[https://nvbugs/6437341][fix] Port sibling fix 38db241414 verbatim (shutil.which short-circuit +…

[trtllm-agent]

Summary

• Root cause: _visual_gen_deps fixture unconditionally runs apt-get update/install as non-root (LOCAL_USER=1) → exit 100 → test errors at setup before LPIPS runs; separately, opencv was removed from the container image by [None][chore] Remove pre-installed opencv #16206 so the LPIPS eval script's cv2 import also fails on the video path.
• Fix: Port sibling fix 38db241 verbatim (shutil.which short-circuit + imageio-ffmpeg fallback) and add pip install opencv-python-headless to the fixture so the LPIPS eval script's lazy _get_cv2() succeeds on the T2V path.
• 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/6437341

Summary by CodeRabbit

• New Features

  • Improved Cosmos3 default generation quality with a comprehensive negative prompt to reduce unwanted visual artifacts.
  • Updated 720p generation defaults for more efficient sequence processing.
  • Streamlined cross-attention behavior for Cosmos3 generation.

• Bug Fixes

  • Improved visual generation test setup by reliably locating or installing required media-processing tools.
  • Updated the waived visual generation test coverage.

[coderabbitai]

📝 Walkthrough

Walkthrough

Cosmos3 defaults now use a shorter 720p sequence length and a populated negative prompt. Cross-attention no longer handles per-sample text lengths. Visual-generation integration setup adds portable OpenCV and ffmpeg provisioning, and the waived LPIPS test entry changes.

Changes

Cosmos3 generation and integration updates

Layer / File(s)	Summary
Generation defaults tensorrt_llm/_torch/visual_gen/models/cosmos3/defaults.py, tensorrt_llm/_torch/visual_gen/models/cosmos3/pipeline_cosmos3.py	The 720p maximum sequence length changes to 1024, and the default negative prompt is populated with video quality and artifact exclusions.
Cross-attention flow tensorrt_llm/_torch/visual_gen/models/cosmos3/transformer_cosmos3.py	Per-sample real_text_lens arguments and slicing are removed; full und/GEN key-value tensors are concatenated and processed in one cross-attention call.
Visual-generation integration setup tests/integration/defs/examples/visual_gen/test_visual_gen.py, tests/integration/test_lists/waives.txt	The fixture installs headless OpenCV, provisions ffmpeg through multiple fallbacks, and replaces the waived Cosmos3 LPIPS test entry.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Cosmos3VFMTransformer
  participant Cosmos3GenDecoderLayer
  participant Cosmos3CrossAttention
  Cosmos3VFMTransformer->>Cosmos3GenDecoderLayer: invoke without real_text_lens
  Cosmos3GenDecoderLayer->>Cosmos3CrossAttention: pass hidden states and und/GEN key-values
  Cosmos3CrossAttention->>Cosmos3CrossAttention: concatenate tensors
  Cosmos3CrossAttention->>Cosmos3CrossAttention: execute full cross-attention

Loading

Possibly related PRs

• NVIDIA/TensorRT-LLM#16139: Overlaps in Cosmos3 visual-generation fixture and waiver updates.
• NVIDIA/TensorRT-LLM#16283: Also changes integration test waiver entries.
• NVIDIA/TensorRT-LLM#16292: Also updates visual-generation-related waiver definitions.

Suggested reviewers: crazydemo, larryxfly, stanleysun639, xinhe-nv

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is clearly related to the PR and summarizes the main fix around the sibling patch and ffmpeg/OpenCV dependency handling.
Description check	✅ Passed	The description includes the required Summary, Test plan, and Links sections and explains the issue, fix, and validation clearly.
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.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

tests/integration/defs/examples/visual_gen/test_visual_gen.py (1)

245-269: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Add a small unit test for the ffmpeg fallback The fallback branch has no direct coverage; mock check_call to raise CalledProcessError and assert the imageio-ffmpeg path creates the ~/.local/bin/ffmpeg symlink.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/integration/defs/examples/visual_gen/test_visual_gen.py` around lines
245 - 269, Add a focused unit test for _ensure_ffmpeg_available that mocks
shutil.which as unavailable and check_call to raise CalledProcessError, then
verifies llm_venv.run_cmd installs imageio-ffmpeg and the fallback creates the
expected ~/.local/bin/ffmpeg symlink using imageio_ffmpeg.get_ffmpeg_exe().

Source: Path instructions

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/integration/defs/examples/visual_gen/test_visual_gen.py`:
- Around line 255-259: Broaden the exception handling around the apt-get calls
in the visual generation fixture to also catch FileNotFoundError, while
preserving the existing CalledProcessError fallback. Ensure either apt-get
failure routes to llm_venv.run_cmd(["-m", "pip", "install", "imageio-ffmpeg"]).

---

Nitpick comments:
In `@tests/integration/defs/examples/visual_gen/test_visual_gen.py`:
- Around line 245-269: Add a focused unit test for _ensure_ffmpeg_available that
mocks shutil.which as unavailable and check_call to raise CalledProcessError,
then verifies llm_venv.run_cmd installs imageio-ffmpeg and the fallback creates
the expected ~/.local/bin/ffmpeg symlink using imageio_ffmpeg.get_ffmpeg_exe().

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: f8781a73-26bb-43e4-9a9f-728c992f6f79

📥 Commits

Reviewing files that changed from the base of the PR and between 17190cf and d9ba787.

📒 Files selected for processing (5)

• tensorrt_llm/_torch/visual_gen/models/cosmos3/defaults.py
• tensorrt_llm/_torch/visual_gen/models/cosmos3/pipeline_cosmos3.py
• tensorrt_llm/_torch/visual_gen/models/cosmos3/transformer_cosmos3.py
• tests/integration/defs/examples/visual_gen/test_visual_gen.py
• tests/integration/test_lists/waives.txt

💤 Files with no reviewable changes (1)

• tests/integration/test_lists/waives.txt

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Broaden the fallback except clause to cover missing apt-get binary.

check_call(["apt-get", ...]) raises FileNotFoundError (not CalledProcessError) if the apt-get binary doesn't exist at all (e.g. non-Debian-based or apt-less containers). That currently isn't caught, so the fixture would crash instead of falling back to imageio-ffmpeg — undermining the portability goal of this fix.

🔧 Proposed fix to also handle a missing apt-get binary

     try:
         check_call(["apt-get", "update", "-y"], shell=False)
         check_call(["apt-get", "install", "-y", "ffmpeg"], shell=False)
-    except subprocess.CalledProcessError:
+    except (subprocess.CalledProcessError, OSError):
         llm_venv.run_cmd(["-m", "pip", "install", "imageio-ffmpeg"])

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	try:
	check_call(["apt-get", "update", "-y"], shell=False)
	check_call(["apt-get", "install", "-y", "ffmpeg"], shell=False)
	except subprocess.CalledProcessError:
	llm_venv.run_cmd(["-m", "pip", "install", "imageio-ffmpeg"])
	try:
	check_call(["apt-get", "update", "-y"], shell=False)
	check_call(["apt-get", "install", "-y", "ffmpeg"], shell=False)
	except (subprocess.CalledProcessError, OSError):
	llm_venv.run_cmd(["-m", "pip", "install", "imageio-ffmpeg"])

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/integration/defs/examples/visual_gen/test_visual_gen.py` around lines
255 - 259, Broaden the exception handling around the apt-get calls in the visual
generation fixture to also catch FileNotFoundError, while preserving the
existing CalledProcessError fallback. Ensure either apt-get failure routes to
llm_venv.run_cmd(["-m", "pip", "install", "imageio-ffmpeg"]).