GLM-4.7 PRISM loads and passes inspect after fixing num_nextn_predict_layers, but MTP generation crashes with AttributeError: LanguageModel has no attribute fa_idx

[godmonero]

mtplx doctor --json

Hi, I’m testing MTPLX 0.3.5 on a Mac Studio M3 Ultra (512 GB RAM) installed via Homebrew.

Environment

• MTPLX 0.3.5
• Apple M3 Ultra
• 512 GiB unified memory
• Python 3.13.13 arm64
• MLX 0.31.2
• mlx_lm 0.31.3

Doctor
I ran mtplx doctor --json and the environment looks healthy overall:

• native arm64 Python: pass
• MLX import/device: pass
• estimated runtime memory well below available memory
• no low power mode
• no thermal warnings
  Only warnings are unrelated (default Qwen cache missing, port 8000 already open, no ThermalForge).

Model
Local model:
/Users/macstudio/.lmstudio/models/mlx-community/GLM-4.7-PRISM-8bit-gs64-mlx

Initial behavior
The model originally showed:

• No MTP head · Glm4MoeForCausalLM

I found that:

• the converted config had "num_nextn_predict_layers": 0
• the original BF16 model had "num_nextn_predict_layers": 1

After adding mtp.safetensors and changing:
"num_nextn_predict_layers": 1

mtplx inspect changed to:

• recognized: true
• can_run: true
• runtime_compatibility: native-family-gated
• mtp_layers: 1
• mtp_tensors_present: 502

So inspect now considers the model runnable.

Runtime result
The model loads successfully in sustained profile and enters:

• Generation mode: MTP
• Native-MTP speed path: draft-only LM head is active

But on the first prompt, generation crashes with:

AttributeError: 'LanguageModel' object has no attribute 'fa_idx'

Stack trace points into:

• generate_mtpk
• forward_ar_capture
• forward_with_gdn_capture
• gdn_capture.py

where it accesses:
cache[inner.fa_idx]

Additional observation
If I run with /mtp off, the model does generate, but it is extremely slow and GPU usage is unstable compared with direct MLX inference.

Question
Is GLM-4.7 PRISM / Glm4MoeForCausalLM expected to work in MTPLX 0.3.5 with MTP, or is this currently a known issue in the GLM runtime path?

doctor_output.json

Exact command

mtplx start --model /Users/macstudio/.lmstudio/models/mlx-community/GLM-4.7-PRISM-8bit-gs64-mlx

Model path or repo id

/Users/macstudio/.lmstudio/models/mlx-community/GLM-4.7-PRISM-8bit-gs64-mlx

Chip, RAM, macOS version

Apple M3 Ultra, 512 GB RAM, macOS 26.4.1