Wan 2.2 Workbench support

Use this skill for the public Wan 2.2 registry candidate and its verified video evidence. Read these files before changing behavior:

• workflows/testing/byof-wan2.2.yaml
• workflows/testing/byof-wan2.2-multigpu.yaml
• npa/src/npa/workflows/wan_rerun.py
• docs/workbench/wan2.2.md

Also load byof-onboard, oss-solution-registry-onboard, author-npa-workflow, real-components, solution-licensing, gpu-selection, nebius-infra, testing-conventions, npa-agent, and agent-visual-feedback when their surfaces are involved.

Ground truth

• Official source: https://github.com/Wan-Video/Wan2.2.git, pinned to 42bf4cfaa384bc21833865abc2f9e6c0e67233dc.
• Official model: Wan-AI/Wan2.2-TI2V-5B, pinned to 921dbaf3f1674a56f47e83fb80a34bac8a8f203e.
• TI2V-5B is a stock generative-video model supporting text and image inputs.
• A historical operator-only validation record accepted the real single-GPU text-to-video path on RTX PRO 6000 Blackwell (sm_120) from immutable image digest sha256:1baa4e2e89999ea26df81891ac786fa99c7498cbf173e5c5abad54c6f1dd1d13, including exact MP4/RRD byte identity.
• A historical operator-only validation record accepted one shared official generation from that same observed image digest on four B200s (sm_100) with world size 4, NCCL, T5 and DiT FULL_SHARD FSDP, Ulysses size 4, and exact MP4/RRD byte identity.
• Those records used Torch 2.7.1/CUDA 12.8 and NCCL 2.27.7. The current acceptance gate is Torch 2.13.0/CUDA 13.0 and NCCL 2.29.7; it requires fresh operator-accepted single- and four-GPU evidence before publication.
• I2V, A14B, speech-to-video, Animate, and training are separate capabilities.
• Stock Wan does not predict robot actions. Never claim that it is action-conditioned.

For changing facts, use only the official Wan repository, official Wan-AI model cards, and primary framework documentation.

Packaging contract

Use workbench.byof.repo; do not add a fake Wan toolRef. Keep the repo and all model inputs immutable. The image may contain pinned source and dependencies but no checkpoint weights, credentials, private code, or user data. The runtime must remain non-root, with /opt/byof and its venv readable and executable.

The single-GPU baseline requests one RTX PRO 6000 Blackwell (sm_120), uses the security-fixed PyTorch 2.13.0 CUDA 13.0 wheel line, and binds pinned Wan attention to native PyTorch SDPA instead of FlashAttention. Record the device, compute capability, driver, CUDA, torch version, compiled arch list, and finite SDPA probe.

The distributed spec uses byof-solution-smoke-wan22-b200-4gpu.yaml and exactly four ranks. Invoke the official path with:

/opt/byof/.venv/bin/python -m torch.distributed.run --standalone \
  --nnodes=1 --nproc_per_node=4 wan22_distributed_wrapper.py
# The generated wrapper instruments all ranks, then executes pinned official
# /opt/byof/generate.py as __main__ with --dit_fsdp --t5_fsdp --ulysses_size 4.

Fail closed unless every rank proves NCCL initialization/all-reduce, a unique local B200, T5 and DiT FULL_SHARD wrappers, live Ulysses distributed-attention and all-to-all calls, the upstream final barrier, observer terminal synchronization, and compute capability 10.0 with sm_100 support.

The single-GPU smoke writes:

• wan2_2_ti2v_5b.mp4
• wan2_2_ti2v_5b_text_to_video.json
• wan2_2_runtime_inventory.json

The distributed smoke additionally writes:

• wan2_2_ti2v_5b_multigpu.mp4
• wan2_2_ti2v_5b_multigpu.json
• wan2_2_multigpu_topology.json
• wan2_2_multigpu_runtime_inventory.json
• wan2_2_multigpu_rank_0.json through wan2_2_multigpu_rank_3.json

Decode every frame and fail on invalid dimensions/count/FPS, a corrupt or empty container, an implausibly small file, or uniform content. Keep capabilities_exercised exact and deferred empty for hard-gated runs.

