April 2026 monthly summary for pytorch-labs/helion (Pallas/Helion backend). Deliveries focused on dynamic control flow, symbolic bounds, and enhanced grid-based tensor operations, underpinned by expanded test coverage and regression safety. Key features delivered include: (1) Dynamic control flow graph enhancements with IfGraphInfo/ElseGraphInfo, improved code generation for dynamic if-conditions, and added tests; unblocked critical tests such as test_grpo_loss_fwd, test_if_new_variable_in_static_range (Pallas TPU), and test_if_arg_indexed_scalar. (2) Symbolic loop bounds support in the Pallas backend, introducing a fori_loop path when loop bounds include non-constexpr symbolic values, with appropriate configuration checks and tests. (3) Tensor indexing, tiling, and grid access enhancements, including SMEM-based scalar indexing, grid() indexing fixes, tile-indexed access with offsets, explicit BlockSpecs for tiling, and alignment handling, backed by expanded test coverage. (4) Testing framework and grid testing improvements, enabling subset testing and additional tests for hl.grid indexing with offsets. These efforts collectively unblock regression scenarios, improve reliability, and expand hardware compatibility.

March 2026 monthly summary for pytorch/pytorch: Delivered Triton kernel fusion with unary epilogues for Inductor, extending TTIR analysis and adding a fusion scheduler node to manage fusion logic. This work reduces tensor allocations and improves runtime performance for fused kernels across user kernels.