Commits

Breadcrumbs

History for

onmain

User selector

All users

Datepicker

All time

Commit History

Commits on Mar 28, 2024

• Add the option for the macro and note (Dao-AILab#893)

  drisspg

  authored
  23e8fa5

  Copy full SHA for 23e8fa5

• Minor fix in compute_attn_1rowblock_splitkv (Dao-AILab#900)

  beginlner

  authored
  3e9414f

  Copy full SHA for 3e9414f

Commits on Mar 15, 2024

• Add in, macrosf for defining __grid_constant__ (Dao-AILab#852)

  drisspg

  authored
  4a73e90

  Copy full SHA for 4a73e90

• Enable paged attention in varlen forward (Dao-AILab#831)

  Show description for 2a15840
  sgrigory

  authored
  2a15840

  Copy full SHA for 2a15840

Commits on Mar 14, 2024

• Make nvcc threads configurable via environment variable (Dao-AILab#885)

  chiragjn

  authored
  50896ec

  Copy full SHA for 50896ec

Commits on Feb 21, 2024

• Enable headdim 256 backward on consumer GPUs (Ampere, Ada)

  tridao

  committed
  2406f28

  Copy full SHA for 2406f28

• Update Cutlass to v3.4.1

  tridao

  committed
  4d6b794

  Copy full SHA for 4d6b794

Commits on Feb 20, 2024

• Don't need to reduce row_sum during online softmax

  tridao

  committed
  b32efb1

  Copy full SHA for b32efb1

Commits on Feb 10, 2024

• Fix dv = torch::empty_like(k) for mha_bwd_varlen as well

  tridao

  committed
  d9a5cb2

  Copy full SHA for d9a5cb2

• fix backward for when query and key have different contiguity (Dao-AILab#818)

  bdhirsh

  authored
  2423cca

  Copy full SHA for 2423cca

Commits on Feb 8, 2024

• Fix Windows build (Dao-AILab#816)

  sgrigory

  authored
  4687936

  Copy full SHA for 4687936

Commits on Jan 30, 2024

• Preprocessor switches to control functionality (Dao-AILab#788)

  Show description for 0658e32
  bottler

  authored
  0658e32

  Copy full SHA for 0658e32

Commits on Jan 23, 2024

• Implement page KV cache

  Show description for 54e80a3
  

  tridao

  and

  beginlner

  committed
  54e80a3

  Copy full SHA for 54e80a3

• Use int64_t instead of uint32_t in kernel_traits.h

  tridao

  committed
  36bc29e

  Copy full SHA for 36bc29e

Commits on Jan 22, 2024

• Use int64_t instead of uint32_t for index_t

  tridao

  committed
  000b67f

  Copy full SHA for 000b67f

Commits on Jan 21, 2024

• Remove configure in bwd kernel launch

  tridao

  committed
  ea8a25c

  Copy full SHA for ea8a25c

• Add split-kv and M<->H swap to varlen forward decoding attention (Dao-AILab#754)

  Show description for af01244
  sgrigory

  authored
  af01244

  Copy full SHA for af01244

• Update cutlass to v3.4.0

  tridao

  committed
  8f4d82c

  Copy full SHA for 8f4d82c

• Move rotary device functions to a separate file

  tridao

  committed
  395e5a0

  Copy full SHA for 395e5a0

• Remove unused kernel_traits file

  tridao

  committed
  3e2c827

  Copy full SHA for 3e2c827

• Refactor masking in fwd pass into 1 object

  tridao

  committed
  66a127a

  Copy full SHA for 66a127a

• Change inline to __forceinline__, use __grid_constant__ param

  tridao

  committed
  ed4959b

  Copy full SHA for ed4959b

Commits on Jan 20, 2024

• Make Softmax an object

  tridao

  committed
  6f706ef

  Copy full SHA for 6f706ef

Commits on Jan 15, 2024

• Make Alibi an object

  tridao

  committed
  4ea866c

  Copy full SHA for 4ea866c

• Move bwd preprocess kernels to a separate file

  tridao

  committed
  5aca153

  Copy full SHA for 5aca153

Commits on Jan 14, 2024

• Move softmax_rescale_o to softmax.h

  tridao

  committed
  df1418f

  Copy full SHA for df1418f

• Move masking to a separate file (mask.h)

  tridao

  committed
  6777336

  Copy full SHA for 6777336

• Remove seqq_parallel backward kernel that's not used

  tridao

  committed
  9448264

  Copy full SHA for 9448264

• Move dropout to a separate file (dropout.h)

  tridao

  committed
  1274ec3

  Copy full SHA for 1274ec3

• apply_dropout now takes tensor of rowcol layout

  tridao

  committed
  10dad61

  Copy full SHA for 10dad61

Commits on Jan 13, 2024

• Remove dead code in philox.cuh

  tridao

  committed
  d9cbcfb

  Copy full SHA for d9cbcfb

• Simplify writing softmax to gmem

  tridao

  committed
  a7b66ae

  Copy full SHA for a7b66ae

Commits on Jan 12, 2024

• Simplify SmemLayoutVtransposed in kernel_traits.h

  tridao

  committed
  8d1b169

  Copy full SHA for 8d1b169

Commits on Jan 5, 2024

• [LayerNorm] Switch from CUDA to Triton implementation

  tridao

  committed
  abbc131

  Copy full SHA for abbc131

Commits on Dec 24, 2023

• Don't dispatch to local if window size >= seqlen_k

  tridao

  committed
  0842ec0

  Copy full SHA for 0842ec0

Pagination

Previous

Next