[https://nvbugs/6248769][fix] In KVCacheManager.__init__, after resolving max_attention_window_vec and…

[tensorrt-cicd]

Summary

• Root cause: max_attention_window_vec is clamped to max_seq_len (512) by _resolve_max_attention_window_vec, but pool_configurations retained the unclamped sliding window (Ministral's 32768), so layers resolving to window 512 had no matching pool key → KeyError: 512.
• Fix: In KVCacheManager.__init__, after resolving max_attention_window_vec and before _build_layer_to_pool_idx(), rebuild self.pool_configurations with window_size=min(pc.window_size, self.max_seq_len), keeping head_dim/dtype and pool indices intact.
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6248769

Summary by CodeRabbit

• Bug Fixes

  • Fixed KV cache memory pool configuration to ensure window size values remain consistent with normalized attention specifications, improving overall cache memory management reliability and stability across different configurations

• Tests

  • Re-enabled accuracy test for model registry functionality, restoring test coverage for model deployment and registration scenarios

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 825a2fd2-894e-4090-833f-64ae56a3715e

📥 Commits

Reviewing files that changed from the base of the PR and between c798fd9 and 324f098.

📒 Files selected for processing (2)

• tensorrt_llm/_torch/pyexecutor/resource_manager.py
• tests/integration/test_lists/waives.txt

💤 Files with no reviewable changes (1)

• tests/integration/test_lists/waives.txt

---

📝 Walkthrough

Walkthrough

KVCacheManager now imports replace from dataclasses to clamp per-pool window sizes to max_seq_len after computing the maximum attention window vector, maintaining consistency between per-pool metadata and clamped window sizes. A related test waiver is removed from the accuracy test skip list.

Changes

KV Cache Manager Window Clamping and Test Configuration Update

Layer / File(s)	Summary
Window size clamping with dataclasses import tensorrt_llm/_torch/pyexecutor/resource_manager.py	replace is added to the dataclasses import to enable remapping of per-pool pool_configurations entries after max_attention_window_vec computation, clamping each window_size to min(original_window_size, self.max_seq_len).
Accuracy test waiver removal tests/integration/test_lists/waives.txt	A specific SKIP waiver entry for test_autodeploy_from_registry targeting mistralai_Ministral-8B-Instruct-2410-False is removed.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#14870: Also removes a waived accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy test entry from waives.txt.
• NVIDIA/TensorRT-LLM#14802: Modifies the waiver for the same accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[mistralai_Ministral-8B-Instruct-2410-False] test case.
• NVIDIA/TensorRT-LLM#14854: Also updates the set of waived/skipped test cases in waives.txt.

Suggested reviewers

• nvchenghaoz
• xinhe-nv

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main fix: clamping per-pool window_size values in KVCacheManager.init to resolve KeyError mismatches.
Description check	✅ Passed	The description adequately explains the root cause, fix details, and test verification. However, PR checklist items are not explicitly checked off as required by the template.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[MrGeva]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1" --disable-fail-fast

[tensorrt-cicd]

PR_Github #51816 [ run ] triggered by Bot. Commit: 324f098 Link to invocation

[tensorrt-cicd]

PR_Github #51816 [ run ] completed with state SUCCESS. Commit: 324f098
/LLM/main/L0_MergeRequest_PR pipeline #41179 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation