[KV cache manager] No functional change intended, simplify block allocation in KVCacheManager::addSequence

eopXD · eopXD · commit 6eda9e8e954d · 2025-08-13T00:35:00.000-07:00
Simply we are allocating enough block to accommodate "window size"
number of tokens.

Signed-off-by: eopXD &lt;yuehtingc@nvidia.com&gt;
diff --git a/cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp b/cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp
@@ -1961,11 +1961,7 @@ void KVCacheManager::addSequence(
 
     for (auto const [windowSize, metadata] : mBlockManager.getWindowSizesMetadata())
     {
-        auto const maxTokenNum = metadata.maxTokenNum;
-        auto const temporaryAttentionWindow = metadata.temporaryAttentionWindow;
-
-        // Consider the temporaryAttentionWindow when allocating blocks.
-        auto const effectiveInputLength = std::min(inputLength, maxTokenNum + temporaryAttentionWindow);
+        auto const effectiveInputLength = std::min(inputLength, windowSize);
         auto const numContextBlocks = tc::ceilDiv(effectiveInputLength, getTokensPerBlock());
         if (!sequence.isCyclic() && mEnableBlockReuse)
         {

Original file line number	Diff line number	Diff line change
`@@ -1961,11 +1961,7 @@ void KVCacheManager::addSequence(`
`1961`	`1961`
`1962`	`1962`	`for (auto const [windowSize, metadata] : mBlockManager.getWindowSizesMetadata())`
`1963`	`1963`	`{`
`1964`		`- auto const maxTokenNum = metadata.maxTokenNum;`
`1965`		`- auto const temporaryAttentionWindow = metadata.temporaryAttentionWindow;`
`1966`		`-`
`1967`		`- // Consider the temporaryAttentionWindow when allocating blocks.`
`1968`		`- auto const effectiveInputLength = std::min(inputLength, maxTokenNum + temporaryAttentionWindow);`
	`1964`	`+ auto const effectiveInputLength = std::min(inputLength, windowSize);`
`1969`	`1965`	`auto const numContextBlocks = tc::ceilDiv(effectiveInputLength, getTokensPerBlock());`
`1970`	`1966`	`if (!sequence.isCyclic() && mEnableBlockReuse)`
`1971`	`1967`	`{`