GLM-OCR fails to load — "Failed to load model" on all configurations #337

GLM-OCR fails with Failed to load model on every configuration I try. The model uses glm4 architecture and was added to llama.cpp in PR ggml-org/llama.cpp#19677 (merged Feb 18, 2026, build ~b8096). Since llama.rn 0.12.0-rc.9 syncs to b8827, I expected the architecture to be supported.

Configs tried (all fail with same error):

• with vision

  {"model": "GLM-OCR-Q8_0.gguf", "n_ctx": 6144, "n_gpu_layers": 99, "use_mlock": true, "ctx_shift": false}

• no ctx_shift

  {"model": "GLM-OCR-Q8_0.gguf", "n_ctx": 6144, "n_gpu_layers": 99, "use_mlock": true}

• CPU only

  {"model": "GLM-OCR-Q8_0.gguf", "n_ctx": 6144, "n_gpu_layers": 0, "use_mlock": true}

• bare minimum

  {"model": "GLM-OCR-Q8_0.gguf", "n_ctx": 2048, "n_gpu_layers": 0, "use_mlock": false}

Model loading failed on all attempts. No additional error details from the native layer.

• Is LLM_ARCH_GLMOCR (and PROJECTOR_TYPE_GLM4V) compiled into the iOS binary?

• Does GLM-OCR's Multi-Token Prediction (MTP) require any special build flag?

• Is there anything else needed to enable this architecture on iOS/Metal?