Commits

Branch selector

 main

User selector

All users

Datepicker

All time

Commit History

Commits on Mar 2, 2024

• Bump to v2.5.6

  tridao

  committed
  6c9e60d

  Copy full SHA for 6c9e60d

• [CI] Change torch 2.3.0.dev20240126 to 20240105 for nvcr 24.02

  tridao

  committed
  6e2fa30

  Copy full SHA for 6e2fa30

Commits on Feb 21, 2024

• Bump to v2.5.5

  tridao

  committed
  87a1277

  Copy full SHA for 87a1277

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

  tridao

  committed
  2406f28

  Copy full SHA for 2406f28

• Bump to v2.5.4

  tridao

  committed
  43950dd

  Copy full SHA for 43950dd

• 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 18, 2024

• Optimize compile to 1: avoid oom 2: minimize swap usage 3: avoid threads starvation when letting ninja decide how many workers to spawn or manual MAX_JOBS "guesses". Logic is to take the min value …

  Show description for f45bbb4
  Qubitium

  authored
  f45bbb4

  Copy full SHA for f45bbb4

Commits on Feb 10, 2024

• Bump to v2.5.3

  tridao

  committed
  5cdabc2

  Copy full SHA for 5cdabc2

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

  tridao

  committed
  d9a5cb2

  Copy full SHA for d9a5cb2

• Add window_size option to ParallelMHA

  tridao

  committed
  a190df0

  Copy full SHA for a190df0

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

  bdhirsh

  authored
  2423cca

  Copy full SHA for 2423cca

Commits on Feb 8, 2024

• Fix Windows build (#816)

  sgrigory

  authored
  4687936

  Copy full SHA for 4687936

Commits on Jan 31, 2024

• Bump to v2.5.2

  tridao

  committed
  61a7772

  Copy full SHA for 61a7772

• [CI] Compile with torch 2.2.0 instead of 2.2.0.dev20231106

  tridao

  committed
  6a5c053

  Copy full SHA for 6a5c053

• Add window_size option to MHA and GPT

  tridao

  committed
  ef0ed10

  Copy full SHA for ef0ed10

Commits on Jan 30, 2024

• [CI] Install torch 2.3 using index

  tridao

  committed
  dc72d96

  Copy full SHA for dc72d96

• Bump to v2.5.1

  tridao

  committed
  daf37a9

  Copy full SHA for daf37a9

• [CI] Compile with pytorch 2.2.0.dev20231106

  tridao

  committed
  aa2eb8d

  Copy full SHA for aa2eb8d

• Preprocessor switches to control functionality (#788)

  Show description for 0658e32
  bottler

  authored
  0658e32

  Copy full SHA for 0658e32

• [CI] Build wheels for Pytorch 2.3 (dev/nightly) (#793)

  Show description for 290596c
  ckadner

  authored
  290596c

  Copy full SHA for 290596c

Commits on Jan 27, 2024

• Updated missing docstrings for args and returns in bert_padding.py (#795)

  Show description for c94cd09
  Avelina9X

  authored
  c94cd09

  Copy full SHA for c94cd09

Commits on Jan 24, 2024

• Add benchmarking code for Alibi (from Sanghun Cho)

  tridao

  committed
  ffc8682

  Copy full SHA for ffc8682

Commits on Jan 23, 2024

• Fixes an error in comment (#785)

  Show description for 204c3c6
  sighingnow

  authored
  204c3c6

  Copy full SHA for 204c3c6

• Bump to v2.5.0

  tridao

  committed
  197f208

  Copy full SHA for 197f208

• Implement page KV cache

  Show description for 54e80a3
  

  tridao

  and

  beginlner

  committed
  54e80a3

  Copy full SHA for 54e80a3

• [LayerNorm] Don't exit early in the backward pass (fix #781)

  tridao

  committed
  bdcae54

  Copy full SHA for bdcae54

• 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

• [CI] Fix CUDA 12.2.2 compilation

  tridao

  committed
  e43a4ce

  Copy full SHA for e43a4ce

• Bump to v2.4.3

  tridao

  committed
  f9d7376

  Copy full SHA for f9d7376

Commits on Jan 21, 2024

• [CI] Use CUDA 12.2.2 instead of 12.2.0

  tridao

  committed
  0399432

  Copy full SHA for 0399432

• 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 (#754)

  Show description for af01244
  sgrigory

  authored
  af01244

  Copy full SHA for af01244

• return z_loss (#768)

  cghawthorne

  authored
  d8aacc5

  Copy full SHA for d8aacc5

Pagination

Previous

Next