Skip to content

cherry-pick: [fix: nvbugs/5355493] Correctly clamp max sequence len to max attention window #5874

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jul 9, 2025
Merged

cherry-pick: [fix: nvbugs/5355493] Correctly clamp max sequence len to max attention window #5874

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5c034d4
[fix: nvbugs/5355493] Correctly clamp max sequence len to max attenti…
netanel-haber Jul 4, 2025
72dc7a4
match cherry-pick to pre https://github.com/NVIDIA/TensorRT-LLM/pull/…
netanel-haber Jul 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion cpp/tensorrt_llm/batch_manager/trtGptModelInflightBatching.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -559,7 +559,7 @@ TrtGptModelInflightBatching::clampWindowSizesToFitAtLeastOneSequence(BlocksPerWi
TLLM_LOG_WARNING("maxAttentionWindowVec too large to fit at least one sequence in kvCache. Old: %s, New: %s",
common::vec2str(getMaxAttentionWindowVec()).c_str(), common::vec2str(newMaxAttentionWindowVec).c_str());
setMaxAttentionWindowVec(newMaxAttentionWindowVec);
if (getMaxSequenceLen() < getMaxAttentionWindow())
if (getMaxSequenceLen() > getMaxAttentionWindow())
{
TLLM_LOG_WARNING("maxSequenceLen is reduced to maxAttentionWindow: %d", getMaxAttentionWindow());
setMaxSequenceLen(getMaxAttentionWindow());
Expand Down
24 changes: 24 additions & 0 deletions cpp/tests/batch_manager/trtGptModelTest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -870,6 +870,11 @@ class TrtGptModelIfbHelper : public TrtGptModelInflightBatching
{
return TrtGptModelInflightBatching::getKVCacheManager();
}

[[nodiscard]] SizeType32 getMaxAttentionWindow() const
{
return TrtGptModelInflightBatching::getMaxAttentionWindow();
}
};

TEST_F(TrtGptModelTest, KVCacheReuseChunked)
Expand Down Expand Up @@ -1201,4 +1206,23 @@ TEST_F(LlamaModelLADTest, SeamlessLookaheadDecoding)
}
}

TEST_F(TrtGptModelTest, ClampSeqLenToAttentionWindow)
{
auto constexpr maxAttentionWindow = 65536;
auto constexpr maxSequenceLen = maxAttentionWindow + 1;

TrtGptModelOptionalParams optionalParams;
optionalParams.kvCacheConfig.maxAttentionWindowVec = std::vector<SizeType32>{maxAttentionWindow};
optionalParams.kvCacheConfig.freeGpuMemoryFraction = 0.0001; // minuscule amount of memory to force a clamp
optionalParams.maxBeamWidth = mBeamWidth;

auto modelConfig = mModelConfig;
modelConfig.setMaxSequenceLen(maxSequenceLen);

auto trtGptModel
= std::make_shared<TrtGptModelIfbHelper>(mLogger, modelConfig, mWorldConfig, *mRawEngine, true, optionalParams);
EXPECT_LT(trtGptModel->getMaxAttentionWindow(), maxAttentionWindow);
EXPECT_EQ(trtGptModel->getMaxSequenceLen(), trtGptModel->getMaxAttentionWindow());
}

} // namespace tensorrt_llm::batch_manager