Commits

Branch selector

 main

User selector

All users

Datepicker

All time

Commit History

Commits on Oct 27, 2024

• Fix custom fwd and bwd for older PyTorch versions (#608)

  Show description for 2017c98
  KokeCacao

  authored
  2017c98

  Copy full SHA for 2017c98

Commits on Oct 25, 2024

• Fix custom fwd and bwd for older PyTorch versions (#596)

  Show description for 83a5c90
  vasqu

  authored
  83a5c90

  Copy full SHA for 83a5c90

Commits on Oct 13, 2024

• Fixed nasty warnings, after torch.cuda.amp went to torch.amp (#584)

  Show description for bc84fb1
  

  fkrasnov2

  and

  fkrasnov

  authored
  bc84fb1

  Copy full SHA for bc84fb1

Commits on Oct 7, 2024

• Fixes typo when indexing after self.conv1d (#580)

  Show description for 6b72c12
  tdgfrost

  authored
  6b72c12

  Copy full SHA for 6b72c12

Commits on Sep 26, 2024

• Fix exploding gradients when ngroups larger than one (#547)

  Show description for 3b0dde5
  

  ilyasch2

  and

  Ilyas.Chahed

  authored
  3b0dde5

  Copy full SHA for 3b0dde5

• fixing softplus bug with _chunk_cumsum_bwd_kernel() triton kernel (#574)

  stephen-youn

  authored
  9259852

  Copy full SHA for 9259852

Commits on Aug 6, 2024

• add support for pytorch 2.4.0 with cuda 12.4 (#504)

  winglian

  authored
  62db608

  Copy full SHA for 62db608

• make properly pip installable without needing build isolation (uv pip install now works too) (#402)

  Show description for 323db26
  

  kszlim

  and

  tridao

  authored
  323db26

  Copy full SHA for 323db26

• fix unpacking error (#507)

  Show description for a07ff1b
  chiennv2000

  authored
  a07ff1b

  Copy full SHA for a07ff1b

Commits on Aug 1, 2024

• Fix residual in MLP when not fused_add_norm

  tridao

  committed
  49ddf83

  Copy full SHA for 49ddf83

Commits on Jul 22, 2024

• Fix bug in mamba2_simple slow path conv1d #393

  albertfgu

  committed
  a71bb5a

  Copy full SHA for a71bb5a

Commits on Jul 16, 2024

• bug fix for test of state update with heads (#472)

  zigzagcai

  authored
  c0a00bd

  Copy full SHA for c0a00bd

• repeat kv for GQA (#462)

  lchu6

  authored
  014c094

  Copy full SHA for 014c094

• Add mixin (#469)

  NielsRogge

  authored
  7fb78a5

  Copy full SHA for 7fb78a5

Commits on Jul 3, 2024

• Fix varlen generation by passing seq_idx to causal_conv1d

  tridao

  committed
  8ffd905

  Copy full SHA for 8ffd905

Commits on Jul 1, 2024

• Fix MambaChunkScanCombinedFn returning incorrect number of gradients

  tridao

  committed
  ddce0c1

  Copy full SHA for ddce0c1

• Bump to v2.2.0

  tridao

  committed
  95a2041

  Copy full SHA for 95a2041

• Drop support for pytorch 1.12 and 1.13

  tridao

  committed
  6e8dfb2

  Copy full SHA for 6e8dfb2

• Implement varlen generation

  tridao

  committed
  03a38fb

  Copy full SHA for 03a38fb

Commits on Jun 30, 2024

• Add tests for layernorm_gated and ssm_update with heads

  tridao

  committed
  3462302

  Copy full SHA for 3462302

• Add swiglu activation

  tridao

  committed
  b85b20d

  Copy full SHA for b85b20d

• Fix ROCm version parsing (#416)

  Show description for c8396a3
  knw01

  authored
  c8396a3

  Copy full SHA for c8396a3

• radeon fix - compile with same launch params as instinct (#419)

  ajassani

  authored
  22c0c87

  Copy full SHA for 22c0c87

Commits on Jun 19, 2024

• ROCm warp size fix [AMD official] (#405)

  Show description for c809b2b
  

  amoskvic

  and

  ajassani

  authored
  c809b2b

  Copy full SHA for c809b2b

Commits on Jun 18, 2024

• Bump to v2.1.0

  tridao

  committed
  c71f86c

  Copy full SHA for c71f86c

• Compile for pytorch 2.4.0.dev20240420

  tridao

  committed
  059030f

  Copy full SHA for 059030f

• Don't need torch.cuda.is_available() to compile

  tridao

  committed
  5ca74b6

  Copy full SHA for 5ca74b6

• Change launch warp size back for CUDA

  tridao

  committed
  e61f105

  Copy full SHA for e61f105

• Move prerequisite

  tridao

  committed
  bed7e5d

  Copy full SHA for bed7e5d

• Force self.A_log to be fp32

  tridao

  committed
  1e6aa5e

  Copy full SHA for 1e6aa5e

• Adding ROCm support [AMD official] (#359)

  Show description for 3c77dcf
  

  7 peopleauthored
  3c77dcf

  Copy full SHA for 3c77dcf

• modify mamba triton kernels compatible to triton version >= 3.0.0 (#377)

  stephen-youn

  authored
  c2568f5

  Copy full SHA for c2568f5

Commits on Jun 12, 2024

• Bump to v2.0.4

  tridao

  committed
  f9dbb4f

  Copy full SHA for f9dbb4f

• Fix mask in _chunk_scan_chunk_state_bwd_dx that could cause NaN

  Show description for bb3a82a
  tridao

  committed
  bb3a82a

  Copy full SHA for bb3a82a

Commits on Jun 11, 2024

• Fix typo in min_p sampling (#385)

  epicfilemcnulty

  authored
  8f42a5e

  Copy full SHA for 8f42a5e

Pagination

Previous

Next