[TLLM-6777][feature] Support SWA KV cache reuse OOW block detach #7922

eopXD · 2025-09-23T08:20:26Z

Summary by CodeRabbit

New Features
- Enhanced KV-cache with sequence-aware management and sliding-window attention support.

Description

NOTE: Please note that the depending MR is still on-going and not merged yet.

This MR is a continuation of #6768. In the previous merge request, OOW (out-of-window) blocks are only detached when reuse is not enabled. This MR enables KV cache manager to detach the block when reuse is enabled.

When reuse is enabled, OOW block gets detached while the manager bookkeep whether the block is overwritten by another sequence. If at the end of the sequence all blocks for it is still clean, then the sequence block will be stored.

Test Coverage

Regarding unit test, KVCacheManagerVariableWindowAttentionWithReuseTest, KVCacheManagerSWAInvalidateReuseTest, and KVCacheManagerMaxAttentionWindowWithReuseTest are added to verify SWA reuse and block invalidating mechanism.

Regarding end-to-end, the existing test for vswa setting for gemma3 for the mmlu and gsm8k benchmark will cover the code paths added in this merge request.

PR Checklist

Please review the following before submitting your PR:

PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...

Provide a user friendly way for developers to interact with a Jenkins server.

Run /bot [-h|--help] to print this help message.

See details below for each supported subcommand.

run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]

Launch build/test pipelines. All previously running jobs will be killed.

--reuse-test (optional)pipeline-id (OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.

--disable-reuse-test (OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.

--disable-fail-fast (OPTIONAL) : Disable fail fast on build/tests/infra failures.

--skip-test (OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.

--stage-list "A10-PyTorch-1, xxx" (OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.

--gpu-type "A30, H100_PCIe" (OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.

--test-backend "pytorch, cpp" (OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.

--only-multi-gpu-test (OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.

--disable-multi-gpu-test (OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.

--add-multi-gpu-test (OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.

--post-merge (OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.

--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" (OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".

--detailed-log (OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.

--debug (OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in the stage-list parameter to access the appropriate container environment. Note: Does NOT update GitHub check status.

For guidance on mapping tests to stage names, see docs/source/reference/ci-overview.md
and the scripts/test_to_stage_mapping.py helper.

kill

kill

Kill all running builds associated with pull request.

skip

skip --comment COMMENT

Skip testing for latest commit on pull request. --comment "Reason for skipping build/test" is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

reuse-pipeline

reuse-pipeline

Reuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

coderabbitai · 2025-09-23T08:30:25Z

📝 Walkthrough

Walkthrough

The change set makes KV-cache management sequence-aware and SWA-aware across headers, implementation, and bindings. It replaces optional max-sequence-length with required parameters, adjusts constructor signatures and call sites, adds per-sequence bookkeeping and block reuse validity, removes cyclic KV-cache logic, updates tests, and refreshes related help text/docs.

