v2.3.2.post1

What's Changed

• Bump version to 2.3.2 by @ko3n1g in #944
• Bump to 2.3.2.post1 by @ko3n1g in #945

Full Changelog: v2.3.2...v2.3.2.post1

v2.3.2

What's Changed

• Fix dtype mismatch in _chunk_scan_bwd_dstates_kernel Triton backward by @fmorgens in #855
• Mamba-3 code release by @aakashlahoti in #858
• fix: add missing init.py files by @pr-twig in #861
• Fix Mamba-3 SISO state shape, Amend README and pyproject.toml by @aakashlahoti in #866
• fix: add missing dependencies for Mamba3 (tilelang, nvidia-cutlass-dsl) by @Ollie-spoon in #863
• Fix dtype mismatch in mamba3_siso_bwd_kernel_dqkv Triton kernel by @tuxunlu in #869
• Revert "Fix dtype mismatch in mamba3_siso_bwd_kernel_dqkv Triton kernel" by @caitWW in #877
• Mamba-3 SISO: always typecast derived input tensors to bf16 by @aakashlahoti in #876
• Mamba-3: imports fail silently by @berlinchen7 in #878
• fixes issue #880 by @darxradi3nt in #881
• fixes issue #880 by @darxradi3nt in #882
• Feat: varlen support for Mamba-3 by @berlinchen7 in #890
• Bump quack-kernels to >=0.3.4 by @caitWW in #897
• Fix Mamba3 Step Fn deprecation warnings by @caitWW in #898
• Add maxnreg autotuning to Mamba-3 Triton kernels by @caitWW in #905
• Fix varlen support for Mamba-3 MIMO (angle cumsum) by @berlinchen7 in #911
• Updated dependency to address Mamba-3 MIMO dependency issue by @berlinchen7 in #918
• Remove duplicate dA assignment in mamba_chunk_scan by @Chessing234 in #921
• Set batch, seqlen, ngroups, and nheads as dynamic arguments for MIMO kernels. by @aakashlahoti in #937
• feat(ci): add weekly scheduled wheel build matrix by @ko3n1g in #941

New Contributors

• @fmorgens made their first contribution in #855
• @aakashlahoti made their first contribution in #858
• @pr-twig made their first contribution in #861
• @Ollie-spoon made their first contribution in #863
• @tuxunlu made their first contribution in #869
• @caitWW made their first contribution in #877
• @berlinchen7 made their first contribution in #878
• @darxradi3nt made their first contribution in #881
• @Chessing234 made their first contribution in #921

Full Changelog: v2.3.1...v2.3.2

v2.3.1

What's Changed

• ci: Free up containerd path by @ko3n1g in #847
• Fix ROCm 7.0+ compatibility: constexpr WARP_THREADS and lane_id mask for 64-wide wavefronts by @AndreasKaratzas in #831

New Contributors

• @AndreasKaratzas made their first contribution in #831

Full Changelog: v2.3.0...v2.3.1

v2.3.0

Merge pull request #840 from ko3n1g/ko3n1g/ci/build-arm-wheels

ci: Publish wheels also on failure

v2.2.6.post3

Bump to v2.2.6.post3

v2.2.6.post2

Bump to v2.2.6.post2

v2.2.6.post1

Bump to v2.2.6.post1

v2.2.5

Bump to v2.2.5

v2.2.4

Bump to v2.2.4

v2.2.3.post2

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