SenseNova U1.5 canonical ONNX workflow
Hashless ONNX GenAI inference-metadata package for the pinned
sensenova/SenseNova-U1.5-8B-MoT revision
1f6ec60423d29939dde4202fd82ae340b144e280.
The package contains five exported components: fp16 embedding, vision encoder,
and understanding decoder, plus fp32 image-generation embedding and denoiser.
inference_metadata.yaml drives conditional/unconditional frozen KV sharing,
fp16-to-fp32 state casts, deterministic counter-based seeded noise, CFG,
x0-to-flow velocity, Euler integration, and clamped RGB output.
H200 evidence
Executed by ONNX GenAI's generic Engine::from_pipeline_dir metadata engine
with ONNX Runtime 1.28 CUDA on an NVIDIA H200:
- text prefill logits:
[1, 28, 151936], fp16 - 512x512 text-to-image, 20 steps: 6.99 s
- 512x512 reference-image edit, 20 steps: 16.68 s
- peak observed VRAM: 91,256 MiB
Requests, raw outputs, PNGs, timing manifests, and VRAM samples are under
evidence/. Production source and model license are preserved in
UPSTREAM_MODEL_CARD.md and LICENSE.
The current workflow exposes text prefill logits rather than a complete
text-token decode loop. Generic OpenAI Images/A1111 serving also needs a
path-specific image-generation chat template and automatic reference-image
placeholder mask construction; use the typed request specs in evidence/
until those server contracts are available.
Annotated inference metadata
Review inference_metadata.annotated.yaml for inline explanations of this package's workflow, tensor/state/cache contracts, and fail-closed omissions. inference_metadata.yaml remains the canonical machine-authored contract; automated validation confirms both files parse to the same metadata object.