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

Add sender timeout checks

Browse Source
This commit is contained in:
Chenggang Zhao 2025-06-23 15:41:49 +08:00
parent be96674e94
commit c0da8eaba5
1 changed files with 13 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
csrc/kernels/internode.cu
View File

@ -475,7 +475,7 @@ dispatch(int4* recv_x, float* recv_x_scales, int64_t* recv_topk_idx, float* recv
// Acquire a tail
int rdma_tail_idx = -1;
if (is_token_in_rank_uint64 != 0) {
do {
while (true) {
// Acquire lock first
acquire_lock(rdma_send_channel_lock + lane_id);
@ -495,11 +495,21 @@ dispatch(int4* recv_x, float* recv_x_scales, int64_t* recv_topk_idx, float* recv
// Release lock
release_lock(rdma_send_channel_lock + lane_id);
} while (rdma_tail_idx == -1);
break;
}
// Wait the remote buffer to be released
while (rdma_tail_idx - cached_rdma_channel_head >= num_max_rdma_chunked_recv_tokens)
auto start_time = clock64();
while (rdma_tail_idx - cached_rdma_channel_head >= num_max_rdma_chunked_recv_tokens) {
cached_rdma_channel_head = static_cast<int>(ld_volatile_global(rdma_channel_head.buffer(lane_id)));
// Timeout check
if (clock64() - start_time >= NUM_TIMEOUT_CYCLES) {
printf("DeepEP dispatch RDMA sender timeout, channel: %d, RDMA: %d, nvl: %d, dst RDMA lane: %d, head: %d, tail: %d\n",
channel_id, rdma_rank, nvl_rank, lane_id, cached_rdma_channel_head, rdma_tail_idx);
trap();
}
}
}
__syncwarp();