[TRTLLM-14268][feat] Add VisualGen serving benchmark

[ishovkun]

@coderabbitai summary

Description

Add a portable online-serving benchmark for VisualGen, with one combined
server/client lifecycle wrapper and independently usable server and client
scripts.

The benchmark:

• covers Cosmos3 text-to-image, text-to-video, image-to-video,
  video-to-video, Transfer, text-to-audio-video,
  text-and-image-to-audio-video, and Action policy/forward/inverse dynamics;
• refuses to reuse an occupied server port and verifies the model exposed by a
  standalone server before sending requests;
• preserves checkpoint generation defaults unless an option is explicitly
  supplied;
• reports request throughput, per-GPU throughput, latency and generation
  distributions, diffusion time, denoising-step timing, vision decode time,
  and total pipeline time when the server exposes them;
• retains the complete run, server, health, client command, benchmark,
  metrics, result, and metadata artifacts;
• optionally saves one generated media file per measured request on the
  client while keeping server-side temporary media isolated and disposable;
• supports one or more Cosmos3 Transfer controls through a single
  TRANSFER_CONTROLS JSON mapping, including client-derived edge/blur and
  precomputed depth/segmentation/world-space-map controls;
• sends current typed image_reference and video_reference multipart
  fields while keeping INPUT_REFERENCE as the shell file-path interface;
• includes tuned Cosmos3 Super 4-GPU and 8-GPU serving configurations.

End-to-end Action validation also exposed and fixes two integration issues:

• standalone clients now accept the basename reported by /v1/models when
  the requested model is an equivalent local checkpoint path;
• Cosmos3 Action preprocessing now decodes uploaded image bytes instead of
  rejecting the typed multipart image_reference payload.

The runtime change is internal to Cosmos3 Action preprocessing. This PR does
not change the public VisualGen API, model equations, or runtime dependencies.

Test Coverage

• tests/unittest/scripts/test_benchmark_visual_gen.py
  • request payloads and typed multipart reference fields
  • all benchmark modes, Action trajectories, and Transfer controls
  • server/client standalone behavior and combined lifecycle behavior
  • occupied-port rejection and benchmark-owned media cleanup
  • metrics extraction, retained artifacts, and shell argument validation
• tests/unittest/_torch/visual_gen/test_cosmos3_action.py
  • uploaded image-byte decoding and corrupt-upload rejection
  • Action modes, embodiment defaults, conditioning, and output contracts
• tests/integration/defs/visual_gen/test_visual_gen_benchmark.py
  • reported GPU count in the online benchmark result
• tests/integration/defs/perf/test_visual_gen_perf_sanity.py
  • VisualGen performance database fields and generation-mode bucketing
• Current tip validation:
  • merged upstream/main at 3503e3f9bb; the latest upstream delta has no
    Cosmos3 or VisualGen overlap
  • native post-merge build passed (7,615 Ninja tasks)
  • focused benchmark unit suite: 77 passed
  • complete Cosmos3 Action unit suite: 126 passed
  • focused pre-commit hooks pass
  • one-request Cosmos3-Nano Action policy benchmark passed on 1x B300:
    1/1 requests, 0.8446 s diffusion, 1.1406 s generation, 1.3889 s latency
  • retained Action output verified as request-0001.safetensors containing
    video (17, 480, 832, 3), action (16, 10), and scalar frame_rate

The one-request Action run is functional validation, not an official
performance baseline.

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)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.