Rerun evidence contract

Every successful named Wan solution smoke is postprocessed by npa.workflows.wan_rerun. The postprocessor runs after the existing BYOF S3 upload and must fail the parent command if source validation, RRD generation, local parsing, rerun rrd verify, upload, S3 byte verification, remote parsing, or manifest verification fails.

Use Rerun SDK 0.31.4, matching the agent-compatible npa[viz] extra. Embed the exact MP4 at /wan2_2/video/asset, log one timestamped VideoFrameReference per decoded frame at /wan2_2/video/frame, and use the video_time duration timeline. Static JSON facts belong in the summary, validation, runtime, distributed, rank, and metric entities; do not invent a time series.

The distributed filenames are:

• wan2_2_ti2v_5b_multigpu.rrd
• wan2_2_ti2v_5b_multigpu_rrd_manifest.json

The manifest must contain source object URIs plus ETags, byte sizes, and SHA-256 values; RRD URI/hash/size/version/entities; embedded-video identity; and local plus remote verification. Only a successfully uploaded and remotely verified manifest may name wan2.2_verified_rerun_recording.

Capability status

Capability	Status
wan2.2_ti2v_5b_text_to_video	accepted current evidence; exact public-dev digest ran the Torch 2.13.0/CUDA 13.0 closure on RTX PRO 6000
wan2.2_decoded_mp4_validation	accepted current evidence; 17 decoded 1280×704 frames at 24 fps
wan2.2_ti2v_5b_text_to_video_multigpu_fsdp_ulysses	accepted historical evidence; current runtime needs a fresh 4×B200 official run
wan2.2_distributed_rank_topology_validation	accepted historical evidence; four unique ranks/devices and collective/barrier evidence
wan2.2_verified_rerun_recording	accepted current single-GPU evidence; exact MP4 identity and uploaded RRD were independently re-verified
wan2.2_ti2v_5b_image_to_video	deferred
A14B / S2V / Animate	deferred
official TI2V fine-tuning	deferred; no pinned-source entrypoint
stock Wan action prediction	rejected as an upstream capability

Licensing

Track official source, baked dependencies, runtime-fetched CUDA software, run-time model/tokenizer, and data separately. Source/model declarations do not classify a built image. The promoted first-class npa-wan2-2 contract is public eligible only when a pushed digest proves all nvidia-*, CUDA/cuDNN/NCCL, checkpoint, credential, and cache bytes absent from every layer and history via npa/scripts/scan_image_wan_payload.py. CUDA/PyTorch installation and use remain governed by the upstream package terms; NPA adds no per-run consent variable. Model/tokenizer acquisition remains runtime-only. Do not treat access as permission beyond the applicable licenses and never publish merely because the Dockerfile looks clean. HF_TOKEN is optional for public assets and remains a submission secret when supplied.

Validation

Use the repository venv, never bare Python:

npa/.venv/bin/npa workbench workflow validate-spec \
  workflows/testing/byof-wan2.2.yaml
npa/.venv/bin/npa workbench workflow plan-spec \
  workflows/testing/byof-wan2.2.yaml --run-id wan22-plan
npa/.venv/bin/npa workbench workflow validate-spec \
  workflows/testing/byof-wan2.2-multigpu.yaml
npa/.venv/bin/npa workbench workflow plan-spec \
  workflows/testing/byof-wan2.2-multigpu.yaml \
  --run-id wan22-multigpu-plan
npa/.venv/bin/python -m pytest npa/tests/workflows/test_wan_rerun.py -q
npa/.venv/bin/python -m pytest npa/tests/workflows/test_byof_solution_smokes.py -q
npa/.venv/bin/python -m pytest npa/tests/guardrails/test_skills_index.py -q
npa/.venv/bin/python -m pytest npa/tests/smoke/test_all_workflow_yamls.py -q

The gated live tests are npa/tests/e2e/test_byof_wan22_live_e2e.py and npa/tests/e2e/test_byof_wan22_multigpu_live_e2e.py. Future compatibility changes require fresh live evidence rather than inference from an older run.