[https://nvbugs/6248764][fix] Normalize non-sliding KV windows to full attention in AutoDeploy

[eopXD]

A KV-cache layer whose sliding_window >= max_seq_len can never slide and is functionally full attention. The AutoDeploy kvcache transform was keying pool grouping off the raw sliding_window, so such a layer forked its own pool and tripped the trtllm backend's uniform-cache requirement.

For google/gemma-3-1b-it at max_seq_len=511, the local layers' sliding_window=512 produced a window=512 pool alongside the global window=511 pool, failing with: KV resources are not uniform ... requires all KV caches to share a single pool.

This normalizes sliding_window to 0 (full attention) when it is >= max_seq_len before grouping, so the layer shares the full-attention pool.

Test coverage

• Unwaives test_autodeploy_from_registry[google_gemma-3-1b-it-False].

Summary by CodeRabbit

• Bug Fixes

  • Fixed KV cache resource pool grouping for layers with large sliding windows.

• Tests

  • Resolved previously failing Gemma 3.1B autodeploy accuracy test.

[coderabbitai]

📝 Walkthrough

Walkthrough

The PR fixes KV cache pool grouping by detecting and normalizing oversized sliding_window values in KVPagedResourceHandler that cannot slide within max_seq_len, rewriting them to 0 to force full-attention pooling semantics. A previously skipped test now passes with this fix.

Changes

KVCache Pool Grouping Normalization

Layer / File(s)	Summary
KVCache sliding window normalization tensorrt_llm/_torch/auto_deploy/transform/library/kvcache.py	Detects KVPagedResourceHandler instances where sliding_window >= max_seq_len and rewrites sliding_window to 0 to ensure they use full-attention pooling instead of creating redundant separate pools.
Test waiver removal tests/integration/test_lists/waives.txt	Removes SKIP waiver for test_autodeploy_from_registry[google_gemma-3-1b-it-False], validating that the sliding window normalization fix resolves the previously failing test.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#14545: Also removes SKIP waiver entries from tests/integration/test_lists/waives.txt.
• NVIDIA/TensorRT-LLM#14789: Adds SKIP waivers including the same test entry removed in this PR.
• NVIDIA/TensorRT-LLM#14839: Adds related SKIP waiver entries in the same waivers file.

Suggested reviewers

• greg-kwasniewski1

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately and specifically describes the main change: normalizing KV windows to full attention in AutoDeploy, matching the core fix in the changeset.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
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.
Description check	✅ Passed	PR description clearly explains the issue, solution, and test coverage with specific details about the problem and fix.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

Warning

Review ran into problems

🔥 Problems

Git: Failed to clone repository. Please run the @coderabbitai full review command to re-trigger a full review. If the issue persists, set path_filters to include or exclude specific files.

---

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

[eopXD]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #51834 [ run ] triggered by Bot. Commit: b3bf607 Link to invocation

[tensorrt-cicd]

PR_Github #51834 [ run ] completed with state SUCCESS. Commit: b3bf607
/LLM/main/L0_MergeRequest_PR pipeline #41196 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

[tburt-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #51894 [ run ] triggered by Bot. Commit: b3bf607 Link to invocation

[tensorrt-cicd]

PR_Github #51894 [ run ] completed with state FAILURE. Commit: b3bf607
/LLM/main/L0_MergeRequest_PR pipeline #41252 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

[eopXD]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #52199 [ run ] triggered by Bot. Commit: b3bf607 Link to invocation

[tensorrt-cicd]

PR_Github #52199 [ run ] completed with state SUCCESS. Commit: b3bf607
/LLM/main/L0_MergeRequest_PR pipeline #41520 completed with status: 'SUCCESS'

CI Report

Link to invocation