50 Pull requests merged by 2 people

• Exchange pairwise-concat optimisation pass

  #345 merged Mar 29, 2025

• Optimise CPU backend

  #344 merged Mar 29, 2025

• Make examples/extra/fusion.rs an example for bullet_core (removes needless dependencies)

  #343 merged Mar 28, 2025

• Improve IR output & display options

  #342 merged Mar 28, 2025

• Use viriformat from crates.io

  #341 merged Mar 27, 2025

• Propogate batching

  #340 merged Mar 27, 2025

• Refactor Activation into DiffableFromOutput and friends internally,

  #339 merged Mar 27, 2025

• Add some operations with scalars

  #338 merged Mar 26, 2025

• Clarify backend selection

  #337 merged Mar 25, 2025

• Restructure bullet_lib files

  #335 merged Mar 25, 2025

• Update gitignore

  #334 merged Mar 25, 2025

• Update sfbinpack version

  #333 merged Mar 24, 2025

• Update viriformat version

  #332 merged Mar 23, 2025

• Make i32 quant more consistent with others

  #331 merged Mar 19, 2025

• Marlinformat adapter + Fix edge case in linear combination backprop

  #330 merged Mar 18, 2025

• Print the id of the layer which failed to quantise

  #329 merged Mar 18, 2025

• Support multiple files in SfBinpackLoader

  #328 merged Mar 18, 2025

• Add support for chaining transforms to SavedFormat

  #327 merged Mar 17, 2025

• Improve upon previous commit

  #326 merged Mar 16, 2025

• Properly terminate threads in SfBinpackLoader

  #325 merged Mar 16, 2025

• Pass -arch=native to NVCC

  #324 merged Mar 16, 2025

• Catch device errors earlier

  #323 merged Mar 16, 2025

• Make workflows check examples

  #322 merged Mar 15, 2025

• Remove IR emission in advanced example

  #321 merged Mar 15, 2025

• GraphCompileArgs has fusion enabled by default, so add method to disable it

  #320 merged Mar 15, 2025

• Clarify output format using TrainerBuilder in training preamble

  #319 merged Mar 15, 2025

• Update docs & README, remove some commented out code

  #318 merged Mar 15, 2025

• Sparse affine dual fusion pass

  #317 merged Mar 15, 2025

• Add first automatic operator fusion passes

  #316 merged Mar 14, 2025

• Emit ir

  #315 merged Mar 14, 2025

• Move frontend into bullet_core, improve graph IR

  #314 merged Mar 14, 2025

• Support Viridithas binpack format

  #313 merged Mar 10, 2025

• Support compiling bullet_lib without a GPU

  #312 merged Mar 10, 2025

• CPU backend supports all core operations

  #311 merged Mar 10, 2025

• Add more CPU-to-device operation comparisons

  #310 merged Mar 9, 2025

• Improve pos/sec reporting

  #309 merged Mar 8, 2025

• Fix validation batches when batches per superbatch < 32

  #308 merged Mar 8, 2025

• Improve profiling + bugfix for large batch sizes

  #307 merged Mar 8, 2025

• Small sparse affine backprop optimisation

  #305 merged Mar 6, 2025

• Update simple.rs example

  #304 merged Mar 5, 2025

• Add back profiling of graph nodes

  #303 merged Mar 5, 2025

• Support larger batch sizes

  #302 merged Mar 5, 2025

• Optimise

  #301 merged Mar 5, 2025

• Start work on CPU backend, optimise a HIP function

  #300 merged Mar 2, 2025

• Remove unused file

  #299 merged Mar 1, 2025

• Restructure bullet_core

  #298 merged Mar 1, 2025

• Prep fusion

  #297 merged Mar 1, 2025

• .

  #296 merged Feb 28, 2025

• Allow batched biases in SparseAffine

  #295 merged Feb 28, 2025

• Explicitly include cstdint in HIP kernels

  #293 merged Feb 28, 2025

4 Issues closed by 1 person

• Problems running CUDA backend

  #336 closed Mar 25, 2025

• TrainerBuilder output doc should be updated

  #294 closed Mar 15, 2025

• Division by zero when doing validation with small batches/superbatch

  #306 closed Mar 8, 2025

• Optimise Sparse Matmuls

  #286 closed Feb 28, 2025

4 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Project Roadmap

  #252 commented on Mar 15, 2025 • 0 new comments

• Multi-Device support

  #262 commented on Mar 15, 2025 • 0 new comments

• New CUDA backend using `cudarc`

  #265 commented on Mar 13, 2025 • 0 new comments

• Release `bullet_core` v0.1.0

  #270 commented on Mar 13, 2025 • 0 new comments