Changes

Cohort / File(s)	Summary of changes
KV-cache core API and implementation `cpp/include/tensorrt_llm/batch_manager/kvCacheManager.h`, `cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp`	Introduces per-sequence block ownership/validity, SWA flag handling, front-block detachment/adjustment, and sequence-aware store/onboard/release paths. Adds kSWAExtraBlock, maxBlocksPerSeq, getAllSequenceBlocks. Changes many method signatures to thread GenerationRequest. Makes maxSequenceLength required.
Inflight batching integration `cpp/tensorrt_llm/batch_manager/trtGptModelInflightBatching.cpp`	Removes cyclic KV-cache logic and “use one more block” path. Updates KVCacheManager construction to pass concrete max sequence/encoder length and new reuse options. Adjusts RewindInputs to fixed false.
Python bindings `cpp/tensorrt_llm/nanobind/batch_manager/kvCacheManager.cpp`	Updates KVCacheManager constructor binding: previously-optional SizeType32 is now required; argument names aligned with new signature, including enable_partial_reuse and copy_on_partial_reuse.
C++ unit tests (constructor alignment) `cpp/tests/unit_tests/batch_manager/cacheTransBufferTest.cpp`, `cpp/tests/unit_tests/batch_manager/capacitySchedulerTest.cpp`, `cpp/tests/unit_tests/executor/agentCommTest.cpp`, `cpp/tests/unit_tests/multi_gpu/cacheTransceiverTest.cpp`	Updates KVCacheManager call sites to pass required maxNumTokens/maxSequenceLength argument (replacing std::nullopt and prior parameter clusters). Adjusts parameters ordering; disables some window-related multi-GPU test instantiations.
Executor PyTorch util `tensorrt_llm/_torch/pyexecutor/_util.py`	Unifies max_tokens computation; removes VSWA-specific branch and always derives from kv_cache_max_memory.
Functional docstrings `tensorrt_llm/functional.py`	Updates wording for sliding-window-attention description; no functional changes.
Examples help text `examples/models/core/llama/summarize_long.py`, `examples/models/core/qwen2audio/utils.py`, `examples/utils.py`	Removes “cyclic” wording from --max_attention_window_size help strings.
Integration tests (VSWA variants) `tests/integration/defs/accuracy/test_llm_api_pytorch.py`, `tests/integration/test_lists/test-db/l0_h100.yml`	Replaces a skipped VSWA test with four explicit VSWA variants (reuse/non-reuse, chunked prefill on/off). Updates test list entries accordingly.

Sequence Diagram(s)

sequenceDiagram
  autonumber
  actor App as Caller
  participant KM as KVCacheManager
  participant BM as BlockManager
  participant WBM as WindowBlockManager
  participant Seq as GenerationRequest
  rect rgba(230,240,255,0.5)
    note over App,KM: Add sequence and allocate context
    App->>KM: addSequence(Seq, ...)
    KM->>BM: holdSequence(Seq.id)
    KM->>BM: adjustBlocksIfNeeded(Seq, enableBlockReuse)
  end
  rect rgba(240,255,240,0.5)
    note over App,WBM: Onboard a new block (sequence-aware)
    App->>KM: onboardBlock(Seq, windowSize, ...)
    KM->>BM: onboardBlock(Seq, ...)
    BM->>WBM: onboardBlock(Seq, ...)
    WBM->>WBM: getFreeBlock(Seq, ...)
    WBM-->>BM: BlockPtr
    BM-->>KM: BlockPtr
    KM-->>App: BlockPtr
  end
  alt SWA sequence
    note right of WBM: SWA guards: skip store/detach on-the-fly
  else Non-SWA
    rect rgba(255,250,230,0.6)
      note over WBM: Detach/adjust front blocks when needed
      KM->>BM: adjustBlocksIfNeeded(Seq, enableBlockReuse)
      BM->>WBM: adjustBlocksIfNeeded(Seq, enableBlockReuse)
      WBM->>WBM: detachFrontBlock(Seq, enableBlockReuse)
    end
  end
  rect rgba(255,235,235,0.6)
    note over App,WBM: Store or release blocks (sequence validity tracked)
    App->>KM: releaseBlocks(Seq, ...)
    KM->>BM: releaseBlocks(Seq, ...)
    BM->>WBM: storeBlocks(Seq, ...) / release(...)
    WBM->>BM: update block->sequence mappings
  end
  KM->>BM: releaseSequence(Seq.id)

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~75 minutes

Pre-merge checks and finishing touches

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 20.48% which is insufficient. The required threshold is 80.00%.	You can run `@coderabbitai generate docstrings` to improve docstring coverage.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title Check	✅ Passed	The title "[TLLM-6777][feature] Support SWA KV cache reuse OOW block detach" follows the repository convention by including a ticket and type and succinctly states the primary change: enabling out-of-window (OOW) block detachment for SWA KV-cache reuse. It is specific, on-topic, and free of noise or file lists, so a teammate scanning history will understand the main purpose of the PR.
Description Check	✅ Passed	The PR description follows the repository template by providing a clear Description that explains the issue and solution, a Test Coverage section listing added unit tests and end-to-end coverage, and a PR Checklist; it also notes the dependent MR is still ongoing. The content is specific and actionable for reviewers, and it lists relevant tests that exercise the new SWA reuse and invalidation behavior.

✨ Finishing touches

📝 Generate Docstrings

🧪 Generate unit tests

Create PR with unit tests
Post copyable unit tests in a comment

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coderabbitai

Actionable comments posted: 3

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

cpp/tests/unit_tests/multi_gpu/cacheTransceiverTest.cpp (1)
582-586: Bug: cacheType shadowing prevents kSELFKONLY from being set.

Inner ‘auto cacheType’ shadows the outer variable; SELF is always used even when kvFactor == 1.

Fix:
-        CacheType cacheType = CacheType::kSELF;
-        if (kvFactor == 1)
-        {
-            auto cacheType = CacheType::kSELFKONLY;
-        }
+        CacheType cacheType = (kvFactor == 1) ? CacheType::kSELFKONLY : CacheType::kSELF;

🧹 Nitpick comments (12)

examples/utils.py (1)
356-362: Help text tweak: clarify plural semantics for nargs="+"

Since this arg accepts multiple integers, make the wording plural and add a brief example.
-        'The attention window size that controls the sliding window attention kv cache behavior'
+        'Attention window size(s) controlling sliding-window attention KV-cache behavior. '
+        'Accepts one or more ints (e.g., --max_attention_window_size 4096 or 2048 4096).'
examples/models/core/qwen2audio/utils.py (2)
41-41: Polish help text for clarity and consistency (KV‑cache, hyphenation).

Suggest concise copy and consistent hyphenation/capitalization.
-        'The attention window size that controls the sliding window attention kv cache behavior'
+        'Controls the sliding-window attention KV-cache behavior.'
1-1: Missing SPDX/NVIDIA Apache-2.0 header.

Per guidelines, add the license header to Python sources.
+# SPDX-FileCopyrightText: Copyright (c) 2022-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
 from argparse import BooleanOptionalAction
cpp/tests/unit_tests/batch_manager/cacheTransBufferTest.cpp (1)
59-63: Constructor signature alignment LGTM; prefer explicit SizeType32 for kvMaxNumTokens.

Avoid implicit int → SizeType32 conversions to keep types consistent across platforms.

Apply:
-        auto kvMaxNumTokens = tokensPerBlock * maxBlocksPerSeq;
+        SizeType32 kvMaxNumTokens = static_cast<SizeType32>(tokensPerBlock) * maxBlocksPerSeq;
cpp/tests/unit_tests/executor/agentCommTest.cpp (1)
84-88: Constructor update LGTM; minor type tightening suggested.

Use SizeType32 for kvMaxNumTokens to match manager expectations.
-        auto kvMaxNumTokens = tokensPerBlock * maxBlocksPerSeq;
+        SizeType32 kvMaxNumTokens = static_cast<SizeType32>(tokensPerBlock) * maxBlocksPerSeq;
cpp/tests/unit_tests/multi_gpu/cacheTransceiverTest.cpp (2)
621-623: Constructor update LGTM; consider explicit type for maxNumTokens.

Use SizeType32 to avoid implicit narrowing and mirror other callsites.
-            dataType, sinkTokenLength, stream, maxNumTokens, enableBlockReuse, onboardBlocks, CacheType::kSELF,
+            dataType, sinkTokenLength, stream, static_cast<SizeType32>(maxNumTokens), enableBlockReuse, onboardBlocks, CacheType::kSELF,
1316-1337: Temporarily waiving isWindow tests reduces coverage.

Add a TODO and plan to re‑enable once SWA reuse/OOW detach stabilizes; otherwise we may miss regressions.
cpp/include/tensorrt_llm/batch_manager/kvCacheManager.h (1)
736-736: Fix inconsistent parameter naming in onboardBlock methods.

