DeepSeek/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP synced 2025-06-26 18:28:11 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

To mitigate incast congestion, shuffle the starting index of target rank for different ranks and channels

Browse Source 
Signed-off-by: wangfakang <fakangwang@gmail.com>
This commit is contained in:
wangfakang 2025-05-09 17:43:01 +08:00
parent 9056a6db95
commit 63c29d06a0
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
csrc/kernels/internode.cu
View File

@ -703,7 +703,8 @@ dispatch(int4* recv_x, float* recv_x_scales, int64_t* recv_topk_idx, float* recv
int last_issued_tail = 0;
while (__any_sync(0xffffffff, num_tokens_to_send > 0)) {
for (int i = 0, synced_num_tokens_to_send; i < kNumRDMARanks; ++ i) {
int dst_rdma_rank = (i + channel_id) % kNumRDMARanks;
// To mitigate incast congestion, shuffle the starting index of target rank for different ranks and channels
int dst_rdma_rank = (i + channel_id + rdma_rank) % kNumRDMARanks;
synced_num_tokens_to_send = __shfl_sync(0xffffffff, num_tokens_to_send, dst_rdma_rank);
if (synced_num_tokens_to_send == 0)
continue;