Should we use master branch or stable version?

[fzyzcjy]

What piece of documentation is affected?

Hi thanks for the library! Usually it is suggested to use stable version of a package (e.g. "release v0.5.2"), but it seems the README here uses the master branch, and the docker image is also mainly about latest master. Therefore, I wonder whether it is OK to use the master branch? i.e. will it be too unstable?

What part(s) of the article would you like to see updated?

(see above)

Additional Information

No response

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this feature has not been requested yet.
• I have provided enough information for the maintainers to understand and evaluate this request.

[winglian]

here's the diff of main vs the last release. There are mostly a lot of test fixes and updating to use some of the more recent upstream dependencies. We should be cutting a new release on Monday after I get some more testing updates done this weekend. Depending on what features you need most of the changes since 0.5.2 are likely to not affect you.

• release 0.5.2 by @winglian in release 0.5.2 #2086
• use pep440 instead of semver by @winglian in use pep440 instead of semver #2088
• Fix duplication of plugin callbacks by @chiragjn in Fix duplication of plugin callbacks #2090
• fix inference when no chat_template is set, fix unsloth dora check by @winglian in fix inference when no chat_template is set, fix unsloth dora check #2092
• Enable Ascend NPU support by @MengqingCao in Enable Ascend NPU support #1758
• Bump liger-kernel requirements to 0.4.2 by @bursteratom in Bump liger-kernel requirements to 0.4.2 #2096
• fix None-type not iterable error when deepspeed is left blank w/ use_… by @bursteratom in fix None-type not iterable error when deepspeed is left blank w/ use_… #2087
• actions/create-release is unmaintained, and doesn't create proper release notes by @winglian in actions/create-release is unmaintained, and doesn't create proper release notes #2098
• updated colab notebook by @bursteratom in updated colab notebook #2074
• .gitignore additions by @tmm1 in .gitignore additions #349
• move shared pytest conftest to top level tests by @winglian in move shared pytest conftest to top level tests #2099
• add finetome dataset to fixtures, check eval_loss in test by @winglian in add finetome dataset to fixtures, check eval_loss in test #2106
• fix: ds3 and fsdp lmbench eval by @NanoCode012 in fix: ds3 and fsdp lmbench eval #2102
• support seperate lr for embeddings, similar to loraplus by @winglian in support seperate lr for embeddings, similar to loraplus #1910
• add e2e tests for Unsloth qlora and test the builds by @winglian in add e2e tests for Unsloth qlora and test the builds #2093
• Add Exact Deduplication Feature to Preprocessing Pipeline by @olivermolenschot in Add Exact Deduplication Feature to Preprocessing Pipeline #2072
• various tests fixes for flakey tests by @winglian in various tests fixes for flakey tests #2110
• build causal_conv1d and mamba-ssm into the base image by @winglian in build causal_conv1d and mamba-ssm into the base image #2113
• make the eval size smaller for the resume test by @winglian in make the eval size smaller for the resume test #2111
• use pytest sugar and verbose for more info during ci by @winglian in use pytest sugar and verbose for more info during ci #2112
• Check torch version for ADOPT optimizer + integrating new ADOPT updates by @bursteratom in Check torch version for ADOPT optimizer + integrating new ADOPT updates #2104
• fix(vlm): handle legacy conversation data format and check image in data by @NanoCode012 in fix(vlm): handle legacy conversation data format and check image in data #2018
• Add ds model card, rebased by @bursteratom in Add ds model card, rebased #2101
• fix so inference can be run against quantized models without adapters by @winglian in fix so inference can be run against quantized models without adapters #1834
• fix merge conflict of duplicate max_steps in config for relora by @winglian in fix merge conflict of duplicate max_steps in config for relora #2116
• feat: add cut_cross_entropy by @NanoCode012 in feat: add cut_cross_entropy #2091
• fix(readme): update cuda instructions during preprocess by @NanoCode012 in fix(readme): update cuda instructions during preprocess #2114
• fix optimizer reset for relora sft by @winglian in fix optimizer reset for relora sft #1414
• prepare plugins needs to happen so registration can occur to build the plugin args by @winglian in prepare plugins needs to happen so registration can occur to build the plugin args #2119
• add missing fixture decorator for predownload dataset by @winglian in add missing fixture decorator for predownload dataset #2117
• replace tensorboard checks with helper function by @winglian in replace tensorboard checks with helper function #2120
• drop unnecessary BNB_CUDA_VERSION env var from docker as it just results in warnings by @winglian in drop unnecessary BNB_CUDA_VERSION env var from docker as it just results in warnings #2121
• update fix_untrained_tokens from unsloth with additional fixes by @winglian in update fix_untrained_tokens from unsloth with additional fixes #2122
• cleanup the readme, add Modal as sponsor by @winglian in cleanup the readme, add Modal as sponsor #2130
• fix license header for fix_untrained_tokens from unsloth-zoo by @winglian in fix license header for fix_untrained_tokens from unsloth-zoo #2129
• CLI Implementation with Click by @djsaunde in CLI Implementation with Click #2107
• auto-versioning and adding axolotl.version by @djsaunde in auto-versioning and adding axolotl.__version__ #2127
• remove accidentally included symlink by @winglian in remove accidentally included symlink #2131
• upgrade bnb 0.45.0 and peft by @winglian in upgrade bnb 0.45.0 and peft #2126
• Fix broken CLI; remove duplicate metadata from setup.py by @djsaunde in Fix broken CLI; remove duplicate metadata from setup.py #2136
• reduce test concurrency to avoid HF rate limiting, test suite parity by @winglian in reduce test concurrency to avoid HF rate limiting, test suite parity #2128
• Fix llama type model check by @chiragjn in Fix llama type model check #2142
• Transformers 4.47.0 by @winglian in Transformers 4.47.0 #2138

[fzyzcjy]

Thank you! I am mainly wondering a general question: should we use master or stable as a routine (not specific to this release)?

[djsaunde]

@fzyzcjy IMO, most users can use the latest stable version. If you want access to bleeding-edge features as they are added, then it might make sense to check out master, but as you noted, it may be less stable.

In short, in depends on how you intend to use the package!

[fzyzcjy]

Thank you! Then I guess the doc here can be updated to clone the latest stable instead of master.

[NanoCode012]

@fzyzcjy , thanks for that recommendation. We have an open PR to update that to install via pip for stable build instead #2137