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

Use one qp per sm for internode normal kernels (#181)

Browse Source 
let the sender SM use the channel_id, and the receiver SM use channel_id + num_channels
This commit is contained in:
Zhicheng Wu
2025-06-13 14:37:59 +08:00
committed by GitHub
parent 21efbe9b48
commit 05df5554ff
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
csrc/kernels/internode.cu
View File

@@ -831,7 +831,7 @@ dispatch(int4* recv_x, float* recv_x_scales, int64_t* recv_topk_idx, float* recv
// Update remote head
if (min_head != std::numeric_limits<int>::max() and min_head >= last_head + num_max_rdma_chunked_send_tokens and lane_id < kNumRDMARanks) {
nvshmemi_ibgda_amo_nonfetch_add(rdma_channel_head.buffer(rdma_rank), min_head - last_head,
translate_dst_rdma_rank<kLowLatencyMode>(lane_id, nvl_rank), channel_id, lane_id == rdma_rank);
translate_dst_rdma_rank<kLowLatencyMode>(lane_id, nvl_rank), channel_id + num_channels, lane_id == rdma_rank);
last_head = min_head;
}
@@ -1563,7 +1563,7 @@ combine(int4* combined_x, float* combined_topk_weights,
min_head = min(min_head, rdma_receiver_rdma_head[i][dst_rdma_rank]);
if (min_head != std::numeric_limits<int>::max() and min_head >= last_rdma_head + num_max_rdma_chunked_send_tokens and lane_id < kNumRDMARanks) {
nvshmemi_ibgda_amo_nonfetch_add(rdma_channel_head.buffer(rdma_rank), min_head - last_rdma_head,
translate_dst_rdma_rank<kLowLatencyMode>(dst_rdma_rank, nvl_rank), channel_id, dst_rdma_rank == rdma_rank);
translate_dst_rdma_rank<kLowLatencyMode>(dst_rdma_rank, nvl_rank), channel_id + num_channels, dst_rdma_rank == rdma_rank);
last_rdma_head = min_head;
}
} else {