1Cat-vLLM/docs/design/native_media/CREATIVE_JOBS.md

Native creative jobs

This change adds a native single-device Z-Image FP16 sampler and asynchronous image jobs alongside the existing H3 video engine. Studio remains responsible for authenticated asset storage and model lifecycle confirmation.

• Image CLI: vllm image --model /local/model --checkpoint z-image-turbo --output-dir /local/output.
• Image jobs: POST /v1/images/jobs, GET /v1/images/jobs/{id}, content at /v1/images/jobs/{id}/content. Synchronous /v1/images/generations remains compatible with URL and base64 results. Queued cancellation is supported; running jobs finish and retain outputs.
• H3 jobs now report stage, stage_started_at, updated_at, stage_progress and denoise_progress. Counts are sampler iterations, not output counts. Existing top-level video progress retains its legacy output-count meaning.
• POST /v1/videos/{id}/cancel never deletes an already completed output.
• Progress crosses the existing worker pipe on rank zero. Reporting introduces no tensor transfer, device synchronization, or additional collective.

Recipes and precision

Z-Image's tokenizer, Qwen encoder, diffusion transformer, scheduler and VAE are loaded exclusively from a local directory downloaded and verified by ModelScope. No remote model code or automatic Hub fallback is enabled. The transformer and text encoder use FP16 weights; latents, classifier-free guidance and VAE decoding use FP32. Projection accumulators and SwiGLU gating retain FP32 range before normalization. Base additionally keeps attention, AdaLN scaling and residuals in FP32 because its learned modulation exceeds FP16 range. It retains FP16 weights and fits one V100 32 GB, but is slower than Turbo. No outlier clamping is used. Turbo uses eight nonzero sampler updates; base uses fifty with CFG 4. Component implementations are pinned to Diffusers 0.40.0 and Transformers 5.15.1. The reference sampler is Tongyi-MAI/Z-Image commit 26f23eda626ffadda020b04ff79488e1d72004cd (Apache-2.0).

Image editing is deliberately outside this capability. The model registry must not advertise image editing, unsupported output sizes or unverified GPU quality.

Acceptance — 2026-09-09

53 CPU media tests pass, covering protocol compatibility, real step counting, queued/running cancellation, restart idempotency, FP16 outlier handling, corrupt saved records and disk failure without a stalled queue. Pre-commit and CI pass.

Actual Studio frontend runs on V100 SXM2 32 GB / PyTorch 2.10 / CUDA 12.8 produced:

• Z-Image Turbo and Base: 1024 × 1024 PNG, seed 42, respectively 8 and 50 updates.
• H3 Turbo 4: text, first/last-frame and image-reference videos, each 1344 × 768, 107 frames at 24 fps (4.458333 s), with synchronous 32 kHz stereo audio.
• The image run uses one explicitly selected V100; H3 uses four. Display GPU is excluded. ABI-matching SM70 extensions from b6d91d61ff are reused; this PR changes no CUDA/C++ source and makes no power-policy changes.

Z-Image Base now retains FP32 attention/modulation/residual range after measured FP16 overflow in layer 25 Q/K/V and later AdaLN scaling. It fits one 32 GB V100, but its 50-update original recipe takes roughly ten minutes in this setup. Turbo is the default for interactive creation. Images were visually checked; video playback/download and keyframe/reference consistency were checked in UI.

Progress parity with fixed inputs/recipes/seeds:

Output	Enabled/disabled reporting
Z-Image Turbo PNG	Byte-identical (a08ca43be383426e…)
Z-Image Base PNG	Byte-identical (1d9db6d2e332c65a…)
Warm H3 keyframe MP4	Byte-identical (fb9dc41853af7778…), including raw audio

First-run H3 audio has a maximum 5.44e-7 sample difference from warm runs, also present with reporting enabled in both runs; enabled/disabled warm outputs match exactly. Video frames match in all three cases. These are correctness checks, not a throughput benchmark or exhaustive validation of every allowed output dimension/duration.

Weights and component code came from the checksum-verified ModelScope catalog: Tongyi-MAI/Z-Image-Turbo, Tongyi-MAI/Z-Image, MiniMax/MiniMax-H3, Comfy-Org/MiniMax-H3, and lightx2v/Minimax-h3-Turbo. The original model revision and complete manifest hashes are retained in deployment acceptance records. No Hub fallback is used by the native image path. Image editing remains disabled.