Commits · c35ed09b668db43da967ddeff88c13d92a5cb02a · OpenDAS / bitsandbytes

01 May, 2023 2 commits
- Double frag 440. · c35ed09b
  Tim Dettmers authored Apr 30, 2023
  
  c35ed09b
- 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 4 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
- 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
19 Apr, 2023 1 commit
- Added nested quantization for blockwise quantization. · 0f9d3020
  Tim Dettmers authored Apr 19, 2023
  
  0f9d3020
04 Apr, 2023 1 commit
- Fixed ParamsIn4 init; fixed PyTorch 2.0 test failure. · 1ccb7bde
  Tim Dettmers authored Apr 03, 2023
  
  1ccb7bde
03 Apr, 2023 1 commit
- Refactor FP4 into 4Bit and integrate NF4 data type. · 4ea489d3
  Tim Dettmers authored Apr 03, 2023
  
  4ea489d3
02 Apr, 2023 4 commits
- First draft of NF4. · 64cc0592
  Tim Dettmers authored Apr 02, 2023
  
  64cc0592
- Added quantization tree generation. · 4ad999d1
  Tim Dettmers authored Apr 02, 2023
  
  4ad999d1
- Added normal with extra value. · 0d332a64
  Tim Dettmers authored Apr 02, 2023
  
  0d332a64
- Generalized FP4 data type. · 2dd5d690
  Tim Dettmers authored Apr 02, 2023
  
  2dd5d690
01 Apr, 2023 1 commit
- Added 8-bit compression to quantization statistics. · 51a21df7
  Tim Dettmers authored Apr 01, 2023
  
  51a21df7
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
05 Feb, 2023 3 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
- Forward matmul_fp4 tests pass. · 160a8358
  Tim Dettmers authored Feb 04, 2023
  
  160a8358
04 Feb, 2023 1 commit
- Added fp4 quant/dequant and dequant optimizations. · 3ac5840c
  Tim Dettmers authored Feb 04, 2023
  
  3ac5840c
19 Nov, 2022 1 commit
- Fixed k-bit quantization maps. · eb028e6e
  Tim Dettmers authored Nov 19, 2022
  
  eb028e6e
08 Nov, 2022 1 commit
- Fixed bug in cpu quant; faster GPU dequant. · 08fa2e7b
  Tim Dettmers authored Nov 07, 2022
  
  08fa2e7b
07 Nov, 2022 2 commits
- Fixed blockwise test and logic. · e0e697b1
  Tim Dettmers authored Nov 06, 2022
  
  e0e697b1
- Added blocksizes 2048, 1024, and 512 to blockwise quant. · 6bc2b992
  Tim Dettmers authored Nov 06, 2022
  
  6bc2b992
06 Nov, 2022 3 commits
- Added k<256 quantile estimate. · 2f2063ba
  Tim Dettmers authored Nov 06, 2022
  
  2f2063ba
- Added k-bit fp8 map. · 98cbc4bc
  Tim Dettmers authored Nov 06, 2022
  
  98cbc4bc
- Added k-bit linear quantization. · caf18325
  Tim Dettmers authored Nov 06, 2022
  
  caf18325
04 Nov, 2022 1 commit
- Added FP8 quantization map. · 1efb87d8
  Tim Dettmers authored Nov 03, 2022
  
  1efb87d8
27 Oct, 2022 1 commit

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

24 Oct, 2022 1 commit
- Isolated CUDASetup logging; all tests green. · df86625a
  Tim Dettmers authored Oct 24, 2022
  
  df86625a
13 Sep, 2022 1 commit
- Fixed cpu blockwise quantization for small input tensors. · c05dd42d
  Tim Dettmers authored Sep 13, 2022
  
  c05dd42d
11 Sep, 2022 1 commit
- Fixed 2^31 max size issue for cpu blockwise quant. · 19a7adca
  Tim Dettmers authored Sep 11, 2022
  
  19a7adca
23 Aug, 2022 1 commit
- Some initial code. Needs to be tested. · 7e0fb655
  Tim Dettmers authored Aug 23, 2022
  
  7e0fb655
16 Aug, 2022 2 commits
- Added fused bias in dequant_mm. · dede3430
  Tim Dettmers authored Aug 16, 2022
  
  dede3430
- Removed storage() from get_ptr; added boilerplate for bias dequant_mm. · 1ed2fa2f
  Tim Dettmers authored Aug 16, 2022
  
  1ed2fa2f
01 Aug, 2022 1 commit
- reran black with linelength 80 for greater readability · ea7c14f8
  Titus von Koeller authored Aug 01, 2022
  
  ea7c14f8