Skip to content

Commit ce401d7

Browse files
committed
fix routed
Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
1 parent cdf67c7 commit ce401d7

1 file changed

Lines changed: 2 additions & 1 deletion

File tree

csrc/fused_moe/trtllm_backend/trtllm_fused_moe_routing_renormalize.cu

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -53,7 +53,8 @@ void run(Data const& data, void* stream) {
5353
FLASHINFER_CHECK(data.mNumExperts % 4 == 0,
5454
"Routing kernel expects #experts %d to be a multiple of 4.", data.mNumExperts);
5555

56-
bool const useSingleBlock = data.mNumTokens <= BlockKernelMaxNumTokens;
56+
bool const useSingleBlock =
57+
data.mNumTokens <= BlockKernelMaxNumTokens && data.mPtrTopKPacked == nullptr;
5758

5859
bool const useSingleCluster =
5960
data.mNumTokens <= ((data.mPtrScores != nullptr || data.mPtrTopKIds != nullptr)

0 commit comments

Comments
 (0)