fix(mps): stale wgp_config regeneration, Z Image float64→float32, smoke test

[huangyebiaoke]

Follow-up to #1768. Addresses remaining issues from @cn0ss's systematic MPS test report in #1750.

Changes

1. setup.py — stale wgp_config.json detection

create_wgp_config() previously skipped generation if the file existed, leaving stale configs with blank attention_mode after branch switches or upgrades. On MPS, blank attention_mode causes the model to load but produce no output files (reported by @cn0ss).

Fix: Validate existing config — if attention_mode or video_profile are missing/empty, force regeneration.

2. pipeline_z_image.py — MPS float64 → float32

Z Image Turbo was the only tested feature still failing on MPS, crashing with:

Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64

Root cause: The DPM-Solver ODE steps use torch.float64 at 5 sites. On CUDA this works; on MPS the backend rejects float64 entirely.

Fix: Detect MPS device and use solver_dtype = torch.float32. For 4-step distilled models this is numerically fine.

3. shared/mps/test_mps_smoke.py — Quick validation script

Runs in <30 seconds without model weights. Validates:

• macOS + Apple Silicon detection
• PyTorch MPS availability, float32/bfloat16 tensor ops
• Float64 rejection (MPS sanity check)
• device_patch application + nn.Parameter duck-typing
• Core model imports (wan_handler, ltx2, z_image)
• attention modes enumeration
• wgp_config.json validity
• Memory reporting

Testing

• Tested on M4 Max 36GB (if @cn0ss or @SquishedSquirrel can re-test Z Image Turbo with this branch)
• Smoke test verified to run on macOS without model weights

[huangyebiaoke]

Closing — most fixes were already in @deepbeepmeep's commit a60b152. Will contribute separately if there's new work to do.