All pull requests

Pull requests

• docs: qualify bf16 reduce-scatter support and fix the Engram NCCL note

  #809·stwilliam503388-creator opened Oct 30, 2026·

  ·

• fix: compute SM/QP counts after handle unpacking in dispatch()

  #808·stwilliam503388-creator opened Oct 30, 2026·

  ·

• Fix shared event completion after deferred epilogues

  #806·ZenAlexa opened Oct 28, 2026·

  ·

• test: support PyTorch 2.10 collective APIs

  #805·bvolpato opened Oct 28, 2026·

  ·

• test(engram): release gathered references before profiling

  #804·bvolpato opened Oct 28, 2026·

  ·

• test: trim redundant profiler warmup callbacks

  #803·bvolpato opened Oct 28, 2026·

  ·

• test(ep): reuse invariant combine reference data

  #802·bvolpato opened Oct 28, 2026·

  ·

• test(pp): drop redundant 16 GiB benchmark allocation

  #801·bvolpato opened Oct 28, 2026·

  ·

• Support optional topk_weights in the reference dispatch

  #800·lsyx6399 opened Oct 27, 2026·

  ·

• Add EP_RDMA_GBS env to declare RDMA link bandwidth when ibstat under-reports it (fixes under-provisioned num_sms on CX8 800G)

  #797·sssssakuraaaaa opened Oct 26, 2026·

  ·

• fix(utils): enhance NCCL library directory resolution and environment robustness

  #796·reginaldalfret opened Oct 25, 2026·

  ·

• fix(build): stop helper scripts after failed commands

  #795·bvolpato opened Oct 25, 2026·

  ·

• Pack channel tail signals by runtime rank count (fix ~2x v2.5 hybrid dispatch regression)

  #794·sssssakuraaaaa opened Oct 23, 2026·

  ·

• Fix HybridEP metadata scan forward progress with cooperative launch

  #792·Autumn1998 opened Oct 23, 2026Collaborator·

  ·

• Make symmetric-window ordering configurable: opt into relaxed ordering to recover ~6x hybrid inter-node EP bandwidth

  #791·sssssakuraaaaa opened Oct 23, 2026·

  ·

• perf: reuse compact dispatch sort permutations for cached handles

  #787·yuweih205 opened Oct 17, 2026·

  ·

• perf: skip unused source-order sorts for expanded deterministic dispatch

  #786·yuweih205 opened Oct 16, 2026·

  ·

• Match GPUs by UUID in NVLink connection check

  #783·lvren1485 opened Oct 15, 2026·

  ·
  1

• Require bucket bandwidths only for hybrid groups

  #782·lvren1485 opened Oct 15, 2026·

  ·

• Document and test ordinary sync-free receive tensor lifetimes

  #780·yuweih205 opened Oct 15, 2026·

  ·

• fix: parse the full /proc/self/maps pathname in check_nccl_so

  #779·li-lizhe opened Oct 13, 2026·

  ·
  1

  2

• fix(build): Discover toolkit NVTX libraries and CCCL headers

  #775·Chlience opened Oct 9, 2026·

  ·

• Add EP_GIN_TYPE to select the NCCL GIN backend

  #773·Yiltan opened Oct 7, 2026Contributor·

  ·

• Add official PyTorch-based Dockerfile for DeepEP

  #771·YidaWeng opened Oct 1, 2026·

  ·

• Fix EP_BUFFER_DEBUG=0 enabling Python-side debug output

  #770·YidaWeng opened Oct 1, 2026·

  ·
  1

  1