All pull requests

Pull requests

• [autotuner] Report search-space cardinality for reduction loops

  CLA Signed
  #4423·bringlein opened Nov 12, 2026Collaborator·

  ·

• [autotuner] Emit the search-space report when autotuning fails

  CLA Signed
  #4422·bringlein opened Nov 12, 2026Collaborator·

  ·

• [benchmarks] Package fixed SDG profiles with ordinary Helion compilation

  CLA Signed
  #4421·jansel opened Nov 12, 2026Contributor·

  ·

• [autotuner] Refine compiler choices with exact search and cache identity

  CLA Signed
  #4420·jansel opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Lower conditional attention and expose configured schedules

  CLA Signed
  #4419·jansel opened Nov 12, 2026Contributor·

  ·

• [autotuner] Add rotating input banks and preserve valid finalist measurements

  CLA Signed
  #4418·jansel opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Schedule proved attention pairs with explicit capture overlap

  CLA Signed
  #4417·jansel opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Lower proved paged block-scaled online attention

  CLA Signed
  #4416·jansel opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Fuse exact S2F6 BF16 decoding on supported targets

  CLA Signed
  #4415·jansel opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Preserve computed MMA ownership and fragment lifetimes

  CLA Signed
  #4414·jansel opened Nov 12, 2026Contributor·

  ·

• [compiler] Preserve logical reductions, captures and graph effects

  CLA Signed
  #4413·jansel opened Nov 12, 2026Contributor·

  ·

• [compiler] Preserve host storage provenance across aliases and effects

  CLA Signed
  #4412·jansel opened Nov 12, 2026Contributor·

  ·

• [compiler] Preserve dynamic host geometry and guarded structural exports

  CLA Signed
  #4411·jansel opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Preserve typed operands and ownership across ordered MMA phases

  CLA Signed
  #4410·jansel opened Nov 12, 2026Contributor·

  ·

• [compiler] Preserve scalar and reduction semantics through lowering

  CLA Signed
  #4409·jansel opened Nov 12, 2026Contributor·

  ·

• [compiler] Preserve tunable, binding and config identity end to end

  CLA Signed
  #4408·jansel opened Nov 12, 2026Contributor·

  ·

• [compiler] Preserve typed input and scalar semantics across backends

  CLA Signed
  #4407·jansel opened Nov 12, 2026Contributor·

  ·

• [triton][gluon] Gluon backend + ring megakernel lowering

  CLA Signed
  #4406·ethche opened Nov 12, 2026Contributor·

  ·

• [cutedsl] Store aux-only outputs from the tcgen05 matmul epilogue

  CLA Signed
  #4405·nullplay opened Nov 11, 2026Contributor·

  ·

• [cutedsl] Size every tcgen05 epilogue store by the widest output dtype

  CLA Signed
  #4404·nullplay opened Nov 11, 2026Contributor·

  ·

• [autotuner] Let build_handoff accept an AccuracyCheck

  CLA Signed
  #4403·yushangdi opened Nov 11, 2026Contributor·

  ·

• torch_tpu: update pin to bfe8373b6e113ca1480119a71cb043449491db0c

  CLA Signed
  #4402·cota opened Nov 10, 2026Collaborator·

  ·

• [autotuner] Drop stale static_ranges hints after shape specialization

  #4395·AckerMan5659 opened Nov 10, 2026·

  ·
  1

  1

• [cutedsl] Stage unaligned output row stores through SMEM

  CLA Signed
  #4393·nullplay opened Nov 10, 2026Contributor·

  ·

• [cutedsl] Prefetch epilogue operands and add a third warp group over short K

  CLA Signed
  #4392·nullplay opened Nov 10, 2026Contributor·

  ·