Commits · 4b88d69de76f4e876d71665f48392b4c12e48867 · OpenDAS / bitsandbytes

09 Jul, 2023 1 commit
- Added abitrary data types; fixed a bug for small matrices. · 4b88d69d
  Tim Dettmers authored Jul 09, 2023
  
  4b88d69d
05 Jul, 2023 1 commit
- Added bfloat16 quantizations and tests. · 02fd80cb
  Tim Dettmers authored Jul 04, 2023
  
  02fd80cb
04 Jul, 2023 2 commits
- Vectorized loads, conflict free NF4; 52 vs 172. · dfe6900b
  Tim Dettmers authored Jul 04, 2023
  
  dfe6900b
- Initial 4-bit naive batch size 1, 81 vs 185. · f89ff93e
  Tim Dettmers authored Jul 03, 2023
  
  f89ff93e
24 May, 2023 1 commit
- Added PagedLion and bf16 Lion. · 1b8772a8
  Tim Dettmers authored May 23, 2023
  
  1b8772a8
06 May, 2023 1 commit
- Added paging. · ec38ba95
  Tim Dettmers authored May 06, 2023
  
  ec38ba95
02 May, 2023 3 commits
- 4-bit draft; 128 vector load 240. · 264a9485
  Tim Dettmers authored May 02, 2023
  
  264a9485
- Shared memory efficient 240. · 77f15fdc
  Tim Dettmers authored May 02, 2023
  
  77f15fdc
- Baseline for debugging. · f9bfea8f
  Tim Dettmers authored May 02, 2023
  
  f9bfea8f
01 May, 2023 6 commits
- 8x32 240 6 warps. · 7bfa09d0
  Tim Dettmers authored May 01, 2023
  
  7bfa09d0
- 16x16 240. · 3d4a2ead
  Tim Dettmers authored May 01, 2023
  
  3d4a2ead
- Warp specalization 362. · 7cc8ff47
  Tim Dettmers authored May 01, 2023
  
  7cc8ff47
- 64 threads, high smem, 434. · 30d03e02
  Tim Dettmers authored Apr 30, 2023
  
  30d03e02
- Slow non-vector 530. · 604bb3fb
  Tim Dettmers authored Apr 30, 2023
  
  604bb3fb
- Slow tensor core solution. · ad07d254
  Tim Dettmers authored Apr 30, 2023
  
  ad07d254
30 Apr, 2023 1 commit
- 4-bit draft. · 21723f79
  Tim Dettmers authored Apr 29, 2023
  
  21723f79
29 Apr, 2023 5 commits
- Added bit template. · cad83994
  Tim Dettmers authored Apr 28, 2023
  
  cad83994
- New implementation for batch size 1. · f3e97ccb
  Tim Dettmers authored Apr 28, 2023
  
  f3e97ccb
- Added fp16 and thread/item template. · f6df4aef
  Tim Dettmers authored Apr 28, 2023
  
  f6df4aef
- Added template refactor. · 3aef7834
  Tim Dettmers authored Apr 28, 2023
  
  3aef7834
- First baseline kernel. · c1bfb210
  Tim Dettmers authored Apr 28, 2023
  
  c1bfb210
27 Apr, 2023 3 commits
- Adedd pipeline draft. · 9cab14a3
  Tim Dettmers authored Apr 27, 2023
  
  9cab14a3
- Added non-cutlass template. · d1c4c205
  Tim Dettmers authored Apr 27, 2023
  
  d1c4c205
- Best attempt at cutlass3. · 0afc8e9e
  Tim Dettmers authored Apr 26, 2023
  
  0afc8e9e
26 Apr, 2023 1 commit
- CUTLASS compiles. · 84964db9
  Tim Dettmers authored Apr 25, 2023
  
  84964db9
25 Apr, 2023 2 commits
- Added cutlass example. · 6e2544da
  Tim Dettmers authored Apr 25, 2023
  
  6e2544da
- Initial template. · 6bfd7a40
  Tim Dettmers authored Apr 25, 2023
  
  6bfd7a40
18 Apr, 2023 1 commit
- Added 32-bit optimizer for bfloat16 gradients. · 7dc198fe
  Tim Dettmers authored Apr 17, 2023
  
  7dc198fe
02 Apr, 2023 1 commit
- First draft of NF4. · 64cc0592
  Tim Dettmers authored Apr 02, 2023
  
  64cc0592
01 Apr, 2023 1 commit
- Added bf16 Adam. · c4cfe4fb
  Tim Dettmers authored Apr 01, 2023
  
  c4cfe4fb
30 Mar, 2023 1 commit
- Added normal quant. · 8645d1f7
  Tim Dettmers authored Mar 29, 2023
  
  8645d1f7
27 Mar, 2023 1 commit
- Some small changes. · 69810521
  Tim Dettmers authored Mar 27, 2023
  
  69810521
10 Mar, 2023 3 commits
- always pass beta2 into all the 1state functions · 6c377b39
  Phil Wang authored Mar 10, 2023
  
  6c377b39
- do the epsilon beta2 switcharoo within the cuda code, and not within the... · c99b44f7
  Phil Wang authored Mar 10, 2023
```
do the epsilon beta2 switcharoo within the cuda code, and not within the python class (so that the state dict still makes sense)
```
  c99b44f7
- swap the order in which momentum and parameters are updated in ops.cu · 8618bed0
  Phil Wang authored Mar 10, 2023
  
  8618bed0
09 Mar, 2023 1 commit
- do a bunch of typical bookkeeping before getting to main lion logic · cb4c3c8c
  Phil Wang authored Mar 09, 2023
  
  cb4c3c8c
14 Feb, 2023 1 commit
- Added more blocksizes for stochastic rounding; fixed dequant blocksize. · 2489d819
  Tim Dettmers authored Feb 14, 2023
  
  2489d819
04 Feb, 2023 1 commit
- Added fp4 quant/dequant and dequant optimizations. · 3ac5840c
  Tim Dettmers authored Feb 04, 2023
  
  3ac5840c
20 Nov, 2022 1 commit
- Added additional blocksizes: {64, 128, 256}. · c059bd28
  Tim Dettmers authored Nov 20, 2022
  
  c059bd28
07 Nov, 2022 1 commit
- Added blocksizes 2048, 1024, and 512 to blockwise quant. · 6bc2b992
  Tim Dettmers authored Nov 06, 2022
  
  6bc2b992