v0.2.0

   Breaking Changes

• Extensible coords / dataflow architecture  -  by @caelyreth (972a4)
• Align sparse stack with MLX-native execution  -  by @caelyreth (2590b)
• Route pooling through relation execution plan  -  by @caelyreth (d3d9e)
• Add relation views for fast sparse backward  -  by @caelyreth (dfff0)
• Pin to mlx 0.31.1 to resolve cuda deps issue  -  by @caelyreth (48607)
• Say goodbye to mlx cuda, night mare  -  by @caelyreth (b4726)

   Features

• Coordinate primitive + baseline kernels  -  by @caelyreth (11b13)
• Feature execution primitive spmm_edges  -  by @caelyreth (f0602)
• MLX flavour lazy graph architecture  -  by @caelyreth (68ad8)
• Neighborhood relations baseline implementation  -  by @caelyreth (e90a7)
• Quantization and voxelization baseline implementation  -  by @caelyreth (b5e3f)
• Target convolution baseline support  -  by @caelyreth (1e23e)
• Fp16 baseline across current backends  -  by @caelyreth (cd262)
• Point-voxel utilities complement  -  by @caelyreth (d1f7f)
• Quantized convolution inference  -  by @caelyreth (c3865)
• conv:
  • Add explicit submanifold relation path  -  by @caelyreth (d7f9f)
• coords:
  • Add morton ordering primitive  -  by @caelyreth (d640f)
  • Add occupancy expansion helpers  -  by @caelyreth (e4bb2)
• cuda:
  • Baseline implementation of cuda operations  -  by @caelyreth (2bdf9)
  • Baseline optim for relation + pool + conv specialist  -  by @caelyreth (ef623)
  • precompiled_cuda_kernel + ptx for mlx  -  by @caelyreth (23c6c)
• entropy:
  • Add native range codec  -  by @caelyreth (506e5)
• native:
  • Edge pooling reductions  -  by @caelyreth (49dfd)
  • Autograd for sparse SpMM and edge pooling  -  by @caelyreth (fc7f3)
  • Re-introduce native conv3d Metal baseline  -  by @caelyreth (85908)
• py:
  • Setup sparse tensor API stub  -  by @caelyreth (07a59)
  • Canonical map layer with spec + view split  -  by @caelyreth (cdbe4)
  • Convolution public API stubs  -  by @caelyreth (fbb14)
  • Pooling public API  -  by @caelyreth (61532)
  • Sparse feature ops  -  by @caelyreth (ab259)
  • Sparse nn wrappers  -  by @caelyreth (9d19b)
• relations:
  • Add neighbor feature gather helper  -  by @caelyreth (0b680)
• tensor:
  • Add mask pruning helper  -  by @caelyreth (a78c9)
  • Add value-aligned sparse algebra  -  by @caelyreth (61f28)

   Bug Fixes

• cuda:
  • Native build gate  -  by @caelyreth (888a8)
  • Add compatible layer for cuda encoder co-mlx  -  by @caelyreth (62d44)
  • Binding + linking issue  -  by @caelyreth (ab8f9)
  • Route relations through precompiled runtime  -  by @caelyreth (883ea)
  • Correct jvp call site  -  by @caelyreth (46096)
  • Correct jvp metadata related processing  -  by @caelyreth (e142a)

