All pull requests

Pull requests

• IB/MLX5: Enable relaxed ordering without DEVX

  #12112·jason136 opened Oct 17, 2026·

  ·
  15

• UCP/RNDV: Use local memory info for unsupported remote memory types

  #12111·yafshar opened Oct 16, 2026Contributor·

  ·
  6

• UCT/ZE/COPY: Log the device count limit at debug level

  #12110·yafshar opened Oct 16, 2026Contributor·

  ·
  5

• UCT/CUDA: Backport #11499 to v1.21.x (fix SIGSEGV in uct_cuda_base_queue_head_ready on flush descriptors)

  #12109·syangx39 opened Oct 16, 2026·

  ·
  1

• UCT/CUDA/CUDA_IPC: Release unused cache entry at rkey release.

  #12108·rakhmets opened Oct 16, 2026Contributor·

  ·
  6

• [MUSA]Feat: musa support

  #12107·superleo opened Oct 16, 2026·

  ·
  3

• UCP/ROCM: Remove rocm_gdr from transport alias and docs

  #12106·edgargabriel opened Oct 16, 2026Contributor·

  ·
  9

• UCS/MEMORY: Support explicit region registration when UCM events are disabled

  #12104·ZhenlongMa opened Oct 16, 2026Contributor·

  ·
  9

• UCP/CORE: Periodically probe path monitor lanes

  #12101·ZhenlongMa opened Oct 16, 2026Contributor·

  ·
  24

• UCP/WIREUP: add dedicated monitor lanes for load-balance paths

  #12100·ZhenlongMa opened Oct 16, 2026Contributor·

  ·
  10

• UCT/CUDA: Fix deadlock with concurrent cuInit

  #12098·samridh10exe opened Oct 15, 2026·

  ·
  1

  3

• UCP: Limit unified mode to 64 transport resources

  #12097·guy-ealey-morag opened Oct 15, 2026Contributor·

  ·
  2

• UCP/CORE: Disable gpu-nic assignment on incomplete gpu class

  #12096·guy-ealey-morag opened Oct 15, 2026Contributor·

  ·
  9

• UCT/CUDA_IPC: POC for simulating different domains on the same node

  #12095·amastbaum opened Oct 15, 2026Contributor·

  ·
  6

• CI: Jenkins integration testing (do not merge)

  WIP-DNM
  #12094·Alexey-Rivkin opened Oct 15, 2026Contributor·

  ·
  303

• UCS/UCP/UCT: Add explicit memtype registry for when memtype cache is disabled

  #12093·ZhenlongMa opened Oct 15, 2026Contributor·

  ·
  9

• UCT/CUDA: Prepare GDRCopy support for cudaMallocAsync

  #12092·nbellalou opened Oct 15, 2026Contributor·

  ·
  34

• UCT/ZE/COPY: Run copies on the device which owns the buffer

  #12090·yafshar opened Oct 14, 2026Contributor·

  ·
  1

• UCT/IB/MLX5: Fix AR support prints in SL selection logging

  #12089·william-gallagher-nv opened Oct 14, 2026Contributor·

  ·
  8

• UCT/IB/MLX5: Keep polling after a TX completion for a removed endpoint.

  #12087·evgeny-leksikov opened Oct 14, 2026Contributor·

  ·
  15

• TEST/UCP: Enable the token failover tests.

  #12086·evgeny-leksikov opened Oct 14, 2026Contributor·

  ·
  15

• UCS/BITMAP: Add static check for bitmap loop index size

  #12080·guy-ealey-morag opened Oct 13, 2026Contributor·

  ·
  5

• UCT/ROCM: Add removal of remote mapping when owning buffer is freed to drop IPC

  #12077·ddebonis-amd opened Oct 12, 2026Contributor·

  ·
  56

• GTEST: Restore and check the current CUDA device after each test

  #12073·guy-ealey-morag opened Oct 12, 2026Contributor·

  ·
  5

• UCT/CUDA_IPC: Restrict MNNVL reachability to the same NVLink domain

  WIP-DNM
  #12072·amastbaum opened Oct 12, 2026Contributor·

  ·
  11