DeepSeek/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP synced 2025-05-05 20:44:48 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Stronger acquire scope for low-latency kernels

Browse Source
This commit is contained in:
Chenggang Zhao 2025-03-27 09:30:36 +08:00
parent 7d52ad7248
commit ffc39ba084
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
csrc/kernels/internode_ll.cu
View File

@ -260,7 +260,7 @@ dispatch(void* packed_recv_x, float* packed_recv_x_scales,
int num_recv_tokens, recv_token_begin_idx;
EP_STATIC_ASSERT(kNumWarpsPerGroup > 1, "Requires more than one warp per group");
if (sub_warp_id == 1 and lane_id == 0) {
while ((num_recv_tokens = ld_acquire_global(rdma_recv_count + local_expert_idx * num_ranks + src_rank)) == 0);
while ((num_recv_tokens = ld_acquire_sys_global(rdma_recv_count + local_expert_idx * num_ranks + src_rank)) == 0);
num_recv_tokens = -num_recv_tokens - 1;
recv_token_begin_idx = atomicAdd(packed_recv_count + local_expert_idx, num_recv_tokens);
shared_num_recv_tokens[warp_group_id] = num_recv_tokens;
@ -450,7 +450,7 @@ combine(void* combined_x,
if (responsible_expert_idx < num_experts) {
EP_STATIC_ASSERT(kNumWarpsPerGroup > 1, "Invalid number of warps per group");
if (sub_warp_id == 0 and lane_id == 0)
while (ld_acquire_global(rdma_recv_flag + responsible_expert_idx) == 0);
while (ld_acquire_sys_global(rdma_recv_flag + responsible_expert_idx) == 0);
}
cg::this_grid().sync();