Commits · 0091f1e2ad2d96acde381d769df58f784f3dbad2 · OpenDAS / AutoAWQ

13 Sep, 2023 1 commit
- Support batch size benchmark, stop if OOM · 0091f1e2
  Casper Hansen authored Sep 13, 2023
  
  0091f1e2
12 Sep, 2023 8 commits
- Falcon 40B attention shapes [WIP] · 28d52d81
  Casper Hansen authored Sep 12, 2023
  
  28d52d81
- xq view key · bc4f93c2
  Casper Hansen authored Sep 12, 2023
  
  bc4f93c2
- Fix Falcon inputs ids · ba4da393
  Casper Hansen authored Sep 12, 2023
  
  ba4da393
- Falcon: Enable fused attention · f03f72de
  Casper Hansen authored Sep 12, 2023
  
  f03f72de
- Create custom attention shape for Falcon 7B · ac770875
  Casper Hansen authored Sep 12, 2023
  
  ac770875
- xk_reshape key · 7f8f9f16
  Casper Hansen authored Sep 12, 2023
  
  7f8f9f16
- Refactor view/reshaping into a predefined dict · a8c9afd5
  Casper Hansen authored Sep 12, 2023
  
  a8c9afd5
- Create Falcon block and model for fusing · 4517b3f2
  Casper Hansen authored Sep 12, 2023
  
  4517b3f2
11 Sep, 2023 9 commits
- Use CUDA stream · e120c9b6
  Casper Hansen authored Sep 11, 2023
  
  e120c9b6
- Remove xformers · 1aa8aebd
  Casper Hansen authored Sep 11, 2023
  
  1aa8aebd
- Update MPTBlock, fuse with MPTModel · dd41a223
  Casper Hansen authored Sep 11, 2023
  
  dd41a223
- Fix attention, support alibi · 7631add1
  Casper Hansen authored Sep 11, 2023
  
  7631add1
- Create MPTModel class · 90f54dbb
  Casper Hansen authored Sep 11, 2023
  
  90f54dbb
- Initialize with device · e80663bb
  Casper Hansen authored Sep 11, 2023
  
  e80663bb
- Remove fusing attention, only blocks · ac3e86df
  Casper Hansen authored Sep 11, 2023
  
  ac3e86df
- Fuse MPT block · 950851b3
  Casper Hansen authored Sep 11, 2023
  
  950851b3
- Switch to torch SDPA · d7badefc
  Casper Hansen authored Sep 11, 2023
  
  d7badefc
09 Sep, 2023 9 commits
- Falcon fused layers · 54f02854
  Casper Hansen authored Sep 09, 2023
  
  54f02854
- Fix alibi, rotary arg, neox, arg · 73c5e2bf
  Casper Hansen authored Sep 09, 2023
  
  73c5e2bf
- Add xformers requirement · 06bea896
  Casper Hansen authored Sep 09, 2023
  
  06bea896
- Fuse MPT · f3695d60
  Casper Hansen authored Sep 09, 2023
  
  f3695d60
- Update module name · 5bd6fbc7
  Casper Hansen authored Sep 09, 2023
  
  5bd6fbc7
- Implement ALiBi. · adc5304b
  Casper Hansen authored Sep 09, 2023
  
  adc5304b
- Default to GEMM · 48be2ee2
  Casper Hansen authored Sep 09, 2023
  
  48be2ee2
- Remove warnings about GEMM · c58ec73b
  Casper Hansen authored Sep 09, 2023
  
  c58ec73b
- Create torch SDPA implementation · ebbbc3a5
  Casper Hansen authored Sep 09, 2023
  
  ebbbc3a5
08 Sep, 2023 13 commits
- GEMM + GEMV compatibility · 890b6aa7
  Casper Hansen authored Sep 08, 2023
  
  890b6aa7
- More detailed benchmark · 5297eccc
  Casper Hansen authored Sep 08, 2023
  
  5297eccc
- New benchmark script. · 341c8868
  Casper Hansen authored Sep 08, 2023
  
  341c8868
- Use apply_rotary_emb · 64e6b3e1
  Casper Hansen authored Sep 08, 2023
  
  64e6b3e1
- Update kernel, remove unused code · a11c313a
  Casper Hansen authored Sep 08, 2023
  
  a11c313a
- Attention works · fbfa9d82
  Casper Hansen authored Sep 08, 2023
  
  fbfa9d82
- Fix MLP module · 75710806
  Casper Hansen authored Sep 08, 2023
  
  75710806
- Fix variables and no contiguous memory for GEMV · e197a733
  Casper Hansen authored Sep 08, 2023
  
  e197a733
- Default to GEMV · 331ff953
  Casper Hansen authored Sep 08, 2023
  
  331ff953
- Fix variables · 6534f5e6
  Casper Hansen authored Sep 08, 2023
  
  6534f5e6
- Implement GEMM/GEMV in quantize function and fused modules · 5db43a7f
  Casper Hansen authored Sep 08, 2023
  
  5db43a7f
- Add deprecation warning · 9b2946b6
  Casper Hansen authored Sep 08, 2023
  
  9b2946b6
- Refactor modules, create separate GEMM and GEMV · fe314160
  Casper Hansen authored Sep 08, 2023
  
  fe314160