@@ -1366,35 +1366,6 @@ struct ncclEpHandle {
13661366 int hidden_int4;
13671367
13681368 union {
1369- struct {
1370- // Both intranode and internode
1371- int * recv_counter;
1372- int * recv_counter_device;
1373- int * internal_recv_expert_counter_host = nullptr ;
1374- int received_token_count = -1 ;
1375- ncclNDTensor_t rank_token_counts;
1376- ncclNDTensor_t expert_token_counts;
1377- ncclNDTensor_t token_rank_mask;
1378- ncclNDTensor_t global_channel_prefix;
1379- ncclNDTensor_t nvl_send_head;
1380- ncclNDTensor_t recv_global_channel_prefix;
1381-
1382- // Internode only
1383- int * rdma_recv_counter;
1384- int * rdma_recv_counter_device;
1385- int rdma_received_token_count = -1 ;
1386- std::optional<ncclNDTensor_t> rdma_rank_token_counts;
1387- ncclNDTensor_t rdma_channel_prefix;
1388- ncclNDTensor_t recv_rdma_rank_prefix;
1389- ncclNDTensor_t recv_global_rank_prefix;
1390- ncclNDTensor_t rdma_send_head;
1391- ncclNDTensor_t recv_source_metadata;
1392- ncclNDTensor_t recv_rdma_channel_prefix;
1393-
1394- // Intranode only
1395- ncclNDTensor_t inter_rank_token_offsets;
1396- ncclNDTensor_t recv_token_source_map;
1397- } ht;
13981369 struct {
13991370 // packed tensors for LL
14001371 ncclNDTensor_t expert_recv_source_indices;
@@ -1539,10 +1510,8 @@ struct ncclEpHandle {
15391510 cached_mode(false ),
15401511 num_scales(0 ),
15411512 hidden_int4(0 ) {
1542- // Zero the entire union (ht, ll, hybridep share memory)
1543- // Use max size to ensure all members are zeroed
1544- constexpr size_t union_size = std::max ({sizeof (ht), sizeof (ll), sizeof (hybridep)});
1545- memset (static_cast <void *>(&ht), 0 , union_size);
1513+ constexpr size_t union_size = std::max (sizeof (ll), sizeof (hybridep));
1514+ memset (static_cast <void *>(&ll), 0 , union_size);
15461515 }
15471516
15481517 ~ncclEpHandle () {
0 commit comments