v0.1.11

   Performance

• pool: Use row offsets for max pooling  -  by @utakotoba (6a2cc)

    View changes on GitHub

v0.1.10

   Features

• coords:
  • Add coordinate set ops  -  by @utakotoba (46f47)
  • Add explicit map reuse  -  by @utakotoba (97e68)
  • Add coordinate query utilities  -  by @utakotoba (67cd4)

    View changes on GitHub

v0.1.9

No significant changes

    View changes on GitHub

v0.1.7

   Features

• Full native code vertical slice  -  by @utakotoba (17d1b)
• Residual maps + submanifold convolution  -  by @utakotoba (0adbd)
• api:
  • Polish sparse tensor usability  -  by @utakotoba (e42c8)
• conv:
  • Add generative transpose conv  -  by @utakotoba (75612)
  • Add sparse conv padding support  -  by @utakotoba (1d8c3)
  • Add sparse conv dilation support  -  by @utakotoba (f2143)
  • Add generic sparse transpose conv  -  by @utakotoba (43ac1)
• core:
  • Add coordinate manager tensor ops  -  by @utakotoba (f701a)
• cuda:
  • Add gated sparse backend kernels  -  by @utakotoba (8cce2)
• metal:
  • Build submanifold coordinate maps  -  by @utakotoba (3bf35)
• native:
  • Add sparse conv gradients  -  by @utakotoba (b2557)
• nn:
  • Add sparse convolution modules  -  by @utakotoba (6d1d9)
  • Add sparse feature ops  -  by @utakotoba (16450)
• pool:
  • Add sparse max pooling  -  by @utakotoba (0914d)
  • Add sparse avg pooling  -  by @utakotoba (9ea01)
  • Add batch global pooling  -  by @utakotoba (735c7)
• sparse:
  • Add tensor API and native conv path  -  by @utakotoba (f71ca)

   Bug Fixes

• build: Rebuild editable native extension  -  by @utakotoba (b49b9)
• conv: Guard sparse conv residual fast path  -  by @utakotoba (968a9)

   Performance

• core: Keep topk row selection on mlx  -  by @utakotoba (c5221)
• metal: Hash submanifold coordinate maps  -  by @utakotoba (e99b1)
• ops: Optimize metal hot sparse kernels  -  by @utakotoba (59225)

    View changes on GitHub