_2_3b

  #971 closed Jan 15, 2025

2 Issues opened by 1 person

• CUDA out of memory in mt training task

  #976 opened Jan 16, 2025

• How to fine-tune NLLB-200 model?

  #969 opened Jan 15, 2025

3 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Refactor parquet dataloader

  #867 commented on Feb 14, 2025 • 22 new comments

• Add BestRQ pretraining

  #873 commented on Jan 27, 2025 • 4 new comments

• using pipeline_builder shared pointer multiple times lead to segfaults

  #369 commented on Feb 6, 2025 • 0 new comments