Spaces:
Running on Zero
fix: use eager transformer instead of incompatible AOTI blocks
After the kernel-loader pin and xlarge allocation were deployed, startup succeeded but actual camera editing still failed.
Reproduction: direct MCP call to Qwen_Image_Edit_Angles_infer_camera_edit using a Qwen-generated apple image URL, rotate_deg=45, seed=42, randomize_seed=false, with the advertised default four inference steps.
Runtime logs show the constructed 45-degree camera prompt, followed by:
CUDA driver error: no kernel image is available for execution on the device
The traceback passes through pipe.transformer -> spaces.zero.torch.aoti -> torch.export.pt2_archive and fails in AOTInductor model execution. This indicates compiled GPU code incompatibility with the allocated device.
Remove the legacy spaces.aoti_blocks_load(..., variant="fa3") call, leaving the existing transformer implementation in eager mode. Keep the xlarge GPU allocation, model weights, fused angle LoRA, camera controls, dependency pins and MCP interface unchanged. This avoids loading the incompatible compiled blocks; eager inference may be slower.
Validation: patched Python source compiles; verified AOTI loading is absent and xlarge allocation remains. GPU generation with this change has NOT yet been validated. After merge/rebuild, repeat the MCP edit with the generated image and inspect the output; further runtime issues remain possible.