Code

Updated to newer base to support qlora#2

Code

Open

stoperro wants to merge 65 commits intoacpopescu:cmake_windowsacpopescu/bitsandbytes:cmake_windowsfrom stoperro:cmake_windowsstoperro/bitsandbytes_windows:cmake_windowsCopy head branch name to clipboard

+2,957-1,000

Commits

Commits on Feb 4, 2023

• Added fp4 quant/dequant and dequant optimizations.

  TimDettmers

  committed
  3ac5840

  Copy full SHA for 3ac5840

Commits on Feb 5, 2023

• Forward matmul_fp4 tests pass.

  TimDettmers

  committed
  160a835

  Copy full SHA for 160a835

• Backward matmul_fp4 passes.

  TimDettmers

  committed
  13c0a4d

  Copy full SHA for 13c0a4d

• Added matmul_fp4 to the benchmark.

  TimDettmers

  committed
  cfe4705

  Copy full SHA for cfe4705

• Fixed matmul_fp4 transpose.

  TimDettmers

  committed
  c361f84

  Copy full SHA for c361f84

• Added bias test for LinearFP4 and basic test.

  TimDettmers

  committed
  c0c352b

  Copy full SHA for c0c352b

• Added backprop test for Linear8bitLt and LinearFP4.

  TimDettmers

  committed
  7f0773a

  Copy full SHA for 7f0773a

Commits on Feb 14, 2023

• Fixed FP4 import and data type conversion in backward.

  TimDettmers

  committed
  c93a90d

  Copy full SHA for c93a90d

Commits on Feb 24, 2023

• Added cast to fp4 layer for speed.

  TimDettmers

  committed
  9851a10

  Copy full SHA for 9851a10

Commits on Feb 27, 2023

• t5 model fix

  artidoro

  committed
  6c31a5f

  Copy full SHA for 6c31a5f

Commits on Mar 27, 2023

• Some small changes.

  TimDettmers

  committed
  6981052

  Copy full SHA for 6981052

Commits on Mar 30, 2023

• Added normal quant.

  TimDettmers

  committed
  8645d1f

  Copy full SHA for 8645d1f

Commits on Apr 1, 2023

• Added bf16 Adam.

  TimDettmers

  committed
  c4cfe4f

  Copy full SHA for c4cfe4f

• Added 8-bit compression to quantization statistics.

  TimDettmers

  committed
  51a21df

  Copy full SHA for 51a21df

Commits on Apr 2, 2023

• Generalized FP4 data type.

  TimDettmers

  committed
  2dd5d69

  Copy full SHA for 2dd5d69

• Added normal with extra value.

  TimDettmers

  committed
  0d332a6

  Copy full SHA for 0d332a6

• Added quantization tree generation.

  TimDettmers

  committed
  4ad999d

  Copy full SHA for 4ad999d

• First draft of NF4.

  TimDettmers

  committed
  64cc059

  Copy full SHA for 64cc059

Commits on Apr 3, 2023

• Refactor FP4 into 4Bit and integrate NF4 data type.

  TimDettmers

  committed
  4ea489d

  Copy full SHA for 4ea489d

Commits on Apr 4, 2023

• Fixed ParamsIn4 init; fixed PyTorch 2.0 test failure.

  TimDettmers

  committed
  1ccb7bd

  Copy full SHA for 1ccb7bd

Commits on Apr 7, 2023

• Some fixed for loading PEFT modules with Params4bit.

  TimDettmers

  committed
  e9fa03b

  Copy full SHA for e9fa03b

Commits on Apr 12, 2023

• Fixed bias conversion in Linear4bit

  TimDettmers

  committed
  b8ea2b4

  Copy full SHA for b8ea2b4

Commits on Apr 18, 2023

• Added 32-bit optimizer for bfloat16 gradients.

  TimDettmers

  committed
  7dc198f

  Copy full SHA for 7dc198f

Commits on Apr 19, 2023

• Added nested quantization for blockwise quantization.

  TimDettmers

  committed
  0f9d302

  Copy full SHA for 0f9d302

Commits on Apr 25, 2023

• Initial template.

  TimDettmers

  committed
  6bfd7a4

  Copy full SHA for 6bfd7a4

• Added cutlass example.

  TimDettmers

  committed
  6e2544d

  Copy full SHA for 6e2544d

Commits on Apr 26, 2023

• CUTLASS compiles.

  TimDettmers

  committed
  84964db

  Copy full SHA for 84964db

Commits on Apr 27, 2023

• Best attempt at cutlass3.

  TimDettmers

  committed
  0afc8e9

  Copy full SHA for 0afc8e9

• Added non-cutlass template.

  TimDettmers

  committed
  d1c4c20

  Copy full SHA for d1c4c20

• Adedd pipeline draft.

