debug

Author: zzx-study

csrc/deepep/deep_ep.cpp

@@ -797,6 +797,15 @@ Buffer::internode_dispatch(
                  recv_data, token_server_idx, token_unique_per_server, ep_rank_token_cnt, recv_tokens_per_expert,
                  src_offset_rank_token_idx, dst_offset_rank_token_idx, token_idx_per_expert, offset_inner, count_outer, expand_idx,
                  total_recv_token);
+    at::Tensor cpu = token_idx_per_expert.cpu();
+auto acc = cpu.accessor<int, 2>();
+
+for (int i = 0; i < cpu.size(0); ++i) {
+    for (int j = 0; j < cpu.size(1); ++j) {
+        std::cout << acc[i][j] << " ";
+    }
+    std::cout << std::endl;
+}
 
     int total_count = total_recv_token.item<int>();
     int num_recv_tokens = (total_count == 0) ? 1 : total_count;

csrc/deepep/ops2/op_kernel/a2/moe_distribute_dispatch_a2_pipeline.h

@@ -527,11 +527,9 @@ __aicore__ inline void MoeDistributeDispatchA2Pipeline<TemplateMC2TypeA2Pipeline
                 SyncFunc<AscendC::HardEvent::S_MTE3>();
                 DataCopyPad(dstRankRecvRingU8Tensor[hccsItemNum * localDstRank + tokenStructLen_ * index], tokenStructInRdmaTensor_, 
                 tokenStructParams);
-                DataCopyPad(rdmaRecvRingU8Tensor_[(i * rdmaItemNum + rdmaHead) * tokenStructLen_], tokenStructInRdmaTensor_, 
-                tokenStructParams);
                 rdmaHead = (rdmaHead + 1) % rdmaItemNum;
-                hcclTail = (hcclTail + 1) % hccsItemNum;
                 rdmaHeadTailTensor_.SetValue(i * RING_BUFFER_HEAD_TAIL + 2, rdmaHead);
+                hcclTail = (hcclTail + 1) % hccsItemNum;
                 localHccsHeadTailTensor.SetValue(1, hcclTail);
                 DataCopy(globalHccsHeadTailTensor[localRankId], localHccsHeadTailTensor, hccsHesdTailParams);
             }