Commits · a24aae30bf0cb912515981192e82fbe5a32e3317 · OpenDAS / bitsandbytes

12 Apr, 2023 2 commits
- Refactored simulated fp8 modules into research.nn. · dd562c24
  Tim Dettmers authored Apr 12, 2023
  
  dd562c24
- Refactored triton into its own folder. Refactored fp8 matmuls. · ec1ea637
  Tim Dettmers authored Apr 12, 2023
  
  ec1ea637
03 Apr, 2023 1 commit
- Refactor FP4 into 4Bit and integrate NF4 data type. · 4ea489d3
  Tim Dettmers authored Apr 03, 2023
  
  4ea489d3
01 Apr, 2023 1 commit
- Added 8-bit compression to quantization statistics. · 51a21df7
  Tim Dettmers authored Apr 01, 2023
  
  51a21df7
23 Feb, 2023 1 commit
- Added matmul_mixed. · c5c38ca1
  Tim Dettmers authored Feb 23, 2023
  
  c5c38ca1
14 Feb, 2023 2 commits
- Added forward/backward tests; removed bias. · ca323658
  Tim Dettmers authored Feb 13, 2023
  
  ca323658
- Added fp8 simulation layer. · 6bdb6c35
  Tim Dettmers authored Feb 13, 2023
  
  6bdb6c35
05 Feb, 2023 4 commits
- Fixed matmul_fp4 transpose. · c361f842
  Tim Dettmers authored Feb 05, 2023
  
  c361f842
- Added matmul_fp4 to the benchmark. · cfe4705e
  Tim Dettmers authored Feb 04, 2023
  
  cfe4705e
- Backward matmul_fp4 passes. · 13c0a4dc
  Tim Dettmers authored Feb 04, 2023
  
  13c0a4dc
- Forward matmul_fp4 tests pass. · 160a8358
  Tim Dettmers authored Feb 04, 2023
  
  160a8358
27 Oct, 2022 3 commits
- Sort imports · 7a3c9af0
  Tom Aarsen authored Oct 27, 2022
```
Via isort
```
  7a3c9af0
- Simplify statements into equivalent, modern variants · 0b078403
  Tom Aarsen authored Oct 27, 2022
```
via pyupgrade --py37-plus. The changes e.g. are subclassing from object, calling super() with super(ThisClass, self), or old-style syntax formatting.
```
  0b078403
- Remove trailing whitespace & ensure newline at EOF · 1eec77d3
  Tom Aarsen authored Oct 27, 2022
  
  1eec77d3
20 Sep, 2022 1 commit
- try fp32 · 76ce9aa6
  justheuristic authored Sep 20, 2022
  
  76ce9aa6
17 Sep, 2022 10 commits
- bump threshold to 0.21 · f6670329
  justheuristic authored Sep 18, 2022
  
  f6670329
- more lenient threshold · fa8e07c7
  justheuristic authored Sep 18, 2022
  
  fa8e07c7
- cast properly · e35e2c66
  justheuristic authored Sep 18, 2022
  
  e35e2c66
- un-fuse bias · d9ca0ed9
  justheuristic authored Sep 17, 2022
  
  d9ca0ed9
- copypaste tolerances · 7facedda
  justheuristic authored Sep 17, 2022
  
  7facedda
- clearer assertions · e29c5f5c
  justheuristic authored Sep 17, 2022
  
  e29c5f5c
- check dtypes first · 9379df85
  justheuristic authored Sep 17, 2022
  
  9379df85
- check dtypes first · 140cdbe8
  justheuristic authored Sep 17, 2022
  
  140cdbe8
- cast to half before double_quant · a9c7953e
  justheuristic authored Sep 17, 2022
  
  a9c7953e
- test_bf16 · 469d5a63
  justheuristic authored Sep 17, 2022
  
  469d5a63
23 Aug, 2022 1 commit
- Some initial code. Needs to be tested. · 7e0fb655
  Tim Dettmers authored Aug 23, 2022
  
  7e0fb655
16 Aug, 2022 1 commit
- Added fused bias to matmullt. · de354f7d
  Tim Dettmers authored Aug 16, 2022
  
  de354f7d
04 Aug, 2022 1 commit
- Fixed bugs in cuda setup. · 8f84674d
  Tim Dettmers authored Aug 04, 2022
  
  8f84674d
03 Aug, 2022 1 commit
- Added fixes for the case that matmullt dim A is zero, e.g. [0, 768]. · 451fd950
  Tim Dettmers authored Aug 03, 2022
  
  451fd950
01 Aug, 2022 2 commits
- reran black with linelength 80 for greater readability · ea7c14f8
  Titus von Koeller authored Aug 01, 2022
  
  ea7c14f8
- ran black and isort for coherent code formatting · bfa0e332
  Titus von Koeller authored Aug 01, 2022
  
  bfa0e332
22 Jul, 2022 1 commit
- Most tests passing. · c771b3a7
  Tim Dettmers authored Jul 22, 2022
  
  c771b3a7