   Performance

• Csr style relation operation + hash path  -  by @caelyreth (f2a28)
• Use input-row relation views for backward  -  by @caelyreth (4bb20)
• cpu:
  • Unify coord hash to fnv-1a  -  by @caelyreth (04ff2)
  • Align relation building, transposed direct path, and hash  -  by @caelyreth (9b651)
  • Specialist conv impls for cpu  -  by @caelyreth (3f468)
• entropy:
  • Add direct probability rans codec  -  by @caelyreth (03347)
• exec:
  • Add relation-backed sparse conv feature primitive  -  by @caelyreth (1725d)
• metal:
  • Native implementation of coords operations  -  by @caelyreth (b0ad0)
  • Add native fast paths for sparse conv and pool  -  by @caelyreth (a1c2f)
  • Parallelize sparse conv relation planning  -  by @caelyreth (6b886)
  • Native hash table for convolution  -  by @caelyreth (45a92)
  • Accelerate cold kernel relation construction  -  by @caelyreth (71403)
  • Parallel neighbor relation path  -  by @caelyreth (e01e0)
  • Non-overlapping pooling hash based downsample  -  by @caelyreth (5e2b6)
  • Parallel autodiff on pooling  -  by @caelyreth (e49d4)
  • Parallel downsample + stabel compaction  -  by @caelyreth (35a0b)
  • Unified coordinate pipeline optimize for union, intersection, downsample, quantize, voxelization  -  by @caelyreth (0f53d)
  • Edge-wise vjp kernels for avg/sum pool  -  by @caelyreth (5c2ea)
  • Reuse voxel quantization maps in fixed workloads  -  by @caelyreth (39f3c)
  • Vectorize sparse conv feature paths  -  by @caelyreth (41e40)
  • Compact kernel relations without regressing pool hot paths  -  by @caelyreth (d1c23)
  • Implicit gemm style conv path on large N  -  by @caelyreth (5324a)
  • Global dweight optim by sharing edge traversal, etc.  -  by @caelyreth (f5967)
  • Parallelize stable coordinate compaction  -  by @caelyreth (f4684)
  • Tile large sparse conv weight gradients  -  by @caelyreth (85b88)
  • Defer and specialize relation execution views  -  by @caelyreth (0de63)
  • Accelerate neighbor relation construction  -  by @caelyreth (997a7)
  • Skip unused small-kernel dweight views  -  by @caelyreth (45f1a)
  • Target convolution coordinate + relation optimized  -  by @caelyreth (81de7)
  • First vertical slice of tensor ops, conv/conv3d_generic_dweight  -  by @caelyreth (3dbb6)
  • Layout specific optimization for Cin=Cout=16  -  by @caelyreth (367c3)
  • Macro generated layout specialist for 32/64 conv  -  by @caelyreth (2e8d4)
  • Tensor-ops accelerated weight grad for backward  -  by @caelyreth (c25a2)
  • Switching dense forward from co4 to co16  -  by @caelyreth (f9f25)
  • Half4 vector load store for fp16 accumulation  -  by @caelyreth (ebc31)
  • Clamp dense conv row edge ranges before traversal  -  by @caelyreth (e75d6)
  • Input-grad grouping optimization  -  by @caelyreth (b5f12)
  • Non-square forward specialization  -  by @caelyreth (7db76)
  • Asymmetric backward specialist for conv  -  by @caelyreth (c8bf0)
  • Fp32 C64->C64 forward specialist  -  by @caelyreth (9aedf)
  • Shared parallel quantization/voxelization optim  -  by @caelyreth (957ed)
  • Add large-row contiguous C64 fp16 forward kernel  -  by @caelyreth (cae9c)
  • Tensor-ops accelerated cin=cout=64 forward path  -  by @caelyreth (9db5b)
  • Gate forward tensor-ops experiments  -  by @caelyreth (083a5)
  • Fp32 CI4/dot-product dense forward variant  -  by @caelyreth (0d98d)
  • Generalized fp32 CI4 subm forward route  -  by @caelyreth (58dfd)
  • Tensor ops implicit GEMM for conv forward  -  by @caelyreth (776da)
  • Kernel major relation tensor ops + true iGEMM  -  by @caelyreth (d5bae)
  • Accelerate non-square convolution weight gradients  -  by @caelyreth (4c834)
  • Implicit GEMM for quantized operation  -  by @caelyreth (36b57)
• native:
  • Native map view constructor  -  by @caelyreth (7250b)

    View changes on GitHub