All pull requests

Pull requests

• DeepEP V2.5

  #763·LyricZhao opened Sep 23, 2026·

  ·
  35

• Fix hybrid combine top-k weight gradients with single reduction

  #762·Gstdioh opened Sep 23, 2026·

  ·
  16

• docs: document mlx5 requirement for legacy IBGDA

  #761·dvd233 opened Sep 22, 2026·

  ·
  1

  8

• Fix pinned host count handling in unpermute

  #758·Autumn1998 opened Sep 21, 2026·

  ·
  7

• docs: stop pointing the PCIe usage guide at a removed test file

  #755·PerryLink opened Sep 19, 2026·

  ·
  10

• docs: point the test instructions at the files that actually exist

  #754·PerryLink opened Sep 19, 2026·

  ·
  2

• DeepEP v2: Kernel-side deterministic dispatch for the non-expand layout

  #753·EricWang008 opened Sep 17, 2026·

  ·
  15

• [HybridEP] Add DMA-BUF fallback for GPU RDMA memory registration

  #750·janbernloehr opened Sep 12, 2026·

  ·
  11

• Fix unknown CMake command "cuda_add_library" in CMakeLists.txt

  #747·PerryLink opened Sep 10, 2026·

  ·
  2

• [HybridEP] Fix aarch64 inter-node JIT, CUDA graph capture, and test EP topology

  #745·QiaoK opened Sep 1, 2026·

  ·
  17

• Parse EP_BUFFER_DEBUG as an integer on the Python side

  #742·yurekami opened Aug 29, 2026·

  ·
  1

  3

• docs: point Development section to develop.sh (fix stale ln -s for V2)

  #741·pengfaming opened Aug 28, 2026·

  ·
  1

• fix: get_nvlink_gbs may return 0 as a divisor

  #740·cxxlink opened Aug 27, 2026·

  ·
  6

• fix: support system NCCL installations

  #738·janbernloehr opened Aug 25, 2026·

  ·

• fix: st.bulk support for size operand with 32-bit length is introduce…

  #737·cxxlink opened Aug 25, 2026·

  ·
  10

• fix: return default_value from get_env<int> on unparsable env values

  #736·shoemoney opened Aug 23, 2026·

  ·
  1

  2

• tests: bounded pressure loops and teardown barrier for test_ep

  #735·Xuan-1998 opened Aug 21, 2026·

  ·
  6

• bug fix: probe RDMA link rate via sysfs, survive probe failure

  #734·Xuan-1998 opened Aug 21, 2026·

  ·
  9

• Add unordered hybrid kernels for transports without ordered delivery (opt-in via EP_HYBRID_KERNEL)

  #732·Xuan-1998 opened Aug 21, 2026·

  ·
  11

• feat(low_latency): support top-16 routing and hidden-dim 3584

  #731·samsja opened Aug 18, 2026·

  ·

• [Runtime] Preflight full intranode P2P matrix

  #730·0z5a opened Aug 17, 2026·

  ·
  1

  15

• docs: fix cached dispatch outputs

  #729·bvolpato opened Aug 16, 2026·

  ·
  2

• utils: only reuse PyTorch's NCCL comm when _comm_ptr() is non-null

  #727·dmvevents opened Aug 14, 2026·

  ·
  1

  8

• Fix legacy normal kernels trapping when hidden exceeds the TMA staging buffer

  #723·yashkgp opened Aug 12, 2026·

  ·
  1

  4

• feat(normal): add optional diagnostic probes for DeepEP normal mode

  #720·Marsssqqq opened Aug 10, 2026·

  ·
  54