v0.6.3.7

Release

odm v0.1.4: GPU accelerated ODM

What's Changed

• Added conditional GPU accelerated ODM by @romitjain in #167

Full Changelog: v0.6.3.6...v0.6.3.7

v0.6.3.6

Release

moe v0.4.5: patch version, lora + moe kernels / EP bug fix on converted safetensors

What's Changed

• fix: lora safe tensors should have the HF peft support format by @kmehant in #166

Full Changelog: v0.6.3.5...v0.6.3.6

v0.6.3.5

Release

mcp v0.1.2: patch version, CP support for mamba layer
odm v0.1.3: patch version, padding free + ODM support

What's Changed

• feat: CP support for mamba layer by @kmehant in #164
• fix: ODM support for padding free collator by @romitjain in #165

Full Changelog: v0.6.3.4...v0.6.3.5

v0.6.3.4

Release

moe v0.4.4: patch version, Some fix patches for FSDP2 + MoE Kernels + EP support
aadp v0.2.2: patch version, Support iterable datasets + padding free
foak v0.7.3: patch version, bug fix for _bitsandbytes_available
odm v0.1.2: patch version, auto-categorization support to run ODM without categories

What's Changed

• fix: incorrect patching of the fsdp2_prepare_model by @kmehant in #160
• fix: state dict patch and _bitsandbytes_available by @kmehant in #161
• fix: Using PaddingFree plugin with Iterable datasets and remove_columns by @dushyantbehl in #135
• feat: ODM without categories by @romitjain in #162

New Contributors

• @dushyantbehl made their first contribution in #135
• @romitjain made their first contribution in #162

Full Changelog: v0.6.3.3...v0.6.3.4

v0.6.3.3

Release

moe v0.4.3: patch version, Some fix patches for FSDP2 + MoE Kernels + EP support

What's Changed

• feat: add missing deps for accelerate patches by @kmehant in #159

Full Changelog: v0.6.3.2...v0.6.3.3

v0.6.3.2

Release

moe v0.4.2: patch version, Now you can use FSDP2 with MoE kernels and expert parallel.

What's Changed

• feat: FSDP2 with MoE kernels and expert parallel by @kmehant in #157

Full Changelog: v0.6.3.1...v0.6.3.2

v0.6.3.1

Release

framework v0.6.4: patch version, version bumped up to indicate patch update to odm plugin.
odm v0.1.1: patch version, now you can resume using odm dataset and sampling weights interface is modified.

What's Changed

• feat: Resume functionality for online data mixing by @kmehant in #155
• Fix: interface change for sampling weights by @kmehant in #156

Full Changelog: v0.6.3...v0.6.3.1

v0.6.3

Release

• framework v0.6.3: patch version, version bumped up to indicate new plugin addition.
• moe v0.4.1: patch version, patch FSDP state dict and warn the user instead of erroring out.
• odm v0.1.0: new plugin, Released online data mixing framework 🎉.

What's Changed

• refactor: patch sharding state dict and warn by @kmehant in #151
• feat: add online data mixing plugin by @kmehant @romitjain @seshapad in #152

New Contributors

• @seshapad made their first contribution in #152
• @romitjain made their first contribution in #152

Full Changelog: v0.6.2.2...v0.6.3

v0.6.2.2

Release

• foak 0.7.2: patch version, variable fixes for latest triton versions

What's Changed

• fix: update datatype to int by @YashasviChaurasia in #150

New Contributors

• @YashasviChaurasia made their first contribution in #150

Full Changelog: v0.6.2.1...v0.6.2.2

v0.6.2.1

Release

• foak 0.7.1: patch version, bound trl>=0.19.1,<0.20.0
• aadp 0.2.1: patch version, fix global variables in unsloth kernels to support latest triton versions

What's Changed

• fix: update global variable syntax since its not support from triton 3.2.0 by @kmehant in #148
• feat: remove owners file by @kmehant in #149

Full Changelog: v0.6.2...v0.6.2.1