[None][fix] Pass max_seq_len to the Qwen-VL attention helper in the MiniCPM-V 4.6 vision encoder

[David-Wu1119]

Description

MiniCPMV4_6VisionModel._make_attn_metadata builds the vision encoder's attention metadata with the shared Qwen-VL helper:

return _prepare_qwen_vl_vision_attn_metadata(seq_lens, attn_metadata)

MiniCPM-V 4.6 support (#15976, 2026-07-28) was written against the helper's original (seq_lens, attn_metadata) signature. A week later #16568 made max_seq_len a required keyword-only argument and updated the Qwen2-VL and Qwen3-VL callers, but not this one. Since then, every MiniCPM-V 4.6 image or video request fails in the vision encoder, both before and after the ViT merger layer:

TypeError: _prepare_qwen_vl_vision_attn_metadata() missing 1 required keyword-only argument: 'max_seq_len'

The fix passes max_seq_len=max(seq_lens). Before #16568, the helper set exactly this value (attn_metadata.max_seq_len = max(seq_lens)), so the vision encoder behaves as it did when MiniCPM-V 4.6 support landed. I kept the pre-#16568 value instead of picking a fixed capacity. This metadata is sized per call (max_num_tokens=sum(seq_lens) + 1), and I didn't want to guess a capacity for MiniCPM. If you want MiniCPM to get the same stable attention-op cache key that #16568 gave Qwen, a fixed value can be a follow-up.

Test Coverage

• New test_vision_attn_metadata_passes_max_seq_len in tests/unittest/_torch/modeling/test_modeling_minicpmv4_6.py, which is already in l0_a10. It follows the existing test_qwen3_vision_prepare_metadata_passes_fixed_max_seq_len: it replaces the helper with one that has the same signature, calls _make_attn_metadata, and checks the arguments. On main the call raises the TypeError above; with this change it passes.
• I don't have a CUDA machine, so I couldn't import tensorrt_llm locally. Instead I pulled _make_attn_metadata and the real _prepare_qwen_vl_vision_attn_metadata out of the source and ran them with a stub metadata class:

main  -> TypeError: _prepare_qwen_vl_vision_attn_metadata() missing 1 required keyword-only argument: 'max_seq_len'
fix   -> max_seq_len = 64, cu_q_seqlens = [0, 64, 80], prepare_encoder_only() called

• pre-commit run passes on both files (ruff, ruff-format, codespell, test-list checks).

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.

I found this with pylint's missing-kwoa (E1125) check. The change was written with AI assistance and I reviewed every line.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: NVIDIA/TensorRT-LLM/.coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d62d80f5-c2f6-4168-9339-aadc7ccd8761

📥 Commits

Reviewing files that changed from the base of the PR and between b88149e and 2d49523.

📒 Files selected for processing (2)

• tensorrt_llm/_torch/models/modeling_minicpmv4_6.py
• tests/unittest/_torch/modeling/test_modeling_minicpmv4_6.py

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

Walkthrough

MiniCPMV4_6VisionModel._make_attn_metadata now passes the maximum sequence length to the vision attention metadata helper. A unit test checks that the helper receives the sequence lengths and their maximum.

Changes

Vision attention metadata

Layer / File(s)	Summary
Pass and verify maximum sequence length tensorrt_llm/_torch/models/modeling_minicpmv4_6.py, tests/unittest/_torch/modeling/test_modeling_minicpmv4_6.py	_make_attn_metadata passes max(seq_lens) as max_seq_len. The test checks the helper arguments for sequence lengths [64, 16].

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~5 minutes

Change: Bug fix

Suggested reviewers: bowenfu

Merge Risk: ⚪ Minimal · up to 2d495

This fixes a crash in MiniCPM-V 4.6 image/video requests by passing a required argument to a shared helper; the fix is small, targeted, and tested.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 5 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
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.
Title check	✅ Passed	The title clearly identifies the fix, the affected MiniCPM-V 4.6 vision encoder, and the missing max_seq_len argument. It follows the required [None][fix] format.
Description check	✅ Passed	The description explains the issue, the fix, test coverage, validation results, and checklist status. It provides sufficient context and is largely complete for this change.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

---

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