All pull requests

Pull requests

• chore(deps): bump the github-actions group across 1 directory with 2 updates

  dependenciesgithub_actionsin-progressiris
  #654·dependabot[bot] opened Nov 12, 2026Bot·

  ·

• fix: set _layout_ = "ms" on packed ctypes structs (#634)

  #653·anish-01 opened Nov 11, 2026·1 task done·

  ·
  1

  13

• Send triggered plans over RDMA: dormant work released by a copy-engine kick

  in-progressiris
  #652·BKP opened Nov 11, 2026Collaborator·

  ·
  6

• Add engines=("rdma",): a multi-host Iris context with GPU-initiated RDMA markdown

  in-progressiris
  #650·BKP opened Nov 11, 2026Collaborator·2/2·

  ·
  6

• Add device-side GPU-initiated RDMA for mlx5 (iris.device.rdma_utils)

  in-progressiris
  #649·BKP opened Nov 11, 2026Collaborator·1/2·

  ·
  10

• Add triggered scatter, gather, broadcast and all-to-all to ctx.ccl

  in-progressiris
  #647·BKP opened Nov 11, 2026Collaborator·2/2·

  ·
  6

• Add lanes: independent host copy-engine queues per peer

  in-progressiris
  #646·BKP opened Nov 11, 2026Collaborator·1/2·

  ·
  9

• Barrier before every CCL collective, not only after

  in-progressiris
  #645·nirvedhmeshram opened Nov 11, 2026Collaborator·

  ·
  17

• [experimental] Wide-EP dispatch/combine across fabric domains over RCCL GIN

  #644·nirvedhmeshram opened Nov 10, 2026Collaborator·

  ·
  10

• Cache get_num_xcc so hipDeviceGetAttribute runs once per device

  #638·jesswsong opened Nov 10, 2026·

  ·
  1

  6

• Declare _layout_ = "ms" on packed ctypes structs to silence the Python 3.14 deprecation

  #635·cdytotoro-ctrl opened Nov 7, 2026·

  ·
  1

  1

• Fix all_gather_matmul_hbm_buffer: partial M groups, small K, int64 offsets

  in-progressiris
  #632·dsidler opened Nov 5, 2026Collaborator·

  ·

• Hoist heap-base loads out of the Triton all-gather store loops

  in-progressiris
  #626·dsidler opened Nov 5, 2026Collaborator·

  ·
  5

• Optimizing two-shot GEMM-AllReduce

  #610·dsidler opened Oct 27, 2026Collaborator·

  ·

• Fix workspace allocation and reuse in matmul_all_reduce

  #581·MashrafeeAryan opened Oct 11, 2026·

  ·
  1

• Add a TDM all-gather and fix gfx1250 enablement

  in-progressiris
  #566·mawad-amd opened Oct 2, 2026Collaborator·

  ·

• Add iris.concurrent (concurrent, independent, fused comm+gemm)

  #544·ryanswann-amd opened Jul 23, 2026Collaborator·1 task·

  ·
  8

• MegaKernels in Iris 🔥

  in-progressiris
  #541·neoblizz opened Jun 24, 2026Member·

  ·
  6

• Support memory sharing in simulation environments

  in-progressiris
  #539·mawad-amd opened Jun 18, 2026Collaborator·1 task done·

  ·
  4

• Add more tests

  #537·micmelesse opened Jun 17, 2026Contributor·1 task·

  ·
  19

• Add vLLM Llama-70B all-reduce reproducer

  #535·micmelesse opened Jun 16, 2026Contributor·1 task·

  ·
  4

• Add graph-capturable one-shot all-reduce

  in-progressiris
  #531·mawad-amd opened May 5, 2026Collaborator·1 task done·

  ·

• Fix matmul_reduce_scatter wrapper passing unsupported bias argument

  in-progressiris
  #530·mawad-amd opened May 2, 2026Collaborator·2 tasks done·

  ·
  3

• Fix hardcoded lock value in fused GEMM+CCL operations

  in-progressiris
  #529·mawad-amd opened May 2, 2026Collaborator·3 tasks done·

  ·
  1

  7

• Cache hip.get_num_xcc() to eliminate 14ms Config overhead

  in-progressiris
  #528·mawad-amd opened May 2, 2026Collaborator·2 of 3 tasks·

  ·
  10