GLM 5.2 - Distributed Inference Fail

### Summary

Distributed inference across two Apple Silicon machines fails during prompt
processing with:

ds4: prompt processing failed: metal GLM layer-slice evaluation failed at pos 0

The failure happens immediately, at position 0, i.e. before any token is
generated (prefill of the very first chunk).

### Environment

- Backend: Metal
- Distributed inference: 2 machines
- Coordinator: Mac Studio M3 Ultra, 256 GB RAM
- Worker: Mac Studio M3 Ultra, 96 GB RAM
- macOS version: 27 Beta 2
- Model: GLM 5.2, quant: Q2_K

### Steps to reproduce

1. Start the worker on the 96 GB machine.
2. Start the coordinator on the 256 GB machine with the distributed route above.
3. Send any prompt (the error triggers at pos 0, so even a trivial prompt reproduces it).

### Expected behavior

Prefill of the first chunk completes and generation starts.

### Actual behavior

Prompt processing aborts at position 0:

ds4: prompt processing failed: metal GLM layer-slice evaluation failed at pos 0