Commit Graph

  • e0eaaf94fb Add the performance data after internode optimization in the Readme file moningchen 2025-04-21 21:30:08 +08:00
  • e2c578485c Revert ibgda_device.cuh and remove some comments. Shangyan Zhou 2025-04-21 17:44:32 +08:00
  • 5ab80c28f3 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 and RoCE network scenarios. moningchen 2025-04-21 15:37:19 +08:00
  • a84a24808f Merge pull request #124 from wplf/patch-1 Shangyan Zhou 2025-04-16 10:57:31 +08:00
  • a2ccc95d78 Fix typo in nvshmem.patch 李金梁 2025-04-16 10:30:38 +08:00
  • a0c69317ab Merge pull request #118 from andylin-hao/main Chenggang Zhao 2025-04-14 15:51:30 +08:00
  • b9bb2bbaf6 Merge pull request #119 from phantom5125/patch-1 Shangyan Zhou 2025-04-14 09:29:46 +08:00
  • 42f617088f Fix typo in nvshmem.patch GreatHato 2025-04-13 00:14:44 +08:00
  • 23c54150ba Fix test combine args Hao Lin 2025-04-11 18:21:09 +08:00
  • 8a0ca8e2ec Merge pull request #116 from alpha-baby/fix-test-result-not-output Chenggang Zhao 2025-04-11 13:23:37 +08:00
  • 0f80da8458 fix: not output result in some linux system fujianhao.fjh 2025-04-10 18:18:30 +08:00
  • 42494864ba Remove useless control metadata for low-latency combine Chenggang Zhao 2025-04-07 09:55:39 +08:00
  • 2a0b3d7a5d Merge pull request #108 from fzyzcjy/patch-2 Chenggang Zhao 2025-04-03 12:18:28 +08:00
  • 218c5a1f96 Update buffer.py fzyzcjy 2025-04-03 10:57:45 +08:00
  • 26fa72d80f Fix zero-copy mode tests Chenggang Zhao 2025-03-28 16:49:33 +08:00
  • c4d12b4f8f Fix compilation Chenggang Zhao 2025-03-28 16:45:10 +08:00
  • dcf46f1c26 Merge pull request #96 from songhexiang/adjust_kNumThreads_of_notify_dispatch Chenggang Zhao 2025-03-28 16:42:21 +08:00
  • 4dd1e68ac8 For the SMs which calculate metadata in notify_dispatch, each warp in the SM is used to calculate the metadata of one channel. The default configuration is 8 warps for 10 channels, which needs two rounds of loop. Maybe the number of warps can be configured to the number of the channels so that one loop is enough. songhexiang 2025-03-28 06:43:29 +00:00
  • e130cc6e7d Remove NVLink low-latency plan Chenggang Zhao 2025-03-27 17:15:01 +08:00
  • cbd92fd0fc Update README Chenggang Zhao 2025-03-27 15:57:59 +08:00
  • ffc39ba084 Stronger acquire scope for low-latency kernels Chenggang Zhao 2025-03-27 09:30:36 +08:00
  • 7d52ad7248 Merge pull request #89 from fzyzcjy/patch-1 Chenggang Zhao 2025-03-25 09:28:44 +08:00
  • ae0eafd2be Remove confusing comments Chenggang Zhao 2025-03-25 09:27:34 +08:00
  • 36b5c27993 Update buffer.py fzyzcjy 2025-03-25 09:12:36 +08:00
  • c4b8ffc37c Merge pull request #79 from deepseek-ai/zero-copy-combine Chenggang Zhao 2025-03-18 15:46:45 +08:00
  • 66465476ae Support zero-copy for low-latency combine Chenggang Zhao 2025-03-18 15:44:26 +08:00
  • dcaf73e5ff Support zero-copy for low-latency combine Chenggang Zhao 2025-03-18 15:41:50 +08:00
  • 82dcf48fd3 Fix bugs for intranode EP kernels Chenggang Zhao 2025-03-14 16:09:23 +08:00
  • 043fa5fa99 Merge pull request #73 from deepseek-ai/p2p-signal Chenggang Zhao 2025-03-14 11:55:17 +08:00
  • 38cdaf390c Fix style. Shangyan Zhou 2025-03-14 11:22:00 +08:00
  • 2d0cf41dd1 Low latency kernels use rdma atomic to support AR. Shangyan Zhou 2025-03-14 11:04:57 +08:00
  • 7128ba3e39 Merge pull request #66 from dzhulgakov/combine-out-arg Chenggang Zhao 2025-03-13 09:18:06 +08:00
  • 50ac280ae7 comments Dmytro Dzhulgakov 2025-03-13 00:42:08 +00:00
  • 0008c6755e Merge pull request #67 from deepseek-ai/roce-support Chenggang Zhao 2025-03-11 09:30:45 +08:00
  • b3b61ef5ef Allow passing output tensor in low_latency_combine Dmytro Dzhulgakov 2025-03-10 22:19:21 +00:00
  • ed7487c15e Support BF16 for low-latency kernels Chenggang Zhao 2025-03-10 17:24:41 +08:00
  • 1fc40d50f3 Improve AR performance Chenggang Zhao 2025-03-06 21:41:19 +08:00
  • 41385ba5b3 Merge pull request #45 from deepseek-ai/ar-support Chenggang Zhao 2025-03-06 09:48:17 +08:00
  • 458cdcb22a Fix AR bugs for normal kernels Chenggang Zhao 2025-03-05 17:13:35 +08:00
  • e995aa22db Update NVSHMEM to v3.2.5. Shangyan Zhou 2025-03-05 16:16:52 +08:00
  • 680e424bdc Bugs fixed Chenggang Zhao 2025-03-05 14:27:45 +08:00
  • 592296cd45 Add some plans Chenggang Zhao 2025-03-04 15:54:46 +08:00
  • 1553fc42bf Improve EP2/4 performance Chenggang Zhao 2025-03-04 15:30:19 +08:00
  • 55cdd9a64f Fix typo Chenggang Zhao 2025-03-04 14:17:58 +08:00
  • 2a3cac903a Add some docs Chenggang Zhao 2025-03-04 10:19:42 +08:00
  • c5b4040502 Enable intranode kernel tests with EP2 and EP4 Chenggang Zhao 2025-03-03 15:01:02 +08:00
  • 6cc3497df8 Remove all raw tensors for better P2P overlapping Chenggang Zhao 2025-03-03 14:25:22 +08:00
  • f60306409a Merge pull request #32 from youkaichao/youkaichao-patch-1 Chenggang Zhao 2025-03-03 09:19:28 +08:00
  • 88b1622e7d update path youkaichao 2025-02-28 17:26:14 +08:00
  • 231e17ebb7 Merge pull request #29 from youkaichao/youkaichao-patch-1 Shangyan Zhou 2025-02-28 17:21:47 +08:00
  • 30e2778d18 Update README.md youkaichao 2025-02-28 16:56:09 +08:00
  • 77bb07aa20 Update some comments and docs Chenggang Zhao 2025-02-27 10:27:22 +08:00
  • 3885404ffb Add NVSHMEM_IB_ENABLE_RELAXED_ORDERING Chenggang Zhao 2025-02-26 17:54:12 +08:00
  • 45f481b87b Update figures Chenggang Zhao 2025-02-26 16:24:59 +08:00
  • 1a0a8bda09 Update prerequisites installation instructions haswelliris 2025-02-25 17:19:07 +08:00
  • 84d3d6fdee Update README.md Chenggang Zhao 2025-02-25 10:59:09 +08:00
  • ebfe47e46f Initial commit Chenggang Zhao 2025-02-24 21:56:14 +08:00