Z-Image inference gets stuck on Loading transformer

[timothybroome]

iris.c % ./iris -d zimage-turbo -p "a fish" -o fish.png                      
MPS: Metal GPU | Apple M4 | 10 cores
Seed: 1771362514
Loading VAE... done (0.0s)
Model: Z-Image-Turbo-6B v1.0 (zimage, 9 steps, guidance 0.0)
Loading Qwen3 encoder... done (1.3s)
Encoding text... done (16.5s)
Loading Z-Image transformer...

Hardware 2025 M4 Macbook Air

Model downloaded as-per readme documentation:

pip install huggingface_hub && python download_model.py zimage-turbo

[martin-frbg]

On my M4 mini (16GB) the mps-enabled iris eventually gets killed in this stage, probably due to running out of memory (exit code is 137).
Both "blas" (Accelerate) and "generic" builds work on this hardware but produce weird mosaics, same as on x86_64.

Loading Z-Image transformer...Process 11519 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGKILL
    frame #0: 0x000000010004f164 iris`get_cached_bf16_as_f16_buffer(weights=0x00000078f7c00000, num_elements=39321600) at iris_metal.m:1047:35
   1044         return nil;
   1045     }
   1046     for (size_t i = 0; i < num_elements; i++) {
-> 1047         f16_data[i] = bf16_to_f16(weights[i]);
   1048     }
   1049
   1050     size_t size = num_elements * sizeof(uint16_t);
Target 0: (iris) stopped.
(lldb)