Zing-0.5 — AOTInductor artifacts
Ahead-of-time (torch.export + AOTInductor) compiled WanAttentionBlock, the 30×
repeated transformer block of seedleap/zing-0.5.
WanAttentionBlock/package.pt2
That is the layout spaces.aoti_blocks_load
expects, so it is consumed with one line:
import spaces
spaces.aoti_blocks_load(pipeline.generator, "hugging-apps/zing-0-5-world-model-aoti")
Weights are not baked into the archive — spaces.aoti_patch feeds each block its
own state_dict() at call time, so this single ~1 MB artifact serves all 30 layers.
Three axes are dynamic: sequence (tokens per generated block, i.e. resolution),
history (KV-cache length) and context (prompt length).
- Produced by
hugging-apps/zing-0-5-world-model-aoti-compile - Consumed by
hugging-apps/zing-0-5-world-model
Built against PyTorch 2.11 / CUDA 13 / spaces 0.51 on ZeroGPU (A10G-class,
sm_80+ via the ZeroGPU Blackwell wheels) and against the vendored zing_v0_5
package as it ships in those two Spaces — it is not portable to a different model
revision or a different PyTorch build.