v0.5.8

Many great updates!

What's Changed

• add missing r1 prompt style by @ali-alshaar7 in #1929
• fix incremental save for PyTorch 2.6 by @t-vi in #1928
• Update pyproject.toml by @syntheticgio in #1939
• fix: resolve failing CI by @Borda in #1944
• handle wrapped thundermodules in generate by @t-vi in #1955
• fix skip condition by @t-vi in #1956
• ci: use HF cache by @Borda in #1958
• nits for CI by @Andrei-Aksionov in #1940
• ci: split HF caching by @Borda in #1960
• bump: PT 2.6 + bitsandbytes & standalone tests by @Borda in #1959
• prune whitespaces for code readability by @Borda in #1962
• fixing various typos in examples & tutorials by @Borda in #1963
• fix n_query_groups for llama-3.1-405b by @ysjprojects in #1946
• tests: make flaky test due to connection issues by @Borda in #1964
• Fix: incorrect gradient accumulation steps bug by @ysjprojects in #1947
• fix: use default num_nodes=1 for back-compatibility by @Borda in #1967
• Do not wrap LoRA layers with FSDP by @janEbert in #1538
• Speculative decoding: Base implementation by @Andrei-Aksionov in #1938
• Better clarity on SFT dataset attributes by @ysjprojects in #1970
• Enforce Consistent Formatting and Validation for YAML Files by @Borda in #1977
• Apply Standard Formatting and Fix Import & Test Name Issues by @Borda in #1981
• Setting config.sliding_window_layer_stride explicity by @ysjprojects in #1972
• feat: add linear rope type by @k223kim in #1982
• feat: update tests for transformers 4.50.2 by @k223kim in #1983
• fix: test_tokenizer_against_hf by @Borda in #1984
• feat: replace sliding window type with offset by @k223kim in #1989
• ci: with pull_request_target by @Borda in #1992
• Phi4 mini by @ysjprojects in #1949
• aggregate val_loss by @ysjprojects in #1971
• feat: add local base freq for rope by @k223kim in #1993
• test: flexible wait for serve start by @Borda in #1996
• fix: replace sliding window configuration parameters to sliding windows indices by @k223kim in #1995
• QwQ-32B by @ysjprojects in #1952
• feat: run thunder tests as part of LitGPT CI by @deependujha in #1975
• try pyupgrade-up py38 by @Borda in #1999
• [1/4] feat: add gemma 3 27b by @k223kim in #1998
• [2/4] add gemma 3 1b by @k223kim in #2000
• feat: add gemma-3-12b by @k223kim in #2002
• [3/4] feat: add gemma 3 4b by @k223kim in #2001
• Add resume for adapter_v2, enable continued finetuning for adapter by @altria-zewei-wang in #1354
• Fix/loading gemma 3 1b by @pquadri in #2004
• feat: add gemma 3 in readme and tutorials by @k223kim in #2005
• add borda as codeowner by @t-vi in #2007
• example for full finetuning with python code by @astrobdr in #1331
• feat: add tests for gemma3 by @k223kim in #2006
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2009
• building tutorials as mkdocs pages by @Borda in #2011
• Add mlflow logger support by @topikachu in #1985
• fix support for litserve>0.2.4 by @ali-alshaar7 in #1994
• Cast tensors in KVCache only when needed by @Andrei-Aksionov in #2017
• feat: load only text weights from multimodal gemma by @pquadri in #2008
• Feature: Adds support for OpenAISpec in litgpt serve by @bhimrazy in #1943
• fix typo by @Lynsoo in #2018
• drop upper bounds in dependencies by @t-vi in #2022
• prepare 0.5.8 by @t-vi in #2023

New Contributors

• @syntheticgio made their first contribution in #1939
• @deependujha made their first contribution in #1975
• @altria-zewei-wang made their first contribution in #1354
• @pquadri made their first contribution in #2004
• @astrobdr made their first contribution in #1331
• @pre-commit-ci made their first contribution in #2009
• @topikachu made their first contribution in #1985
• @bhimrazy made their first contribution in #1943
• @Lynsoo made their first contribution in #2018

Full Changelog: v0.5.7...v0.5.8