  TimDettmers

  committed
  9cab14a

  Copy full SHA for 9cab14a

Commits on Apr 29, 2023

• First baseline kernel.

  TimDettmers

  committed
  c1bfb21

  Copy full SHA for c1bfb21

• Added template refactor.

  TimDettmers

  committed
  3aef783

  Copy full SHA for 3aef783

• Added fp16 and thread/item template.

  TimDettmers

  committed
  f6df4ae

  Copy full SHA for f6df4ae

• New implementation for batch size 1.

  TimDettmers

  committed
  f3e97cc

  Copy full SHA for f3e97cc

• Added bit template.

  TimDettmers

  committed
  cad8399

  Copy full SHA for cad8399

Commits on Apr 30, 2023

• 4-bit draft.

  TimDettmers

  committed
  21723f7

  Copy full SHA for 21723f7

Commits on May 1, 2023

• Slow tensor core solution.

  TimDettmers

  committed
  ad07d25

  Copy full SHA for ad07d25

• Slow non-vector 530.

  TimDettmers

  committed
  604bb3f

  Copy full SHA for 604bb3f

• Double frag 440.

  TimDettmers

  committed
  c35ed09

  Copy full SHA for c35ed09

• Fixed bank conflicts in non-vector load 422.

  TimDettmers

  committed
  e01d4e0

  Copy full SHA for e01d4e0

• 64 threads, high smem, 434.

  TimDettmers

  committed
  30d03e0

  Copy full SHA for 30d03e0

• Halved shared memory 466.

  TimDettmers

  committed
  cabcd9b

  Copy full SHA for cabcd9b

• Warp specalization 362.

  TimDettmers

  committed
  7cc8ff4

  Copy full SHA for 7cc8ff4

• 16x16 240.

  TimDettmers

  committed
  3d4a2ea

  Copy full SHA for 3d4a2ea

• 8x32 240 6 warps.

  TimDettmers

  committed
  7bfa09d

  Copy full SHA for 7bfa09d

Commits on May 2, 2023

• Baseline for debugging.

  TimDettmers

  committed
  f9bfea8

  Copy full SHA for f9bfea8

• Tighter and scaled error analysis.

  TimDettmers

  committed
  9192c9d

  Copy full SHA for 9192c9d

• Initial.

  TimDettmers

  committed
  9aa232c

  Copy full SHA for 9aa232c

• Correct implementation 240.

  TimDettmers

  committed
  394749d

  Copy full SHA for 394749d

• Removed uncessary sync.

  TimDettmers

  committed
  4decb3c

  Copy full SHA for 4decb3c

• A tile multi-tiling.

  TimDettmers

  committed
  89cccd8

  Copy full SHA for 89cccd8

• Shared memory efficient 240.

  TimDettmers

  committed
  77f15fd

  Copy full SHA for 77f15fd

• Warp multi-specialization 240.

  TimDettmers

  committed
  869b7e8

  Copy full SHA for 869b7e8

• 4-bit draft; 128 vector load 240.

  TimDettmers

  committed
  264a948

  Copy full SHA for 264a948

Commits on May 6, 2023

• Added paging.

  TimDettmers

  committed
  ec38ba9

  Copy full SHA for ec38ba9

• Added paged optimizers.

  TimDettmers

  committed
  44d68ff

  Copy full SHA for 44d68ff

Commits on May 7, 2023

• Changed prefetching.

  TimDettmers

  committed
  41a9c70

  Copy full SHA for 41a9c70

• Fixed prefetch bug for non-paged tensors; added benchmark.

  TimDettmers

  committed
  f64cfe6

  Copy full SHA for f64cfe6

• Merge remote-tracking branch 'origin/main' into merge

  TimDettmers

  committed
  675baa7

  Copy full SHA for 675baa7

• Fixed gradient accumulation test.

  TimDettmers

  committed
  4bd1151

  Copy full SHA for 4bd1151

Commits on May 24, 2023

• Fixed Makefile.

  TimDettmers

  committed
  2bce175

  Copy full SHA for 2bce175

• Added PagedLion and bf16 Lion.

  TimDettmers

  committed
  1b8772a

  Copy full SHA for 1b8772a

• Bumped version.

  TimDettmers

  committed
  0f40fa3

  Copy full SHA for 0f40fa3

Commits on May 25, 2023

• Merge remote-tracking branch 'remotes/source/main' into cmake_windows

  Show description for 46b184b
  stoperro

  committed
  46b184b

  Copy full SHA for 46b184b

Commits on May 29, 2023

• Disables cudaMemPrefetchAsync() in unsupported systems (e.g. Windows). Fixes artidoro/qlora#73 and bitsandbytes-foundation#453

  stoperro

  committed
  e02f078

  Copy full SHA for e02f078