v1.10.0

What's Changed

• Revert-deepspeed by @alan-cooney in #189
• Make loading device agnostic by @alan-cooney in #190
• Simplify dataparallel approach by @alan-cooney in #191
• Update dependencies by @alan-cooney in #192
• Add sample GPT training run by @alan-cooney in #194

Full Changelog: v1.9.1...v1.10.0

v1.9.1

What's Changed

• Fix deepspeed for single GPU by @alan-cooney in #187

Full Changelog: v1.9.0...v1.9.1

v1.9.0

What's Changed

• Remove abstract resampler class by @alan-cooney in #183
• Add notebook auto-formatting with vscode by @alan-cooney in #185
• Add deepspeed support by @alan-cooney in #186

Full Changelog: v1.8.0...v1.9.0

v1.8.0

What's Changed

• Default to average metric aggregation across components by @alan-cooney in #180
• Fix number of components in the resampler by @alan-cooney in #181
• Add cosine annealing to LR scheduler options by @alan-cooney in #182

Full Changelog: v1.7.0...v1.8.0

v1.7.0

What's Changed

• Add join run cli command by @alan-cooney in #177

Full Changelog: v1.6.0...v1.7.0

v1.6.0

What's Changed

• Ensure that the adam reset indices are on the correct device by @alan-cooney in #174
• Add hugging face model save/load support by @alan-cooney in #175
• Add partial dataset support by @alan-cooney in #176
• Support training with DataParallel by @alan-cooney in #178

Full Changelog: v1.5.0...v1.6.0

v1.5.0

What's Changed

• Add save and load methods to the model by @alan-cooney in #172

Full Changelog: v1.4.0...v1.5.0

v1.4.0

What's Changed

• Add join sweep method by @alan-cooney in #173

Full Changelog: v1.3.0...v1.4.0

v1.3.0

What's Changed

• Add optional components dimension to SAE model by @alan-cooney in #137
• Change reconstruction score to use change in means. by @HoagyC in #157
• Change loss functions to support a components dimension by @alan-cooney in #158
• Add component dimension support to the optimizer by @alan-cooney in #159
• Make the activation store support multiple component dimensions by @alan-cooney in #160
• Update the metrics to have a components dimension by @alan-cooney in #162
• Make the pipeline support training multiple components by @alan-cooney in #163
• Add unit/int test separation by @alan-cooney in #164
• Add column name support for source datasets by @alan-cooney in #165
• Add context size checks to pre-tokenized datasets by @alan-cooney in #166
• Add pydantic validation by @alan-cooney in #167
• Change the store to require a component dimension by @alan-cooney in #168
• Simplify stateful pipeline iteration by @alan-cooney in #170
• Switch to using named tuples to improve error handling by @alan-cooney in #169
• Reduce model complexity by @alan-cooney in #171

Full Changelog: v1.2.1...v1.3.0

v1.2.1

What's Changed

• Fix activation resampler normalization dimension by @alan-cooney in #155

Full Changelog: v1.2.0...v1.2.1