Skip to content

Commit db697ce

Browse files
chuangz0mikeiovine
authored andcommitted
[https://nvbugs/5578175][fix] Fix block range index (#8470)
Signed-off-by: Chuang Zhu <[email protected]> Signed-off-by: Mike Iovine <[email protected]>
1 parent ddbb116 commit db697ce

File tree

1 file changed

+8
-2
lines changed

1 file changed

+8
-2
lines changed

cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp

Lines changed: 8 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -2142,7 +2142,7 @@ SizeType32 KVCacheManager::getNeededBlocksOneStep(
21422142
return 0;
21432143
}
21442144

2145-
auto const numCurrTokens = mSequences.at(req.mRequestId).getNumTokens();
2145+
auto const numCurrTokens = getSequence(req.mRequestId).getNumTokens();
21462146
auto const generatedTokens = numCurrTokens - req.getPromptLen();
21472147
auto const maxTokensToAddToKVCache = req.mMaxNewTokens - generatedTokens;
21482148
auto const tokensPerStep = req.getNumDraftTokens() + 1;
@@ -2406,7 +2406,13 @@ void KVCacheManager::addSequence(
24062406
void KVCacheManager::storeContextBlocks(LlmRequest const& llmRequest)
24072407
{
24082408
auto const requestId = llmRequest.mRequestId;
2409-
if (mSequences.find(requestId) != mSequences.end())
2409+
bool found = false;
2410+
{
2411+
// protect the mSequences
2412+
std::scoped_lock lock(mSequencesMtx);
2413+
found = mSequences.find(requestId) != mSequences.end();
2414+
}
2415+
if (found)
24102416
{
24112417
auto& sequence = getSequence(requestId);
24122418
if (mEnableBlockReuse && !llmRequest.isDummyRequest())

0 commit comments

Comments
 (0)