All pull requests

Pull requests

• [cutedsl] Aligned-N physical tiles for 16-bit paired GEMMs

  CLA Signed
  #4330·nullplay opened Nov 7, 2026Contributor·

  ·

• [cutedsl] Price the plain pair's (128, 32) C subtile in the paired-pipeline SMEM model

  CLA Signed
  #4329·nullplay opened Nov 7, 2026Contributor·

  ·

• [cutedsl] Direct-entry default: six 64-deep AB stages instead of three 128-deep

  CLA Signed
  #4327·nullplay opened Nov 7, 2026Contributor·

  ·

• [CI] Benchmark MI350X on torch 2.15 with ROCm 10.0

  ciflow/rocmCLA Signedmodule: rocm
  #4326·yushangdi opened Nov 7, 2026Contributor·

  ·
  1

• [noland] Inductor epilogue corpus hillclimb plan and iteration log

  CLA Signed
  #4325·nullplay opened Nov 7, 2026Contributor·

  ·

• [cutedsl] Output-edge TMA-store admission by the SMEM arena model; two-CTA drain after dealloc

  CLA Signed
  #4324·nullplay opened Nov 7, 2026Contributor·

  ·

• [cutedsl] Autotune-off default tile for dense tcgen05 GEMMs

  CLA Signed
  #4323·nullplay opened Nov 7, 2026Contributor·

  ·

• [cutedsl] Persistent and CtaGroup.TWO search for output-edge GEMMs; grouped 2x2 twin seed

  CLA Signed
  #4322·nullplay opened Nov 7, 2026Contributor·

  ·

• [docs] Describe Helion's backends in the README and docs intro

  CLA Signed
  #4321·yushangdi opened Nov 6, 2026Contributor·

  ·
  2

• [cute] Re-stage a partial flex tile's first chunks after a full tile

  CLA Signed
  #4320·yushangdi opened Nov 6, 2026Contributor·

  ·

• [cute] Plan the flex hooks' shared rings over every region of the plan

  CLA Signed
  #4319·yushangdi opened Nov 6, 2026Contributor·

  ·

• improve H100 small BF16 matmul defaults

  CLA Signed
  #4300·coder-2011 opened Nov 6, 2026Contributor·

  ·

• [compiler] Support static filters in device list comprehensions

  CLA Signed
  #4299·coder-2011 opened Nov 6, 2026Contributor·

  ·

• [cute] Give the flex hook's cooperative stream its own KV tile binding

  CLA Signed
  #4298·yushangdi opened Nov 6, 2026Contributor·

  ·

• [cute] Keep score_mod_bwd's gradient literals as nodes in the Jacobian region

  CLA Signed
  #4296·yushangdi opened Nov 5, 2026Contributor·

  ·

• [torch.compile] Emit valid host calls for annotated and callable-arg kernels

  CLA Signed
  #4295·yushangdi opened Nov 5, 2026Contributor·

  ·
  1

• [compiler] Name a callable object by its type in host type annotations

  CLA Signed
  #4294·yushangdi opened Nov 5, 2026Contributor·

  ·

• [autotuner] Drain the HIP launch queue in do_bench's warm-up and timed loops

  CLA Signed
  #4293·yushangdi opened Nov 5, 2026Contributor·

  ·
  1

• [CI] Spread ROCm benchmark shards over the host's GPUs; keep a shard's results past a crashed kernel

  ciflow/rocmCLA Signedmodule: rocm
  #4287·yushangdi opened Nov 5, 2026Contributor·

  ·
  3

• Fix Helion autodiff unsqueeze after PyTorch mean_backward reorder (#4285)

  CLA Signedmeta-exported
  #4285·jansel opened Nov 5, 2026Contributor·

  ·
  1

• Make cute_dense_aligned_n expect the AB pipeline constructor the installed DSL selects (#4284)

  CLA Signedmeta-exported
  #4284·jansel opened Nov 5, 2026Contributor·

  ·
  1

• Pin split barrier init in grouped row-union cluster4 codegen tests (#4283)

  CLA Signedmeta-exported
  #4283·jansel opened Nov 5, 2026Contributor·

  ·
  1

• Accept DSL pipeline constructors in column-major output test (#4282)

  CLA Signedmeta-exported
  #4282·jansel opened Nov 5, 2026Contributor·

  ·
  1

• Support for out of tree (OOT) backends PR3: Let backends declare their own autotuner config keys.

  CLA Signed
  #4281·umechand-amd opened Nov 5, 2026Collaborator·

  ·

• Support for out of tree (OOT) backends PR3: Discover out-of-tree backends from the helion.backends entry points.

  CLA Signed
  #4280·umechand-amd opened Nov 5, 2026Collaborator·

  ·