[Apple M5 Max] Z-Image-Turbo (Metal): tensor-API path silently outputs blank white PNGs (exit 0) — GGML_METAL_TENSOR_DISABLE=1 restores correct output

[ziyu4huang]

Summary

On Apple M5 Max (macOS, Xcode 26 SDK), generating with Z-Image-Turbo GGUF via the Metal backend produces completely blank white images when the Metal tensor API is active — with exit code 0, no errors, and normal-looking speed. Setting GGML_METAL_TENSOR_DISABLE=1 restores fully correct output.

Environment

• stable-diffusion.cpp @ cc515a0 (master, 2026-09-17), ggml submodule = leejet/ggml @ e20c3a1 (branch sd.cpp)
• Apple M5 Max, 128GB unified memory, macOS (Xcode 26 SDK)
• Model: leejet/Z-Image-Turbo-GGUF (Q8_0 and Q4_K both affected), VAE ae.sft, TE Qwen3-4B-Instruct-2507

Repro

# (a) tensor API active → blank white PNG, rc=0, ~20s @ 864×480:
build/bin/sd-cli --diffusion-model weights/z_image_turbo-Q8_0.gguf \
  --vae weights/ae.sft --llm weights/Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
  -p "a cinematic photo of a taxi driver at night" --cfg-scale 1.0 --steps 8 \
  --diffusion-fa -W 864 -H 480 -s 42 -o out.png -v

# (b) same command + GGML_METAL_TENSOR_DISABLE=1 → correct photoreal image (~28s):
GGML_METAL_TENSOR_DISABLE=1 build/bin/sd-cli ...

• Verbose log (a) shows tensor API self-tests passing ("testing tensor API for f16/bfloat support", dummy kernels compile) — no error anywhere; the failure is silent.
• --diffusion-fa on/off makes no difference (both blank without the env guard).
• Same seed two runs are byte-identical (deterministically blank — so downstream checks on file size/dims won't catch it).

Notes

• We initially suspected the MTLCompileOptions.languageVersion issue fixed upstream in llama.cpp PR #27461 and patched options.languageVersion = MTLLanguageVersion4_0 for tensor compiles — the patch changed nothing (still blank). So this looks like a compute-kernel correctness issue in the Metal tensor path on M5, not shader compilation.
• MiniMax-H3 (vid_gen) is NOT affected — H3 runs fine with the tensor API enabled on the same machine/build (verified: correct frames + audio, 22f @ 864×480 in ~2 min with steps 20).
• Evidence set (blank vs correct outputs, A/B logs, timings) can be provided on request.

Workaround

GGML_METAL_TENSOR_DISABLE=1 for Z-Image workloads on M5 until the tensor kernels are fixed.

Thanks for the great project — day-1 Z-Image and H3 support made this whole evaluation possible.