[None][fix] Fix stale sparse attention kwargs

[bobboli]

Summary

1. Fix the stale sparse-attention kwarg in the encoder-decoder cross KV-cache manager path.
2. Fix the encoder attention metadata path to pass lowered sparse_metadata_params instead of the user-facing sparse config.

Details

PR #14687 renamed and split sparse-attention plumbing so AttentionMetadata consumes SparseMetadataParams, while KV-cache manager construction uses sparse_attention_config.

Two stale call sites remained on main:

1. KvCacheCreator._create_cross_kv_cache_manager() still passed sparse_attn_config=None to _create_kv_cache_manager(), whose parameter is now sparse_attention_config.
2. PyTorchModelEngine._prepare_tp_inputs_encoder() still passed sparse_attention_config=self.sparse_attention_config into self.attn_backend.Metadata(...), but TrtllmAttentionMetadata now expects sparse_metadata_params.

This fixes both reported pre-merge CI failures:

• _create_kv_cache_manager() got an unexpected keyword argument 'sparse_attn_config'
• TrtllmAttentionMetadata.__init__() got an unexpected keyword argument 'sparse_attention_config'

Test Plan

1. git diff --check
2. python3 -m py_compile tensorrt_llm/_torch/pyexecutor/_util.py tensorrt_llm/_torch/pyexecutor/model_engine.py
3. pre-commit run --files tensorrt_llm/_torch/pyexecutor/_util.py tensorrt_llm/_torch/pyexecutor/model_engine.py
4. Not run locally: pytest tests/unittest/_torch/executor/test_dual_pool_kv_cache.py -q, because the local Python environment does not have pytest installed.

Summary by CodeRabbit

• Refactor
  • Updated sparse attention configuration handling in encoder-decoder models for improved internal consistency.

[coderabbitai]

Caution

Review failed

An error occurred during the review process. Please try again later.

📝 Walkthrough

Walkthrough

Two small fixes adjust sparse attention handling in encoder-side paths: the cross-attention KV cache manager is now constructed with sparse_attention_config=None, and the encoder attention metadata constructor receives the derived sparse_metadata_params (via to_sparse_metadata_params) rather than the raw sparse_attention_config object.

Changes

Sparse Attention Encoder Path Fixes

Layer / File(s)	Summary
Encoder sparse attention config and cross-KV cache fixups tensorrt_llm/_torch/pyexecutor/_util.py, tensorrt_llm/_torch/pyexecutor/model_engine.py	_create_cross_kv_cache_manager passes None for sparse_attention_config instead of forwarding sparse_attn_config. In _prepare_tp_inputs_encoder, sparse_metadata_params is computed via sparse_attention_config.to_sparse_metadata_params(...) and passed to the encoder attention metadata constructor, replacing the prior direct sparse_attention_config argument.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~5 minutes

Suggested reviewers

• heyuhhh
• yuxianq

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[bobboli]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #54860 [ run ] triggered by Bot. Commit: c0a3c44 Link to invocation

[tensorrt-cicd]

PR_Github #54860 [ run ] completed with state FAILURE. Commit: c0a3c44
/LLM/main/L0_MergeRequest_PR pipeline #43865 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

CI Agent Failure Analysis

Link to invocation

[yuxianq]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #54879 [ run ] triggered by Bot. Commit: c0a3c44 Link to invocation

[tensorrt-cicd]

PR_Github #54879 [ run ] completed with state SUCCESS. Commit: c0a3c44
/LLM/main/L0_MergeRequest_PR pipeline #43884 completed with status: 'SUCCESS'

CI Report

Link to invocation