Commits

Branch selector

 main

User selector

tridao

All time

Commit History

Commits on Jan 6, 2025

• Fix license in pyproject.toml to Apache

  tridao

  committed
  9182c93

  Copy full SHA for 9182c93

Commits on Dec 6, 2024

• Bump to v2.2.4

  tridao

  committed
  95d8aba

  Copy full SHA for 95d8aba

• Don't have triton as dependency

  tridao

  committed
  bee0ae2

  Copy full SHA for bee0ae2

• [CI] Use torch 2.6.0.dev20241001 instead of 20241010 for nvcr 24.11

  tridao

  committed
  5fb943f

  Copy full SHA for 5fb943f

• Add back info in setup.py to upload to PyPI

  tridao

  committed
  b7cdf7b

  Copy full SHA for b7cdf7b

• Bump to v2.2.3

  tridao

  committed
  6763edb

  Copy full SHA for 6763edb

• [CI] Drop python 3.8, torch 2.0, add python 3.13, torch 2.5 & 2.6dev

  tridao

  committed
  ef99f92

  Copy full SHA for ef99f92

• Use absolute import

  tridao

  committed
  fb30432

  Copy full SHA for fb30432

• Reduce torch #include

  tridao

  committed
  b29375c

  Copy full SHA for b29375c

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 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

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

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 3, 2024

• Rename k_activation.py -> k_activations.py

  tridao

  committed
  28f623d

  Copy full SHA for 28f623d

• Add k_activation.py

  tridao

  committed
  f40e2ee

  Copy full SHA for f40e2ee

• Fix imports

  tridao

  committed
  33aeaa0

  Copy full SHA for 33aeaa0

• Mamba-2 code release

  tridao

  committed
  60dadf2

  Copy full SHA for 60dadf2

• Use lm-evaluation-harness from PyPI instead of our submodule

  tridao

  committed
  c59255a

  Copy full SHA for c59255a

Commits on May 26, 2024

• Bump to v1.2.2

  tridao

  committed
  a07faff

  Copy full SHA for a07faff

• [CI] Compile for pytorch 2.4.0.dev20240407 (for nvcr 24.05)

  tridao

  committed
  7fc139e

  Copy full SHA for 7fc139e

• Bump to v1.2.1

  tridao

  committed
  7efe513

  Copy full SHA for 7efe513

• [CI] Compile for pytorch 2.2.2 and 2.3.0

  tridao

  committed
  0e8f8c3

  Copy full SHA for 0e8f8c3

Commits on May 18, 2024

• Support nheads and ngroups in inference kernel

  tridao

  committed
  28b1435

  Copy full SHA for 28b1435