Skip to content

Commit 41ce14a

Browse files
[None][feat] Enable NCCL_SYMMETRIC as default fallback for AllReduce (#9314)
Signed-off-by: Ludwig Schneider <[email protected]>
1 parent d252101 commit 41ce14a

File tree

20 files changed

+2222
-389
lines changed

20 files changed

+2222
-389
lines changed

cpp/tensorrt_llm/common/customAllReduceUtils.h

Lines changed: 2 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -81,7 +81,6 @@ inline AllReduceStrategyType SelectStrategyLP(size_t seq_len, size_t hidden_size
8181
{
8282
return AllReduceStrategyType::ONESHOT;
8383
}
84-
return AllReduceStrategyType::NCCL;
8584
}
8685

8786
// use 1D vector to store the best strategy instead of a map for each sm version
@@ -143,15 +142,15 @@ inline AllReduceStrategyType selectStrategyLookUpTable(
143142
sm_version = 100;
144143
}
145144

146-
// Check if the entry is out of bounds, otherwise return NCCL as fallback
145+
// Check if the entry is out of bounds, otherwise return NCCL_SYMMETRIC as fallback
147146
if (AllReduceBestStrategyTable.find(sm_version) == AllReduceBestStrategyTable.end()
148147
|| tp_index >= AllReduceBestStrategyTable.at(sm_version).size()
149148
|| fusion_op_index >= AllReduceBestStrategyTable.at(sm_version).at(tp_index).size()
150149
|| hidden_size_index >= AllReduceBestStrategyTable.at(sm_version).at(tp_index).at(fusion_op_index).size()
151150
|| num_token_index
152151
>= AllReduceBestStrategyTable.at(sm_version).at(tp_index).at(fusion_op_index).at(hidden_size_index).size())
153152
{
154-
return AllReduceStrategyType::NCCL;
153+
return AllReduceStrategyType::NCCL_SYMMETRIC;
155154
}
156155

157156
return static_cast<AllReduceStrategyType>(

0 commit comments

Comments
 (0)