v2.7.0

What's Changed

• Kernel autotuners by @liuanji in #54
• Blockscale sampling by @liuanji in #55
• Fix illegal memory access for product nodes with more than 2048 children by @prateekanand2 in #56
• Blockscale z correction by @liuanji in #57
• Pc perf followups by @liuanji in #58
• Remove partial evaluation; fix conditional target_vars for DiscreteLogistic by @liuanji in #59
• Fix NaN from the small-batch CUDA sum forward when a child is -inf by @liuanji in #60
• Fix repeated and block-sparse sum edges: random constructor, dense views, set_params by @liuanji in #61
• feat(nodes)!: BlockScaleSumParams applies the z-correction (dual-flow EM) by default by @liuanji in #63
• Sum-layer precision option, and dense sum layers through cuBLAS by @liuanji in #62
• Exact constrained inference for PCs (pyjuice.constraints) by @liuanji in #67

New Contributors

• @prateekanand2 made their first contribution in #56

Full Changelog: v2.6.3...v2.7.0