The onboardBlock methods use inconsistent parameter naming - copyOnPartialReuse in some signatures and copyOnpartialReuse in others.

Apply this diff to fix the inconsistent casing:
-        bool copyOnpartialReuse = true,
+        bool copyOnPartialReuse = true,
-        bool copyOnpartialReuse = true,
+        bool copyOnPartialReuse = true,
Also applies to: 828-829, 951-951, 988-989, 1053-1055, 1060-1060
cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp (4)
324-327: Remove or relocate debug logging from production code.

The debug logging before the assertion seems misplaced. Consider removing it or wrapping it in a debug-only build flag.
-    if (!hasRefs())
-    {
-        TLLM_LOG_DEBUG("::decRefCount Going to hit assertion for block %d", getBlockId());
-    }
     TLLM_CHECK_WITH_INFO(hasRefs(), "Can't remove link from block that is not allocated");
507-519: Consider using the intended calculateWindowSizeToShare logic.

The comment at line 519-520 indicates that the current uniform distribution (1.0f / windowSizeToLayers.size()) is temporary. The actual proportional calculation based on window sizes is already implemented but commented out. Consider enabling the intended logic if testing has been completed.

The function calculateWindowSizeToShare appears to have sophisticated logic for proportional distribution, but it's not being used. Consider tracking this as a TODO if the uniform distribution is indeed temporary:
+    // TODO(TLLM-XXXX): Enable proportional block allocation based on window sizes
+    // once testing is complete. Currently using uniform distribution for identical performance.
     std::map<SizeType32, float> windowSizeToShare;
1091-1096: Consider documenting the offloading behavior trade-offs.

The comment correctly identifies that the current default behavior may lead to unnecessary traffic. Consider tracking this as a technical debt item.

Add a TODO to track the improvement opportunity:
+    // TODO(TLLM-XXXX): Delegate offloading decisions to the eviction policy
+    // to avoid unnecessary traffic from offloading blocks that won't be reused.
     // The current default behavior is to offload the out-of-window block
1485-1547: Document the return value usage of storeBlocks.

The method now returns the number of blocks stored, but this information appears to be used only for logging. Consider documenting this behavior.

Add a comment explaining the return value usage:
 SizeType32 WindowBlockManager::storeBlocks(
     std::vector<BlockKey> const& blockKeys, std::vector<KVCacheBlock::IdType> const& blockIds)
 {
+    // Returns the number of blocks actually stored (excludes already-matched blocks)
     SizeType32 numBlocksStoredForReuse = 0;

📜 Review details

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between bb64e74 and 3da009f.

📒 Files selected for processing (15)

cpp/include/tensorrt_llm/batch_manager/kvCacheManager.h (22 hunks)
cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp (42 hunks)
cpp/tensorrt_llm/batch_manager/trtGptModelInflightBatching.cpp (2 hunks)
cpp/tensorrt_llm/nanobind/batch_manager/kvCacheManager.cpp (1 hunks)
cpp/tests/unit_tests/batch_manager/cacheTransBufferTest.cpp (1 hunks)
cpp/tests/unit_tests/batch_manager/capacitySchedulerTest.cpp (1 hunks)
cpp/tests/unit_tests/executor/agentCommTest.cpp (1 hunks)
cpp/tests/unit_tests/multi_gpu/cacheTransceiverTest.cpp (3 hunks)
examples/models/core/llama/summarize_long.py (1 hunks)
examples/models/core/qwen2audio/utils.py (1 hunks)
examples/utils.py (1 hunks)
tensorrt_llm/_torch/pyexecutor/_util.py (1 hunks)
tensorrt_llm/functional.py (1 hunks)
tests/integration/defs/accuracy/test_llm_api_pytorch.py (1 hunks)
tests/integration/test_lists/test-db/l0_h100.yml (1 hunks)

🧰 Additional context used

📓 Path-based instructions (8)

**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}