8 Commits
Author SHA1 Message Date
skyandClaude 0132848349 Revert "[Fix] Disable --enable-symm-mem under CUDA graphs on Kimi hybrid models" (#34842)
Signed-off-by: wangfakang <fakangwang@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
2026-08-27 14:19:27 -07:00
skyandBrayden Zhong 980acd6eca Fix MoE TP allreduce to use NCCL symmetric memory via in-pool output allocation (#29007)
Signed-off-by: wangfakang <fakangwang@gmail.com>
Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>
2026-07-15 15:06:37 +08:00
sky 409d350fb6 Bugfix: fix symm not enabled due to incorrect registration of comm (#19329)
Signed-off-by: wangfakang <fakangwang@gmail.com>
2026-05-13 10:01:04 +08:00
sky bfc1aeae13 [CP] Register KV cache allgather buffer with symmetric memory (#24040)
Signed-off-by: wangfakang <fakangwang@gmail.com>
2026-05-06 23:24:36 +08:00
sky c8bc23522f Refactor: decouple segment tracking from comm registration (#21392)
Signed-off-by: wangfakang <fakangwang@gmail.com>
2026-05-06 17:07:58 +08:00
sky 3d58cd16d9 [DP Attention] Optimize dp_padding_mode selection for dp_size=1 in extend mode (#20406)
Signed-off-by: wangfakang <fakangwang@gmail.com>
2026-03-16 18:44:42 +08:00
sky 17119a697d Optimization: Reduce the number of D2H operations (#19424)
Signed-off-by: wangfakang <fakangwang@gmail.com>
2026-03-04 16:32:42 +08:00
sky 72c1526657 Register cp-atten-allgather buffers with symm memory (#17756)
Signed-off-by: wangfakang <fakangwang@gmail.com>
2026-02-11 15:26:37 +08:00