Spaces:

---

Duplicated from  linoyts/Qwen-Image-Edit-Angles

mcp-tools

/

Qwen-Image-Edit-Angles

Like 3

Running on Zero

The Space currently fails during startup at spaces.aoti_blocks_load(pipe.transformer, "zerogpu-aoti/Qwen-Image", variant="fa3"). The MCP endpoint consequently returns HTTP 503.

Root cause: config.fa3.json references kernels-community/flash-attn3 at revision 6179fd55cfb8d901207336275e8cfc2c7b439fd1. That revision exists in the model repository (verified HTTP 200 and listed its torch28 builds using hf CLI), but installed kernels 0.16.1 explicitly requests repo_type="kernel", where the revision returns HTTP 404. This is a repository-type compatibility issue, not a deleted model-repository revision.

Changes:

• Pin kernels==0.11.0, whose download implementation uses the model repository path expected by this legacy AOTI artifact.
• Set @spaces.GPU(size="xlarge") for infer_camera_edit to request 96 GB instead of default 48 GB. This is a preventive memory adjustment for the full BF16 pipeline; the current startup failure prevents measuring this Space’s inference memory usage. xlarge consumes 2x daily GPU quota.
• Preserve the compiled blocks and existing torch==2.8 pin; no uncompiled fallback or inference behavior changes.

Validation:

• Inspected installed-version kernel loader source and the proposed pinned-version download implementation.
• Verified the referenced revision via the model endpoint and confirmed failure via the kernel endpoint.
• Patched app.py compiles syntactically.
• No end-to-end GPU generation has yet been verified with this patch. After merge/rebuild, check startup and invoke the camera-edit MCP tool with a sample image. Further AOTI/runtime compatibility issues may still need resolution.