Skip to content

Commit 50e5e72

Browse files
authored
[https://nvbugs/5412456][fix] Fix an illegal instruction was encountered (#6776)
Signed-off-by: Zhou Yuxin <[email protected]>
1 parent 2e0081b commit 50e5e72

File tree

1 file changed

+5
-0
lines changed

1 file changed

+5
-0
lines changed

cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/fmhaRunner.cpp

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -297,6 +297,11 @@ void FusedMHARunnerV2::setupLaunchParams(MHARunnerParams runnerParams)
297297
= mFixedParams.isSPadded ? runnerParams.b * runnerParams.qSeqLen : runnerParams.totalQSeqLen;
298298
mLaunchParams.total_kv_seqlen
299299
= mFixedParams.isSPadded ? runnerParams.b * runnerParams.kvSeqLen : runnerParams.totalKvSeqLen;
300+
// Workaround for nvbug 5412456: total_kv_seqlen fallbacks to total_q_seqlen if it's zero.
301+
if (mLaunchParams.total_kv_seqlen == 0)
302+
{
303+
mLaunchParams.total_kv_seqlen = mLaunchParams.total_q_seqlen;
304+
}
300305

301306
TLLM_CHECK_WITH_INFO(mFixedParams.headSize > 0, "Head size should be greater than 0.");
302307
// Pad head size to next power of 2.

0 commit comments

Comments
 (0)