Chenggang Zhao
74f4ef7b22
Remove the low-latency usage flag
2025-06-16 13:28:24 +08:00
Zhicheng Wu
05df5554ff
Use one qp per sm for internode normal kernels ( #181 )
...
let the sender SM use the channel_id, and the receiver SM use channel_id + num_channels
2025-06-13 14:37:59 +08:00
Shifang Xu
21efbe9b48
Support UE8M0 data format. ( #206 )
2025-06-12 09:38:19 +08:00
Chenggang Zhao
b8d90fb753
Support Ampere architecture ( #204 )
...
* Update README
* Update `setup.py`
* Fix headers
* Add `DISABLE_NVSHMEM` for APIs
* Fix launch
* Fix TMA settings
* Fix TMA usages
* Fix dlink
* Separate layout kernels
* Update version
* Add `is_sm90_compiled`
* Fix tests
* Add NVLink connection checks
* Update README
* Fix tests
* Add some comments
* Minor fix
* Minor fix
* Fix bugs
2025-06-11 15:48:18 +08:00
Chenggang Zhao
dd13c7145c
Check the empty list
2025-06-11 11:14:30 +08:00
Chenggang Zhao
a8299ca7c2
Support CUDA graph for intranode normal kernels ( #203 )
2025-06-11 11:08:54 +08:00
Chenggang Zhao
5a2e37fa28
Support statistics tensor for low-latency kernels ( #196 )
2025-06-09 15:50:56 +08:00
Chenggang Zhao
0d1a855d81
Add low-latency kernel PCIe usage flag ( #195 )
...
* Add low-latency kernel usage flag
* Update comments
2025-06-09 14:37:13 +08:00
Chenggang Zhao
c8dceba110
Use TMA instead of LD/ST for intra-node normal kernels ( #191 )
...
* Update CMake files
* Use TMA instead of LD/ST for intranode dispatch
* Use TMA instead of LD/ST for intranode combine
* Adjust configs
* Test default configs as well
* More warps for combine
* Add inter-thread fence
* Enable more warps
* Do not use TMA for senders
* Update configs
* Remove useless wait
2025-06-06 15:40:17 +08:00
Shangyan Zhou
20b2aaaf9e
Refactor some code.
2025-04-22 10:22:30 +08:00
moningchen
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.
...
In our optimized Internode Normal Kernel, we implemented multiple QPs for data transmission between two GPUs, setting a different QP for each channel. Additionally, we modified the transmission method from IBRC to IBGAD.
Through these optimizations, the Internode Normal Kernel achieves optimal performance in both H800 and H20 environments, with RDMA transmission performance nearly reaching the physical network performance limit. Using the current default statistical method, in 4-node H800 and H20 environments, RDMA performance can reach 60GB/s+.
2025-04-21 15:50:39 +08:00
Hao Lin
23c54150ba
Fix test combine args
...
Signed-off-by: Hao Lin <linhaomails@gmail.com>
2025-04-11 18:21:09 +08:00
fujianhao.fjh
0f80da8458
fix: not output result in some linux system
2025-04-10 18:18:30 +08:00
Chenggang Zhao
42494864ba
Remove useless control metadata for low-latency combine
2025-04-07 09:55:39 +08:00
Chenggang Zhao
26fa72d80f
Fix zero-copy mode tests
2025-03-28 16:49:33 +08:00
Chenggang Zhao
ae0eafd2be
Remove confusing comments
2025-03-25 09:27:34 +08:00
Chenggang Zhao
dcaf73e5ff
Support zero-copy for low-latency combine
2025-03-18 15:41:50 +08:00
Dmytro Dzhulgakov
b3b61ef5ef
Allow passing output tensor in low_latency_combine
2025-03-10 22:19:21 +00:00
Chenggang Zhao
ed7487c15e
Support BF16 for low-latency kernels
2025-03-10 17:24:41 +08:00
Chenggang Zhao
458cdcb22a
Fix AR bugs for normal kernels
2025-03-05 17:13:35 +08:00
Chenggang Zhao
1553fc42bf
Improve EP2/4 performance
2025-03-04 15:34:33 +08:00
Chenggang Zhao
c5b4040502
Enable intranode kernel tests with EP2 and EP4
2025-03-03 15:01:02 +08:00
Chenggang Zhao
ebfe47e46f
Initial commit
2025-02-25 09:07:53 +08:00
