[ConstBob]

Summary

• Replays the Cosmos3-Edge TensorRT-LLM Reasoner cookbook onto current main, after the generate-media skill migration in cosmos3-generate-media skill (cookbook → skill migration) with eval-derived fixes, cosmos3-skill-authoring guide, opt-in skill-eval CI #324.
• Cosmos3-Nano and Cosmos3-Super stay on nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc22.
• Cosmos3-Edge is served from a TensorRT-LLM main build at or after 39a44fab (the merge of [None][feat] Add Cosmos3-Edge Reasoner support TensorRT-LLM#20307). Release images through v1.4.0rc1 do not include Cosmos3EdgeForConditionalGeneration.
• Keeps the cosmos3-generate-media skill (cookbook → skill migration) with eval-derived fixes, cosmos3-skill-authoring guide, opt-in skill-eval CI #324 distilled-reference link, the FP8 quantization note, and the separate TensorRT-Edge-LLM sections.
• Supersedes Add Cosmos3-Edge to the TensorRT-LLM Reasoner cookbook #299.

Testing

• The notebook content has been reviewed in Add Cosmos3-Edge to the TensorRT-LLM Reasoner cookbook #299. Its process, port, and Docker checks were already run locally.

[satishra-ai]

PR #347 tested at 0eee489: PARTIAL PASS.

• Edge image/video inference: PASS with setup workarounds.
• Nano image/video regression: PASS.
• CPU checks: 43 passed, 2 failed.
• Bugs filed: 7033793, outdated build pin; 7033796, incorrect Python-path assumption. Both P2.
  Used ≤0.60 / 4 GPU-hours on GB200

@lfengad

[ConstBob]

Thanks @lfengad @satishra-ai . Both items from the review are fixed in 6cf8beb.

The Edge section now has its own checkout, git checkout 39a44fab74f4cfc0d4d498be385e746fdd182872, and says to run that instead of the git checkout line in the Generator source-build block. The Generator block still checks out cc63e66 for VisualGen. Nano and Super stay on release:1.3.0rc22. The notebook carries the same Edge checkout.

The Edge launch cell no longer requires a python executable beside trtllm-serve. It uses that interpreter when it exists, and otherwise uses the absolute interpreter in the trtllm-serve shebang, including #!/usr/bin/env python3. It does not fall back to whichever python is first on PATH. The README install command uses the same steps.

I checked these paths with local stand-ins for trtllm-serve. The serve command is unchanged, so I did not rerun the GB200 inference.

[pengcuo]

Review notes (docs/notebook only; I did not run the notebook).

1. The "release images through v1.4.0rc1" framing is stale. v1.4.0rc2 was published 2026-10-28, two days before this PR, and it is 23 commits ahead of 39a44fab. Its docs/source/models/supported-models.md lists Cosmos3EdgeForConditionalGeneration; v1.4.0rc1 does not. Edge therefore no longer needs a source build at a raw commit. I'd say "Edge requires TensorRT-LLM >= v1.4.0rc2" and, if an rc2 NGC image exists, use it. Keep the commit checkout only as a fallback. I could not check NGC from here. The same text is repeated in README.md, reasoner/README.md and four places in the notebook, so it needs updating in each.

2. reasoner/README.md quickstart changed for every tier. The TRT-LLM quickstart is shared by Nano, Super and Edge. This PR changes its prompt from "Caption the image in detail." to a one-sentence prompt, max_tokens 4096 -> 1024, and adds temperature=0, "to match the Cosmos Framework Edge example". That makes the Nano/Super quickstart worse and is unrelated to adding Edge. I'd leave the quickstart as is, or show the Edge variant separately.

3. Low: the trtllm-serve interpreter lookup. The ~25-line shell block that resolves the interpreter from trtllm-serve's shebang is copied into the README and the notebook. python -m pip install opencv-python-headless in the same environment as the source build is what the old text said and is much simpler to maintain. If the shebang case matters, one sentence would do.

[ConstBob]

Thanks @pengcuo . Addressed in 6b84b7a.

Edge now requires TensorRT-LLM >= v1.4.0rc2 and runs in nvcr.io/nvidia/tensorrt-llm/release:1.4.0rc2. v1.4.0rc2 lists Cosmos3EdgeForConditionalGeneration, and the NGC tag is published. v1.4.0rc1 and release:1.3.0rc22 do not include it. The 39a44fab checkout remains only as a fallback when that image is unavailable. Nano and Super stay on release:1.3.0rc22.

The shared TensorRT-LLM quickstart is back to "Caption the image in detail." with max_tokens=4096. The one-sentence Edge example, max_tokens=1024 and temperature=0, is a separate note.

OpenCV is installed with python -m pip install opencv-python-headless in the same environment as trtllm-serve. The shebang lookup is gone. The Edge notebook cell installs it inside the v1.4.0rc2 container, the same way Nano and Super do.