v0.3.4

What's Changed

• Bump setuptools from 68.2.2 to 70.0.0 by @dependabot in #81

New Contributors

• @dependabot made their first contribution in #81

Full Changelog: v0.3.3...v0.3.4

v0.3.3

What's Changed

• first draft of pypi action by @thecharlieblake in #79
• Switch to pyproject.toml by @thecharlieblake in #80

Full Changelog: v0.3.0...v0.3.3

v0.3.2

Full Changelog: v0.3.1...v0.3.2

v0.3.1

Full Changelog: v0.3.0...v0.3.1

v0.3.0

What's Changed

• Linear by @thecharlieblake in #1
• Add CI for public repo by @thecharlieblake in #2
• Docs by @thecharlieblake in #3
• Remove unneeded docker setup by @thecharlieblake in #4
• Add MLP and supporting fns by @thecharlieblake in #6
• Add analyse_module() utils function by @thecharlieblake in #7
• Add necessary IP requirements for open-source by @thecharlieblake in #8
• Implement self-attention and update utils by @thecharlieblake in #9
• Add transformer layer by @thecharlieblake in #10
• Full transformer by @thecharlieblake in #13
• User guide by @thecharlieblake in #14
• Remove poptorch/pea by @thecharlieblake in #15
• Fixes to run on python3.10 by @thecharlieblake in #17
• Alpha docs by @thecharlieblake in #18
• Add scaling fusion analysis and update docs by @thecharlieblake in #21
• Doc review by @JamesRandom in #20
• Add ops for demo by @thecharlieblake in #24
• Change constraints to strings by @thecharlieblake in #26
• Format simulation (ready for review) by @thecharlieblake in #25
• Add unit_scale transform by @thecharlieblake in #27
• Quantisation for IPU by @DouglasOrr in #28
• Fix setuptools packages by @thecharlieblake in #30
• Changes grad_w default scaling to be 0.75 by @thecharlieblake in #31
• Prepare for beta release by @thecharlieblake in #34
• Support sdist builds by @DouglasOrr in #33
• Fix a torch FX error on torch-2.0.1 'annotations must be set to a dict object' by @DouglasOrr in #35
• Update scaled_dot_product_attention scaling factor from our derivation. by @DouglasOrr in #36
• Add separate tau scaling by @thecharlieblake in #37
• Visualisation tool by @thecharlieblake in #32
• Fix setup dependencies by @thecharlieblake in #39
• Add compile transform by @thecharlieblake in #40
• Tidy plots by @thecharlieblake in #41
• Tidy docs for beta release by @thecharlieblake in #42
• Fix torch.nn functional issue in docs by @thecharlieblake in #43
• Update limitations and add link to notebook by @thecharlieblake in #44
• PyTorch 2.1 fixes by @thecharlieblake in #47
• Add blog draft for 'almost scaled dot product attention' by @DouglasOrr in #46
• [almost-scaled blog] Introduce terms {d_seq, t} by @DouglasOrr in #48
• Add dependencies and dataset download to almost_scaled_dot_product_attention blog by @DouglasOrr in #49
• Fix torch version at 2.1 for now by @thecharlieblake in #50
• Revert "Changes grad_w default scaling to be 0.75" by @thecharlieblake in #51
• Proposal to refactor a state-carrying closure to a class by @awf in #53
• Update to PyTorch 2.2 by @awf in #52
• Update to PyTorch 2.2 addendum: Fixing a doc typo and logic mismatch by @awf in #54
• Update license and numpy by @thecharlieblake in #55
• Parametrize the number of random bits used in stochastic rounding by @awf in #56
• Add a package version and unit_scaling.version (0.1) by @DouglasOrr in #59
• Fix recursion in torch_nn_modules_to_user_modules() by @DouglasOrr in #57
• Updates to support u-muP, as the new default behaviour by @DouglasOrr in #58
• Fix unit_scaling setup.py install (#60) by @DouglasOrr in #61
• Migrate blog post 'almost scaled dot product attention' to the graphcore blog by @DouglasOrr in #64
• Add initial how_to_scale_op notebook by @DouglasOrr in #65
• Conv1d by @thecharlieblake in #67
• Import Conv1d into top-level unit_scaling (fixes #69) by @DouglasOrr in #71
• Remove the constraint argument in the container modules uu.{MLP, MHSA} by @DouglasOrr in #72
• fix sgd bug and add demo notebook by @thecharlieblake in #74
• add umup slides by @thecharlieblake in #75
• fix paper link in docs by @thecharlieblake in #76
• upgrade transformers dependency (security) by @thecharlieblake in #77
• Add docker dev setup and update requirements by @thecharlieblake in #78

New Contributors

• @JamesRandom made their first contribution in #20
• @DouglasOrr made their first contribution in #28
• @awf made their first contribution in #53

Full Changelog: https://github.com/graphcore-research/unit-scaling/commits/v0.3.0