Code

Support multi-QP for normal kernels#130

Code

Merged

LyricZhao merged 9 commits intomaindeepseek-ai/DeepEP:mainfrom

trmt/internode_multi_qpdeepseek-ai/DeepEP:trmt/internode_multi_qpCopy head branch name to clipboard

Apr 22, 2025

Commits

Commits on Apr 21, 2025

• In the Internode Normal Kernel, when using nvshmem ibrc for RDMA data transmission, a single QP is used for data transfer between two GPUs, which limits kernel performance in network card dual-port…

  Show description for 5ab80c2
  moningchen

  committed
  5ab80c2

  Copy full SHA for 5ab80c2

• Revert `ibgda_device.cuh` and remove some comments.

  sphish

  committed
  e2c5784

  Copy full SHA for e2c5784

• Add the performance data after internode optimization in the Readme file

  moningchen

  committed
  e0eaaf9

  Copy full SHA for e0eaaf9

• Merge branch 'trmt/internode_multi_qp' of github.com:deepseek-ai/DeepEP into trmt/internode_multi_qp

  moningchen

  committed
  c07fdd1

  Copy full SHA for c07fdd1

Commits on Apr 22, 2025

• Refactor some code.

  sphish

  committed
  20b2aaa

  Copy full SHA for 20b2aaa

• Normal kernels always use IBGDA mode.

  sphish

  committed
  3e54b78

  Copy full SHA for 3e54b78

• Several code lints

  LyricZhao

  committed
  edbb1bc

  Copy full SHA for edbb1bc

• Fix the performance data.

  sphish

  committed
  3b1045d

  Copy full SHA for 3b1045d

• Use `put_nbi_warp`.

  sphish

  committed
  e255d57

  Copy full SHA for e255d57