feat(vlm): MLX engine for Nemotron-3.5-Parse #4932

Requested feature

Add the MLX engine to the nemotron_parse_v3_5 preset (nvidia/Nemotron-3.5-Parse, added in #4897), so the model runs on Apple Silicon. Today the preset supports only TRANSFORMERS and API, and #4897 says: "MLX is not supported: no MLX port exists."

State on 2026-10-21:

• No MLX conversion of Nemotron 3.5 Parse exists on Hugging Face.
• mlx-vlm has a nemotron_parse model (Add Nemotron-Parse (C-RADIO encoder + mBART decoder) model support Blaizzy/mlx-vlm#1866), which mlx-community/Nemotron-Parse-2.0-{4,8}bit and Nemotron-Parse-v1.2-* use. That port is for the 2.0 encoder-decoder (C-RADIO + mBART with cross-attention).
• 3.5 uses a different architecture: RadioMBartPrefixForCausalLM, a decoder-only mBART that reads the vision tokens as a causal prefix (is_encoder_decoder: false). It also has C-RADIOv4-SO400M, a 1D-conv adapter, dynamic NaViT preprocessing, grouped-query attention (16 query / 8 KV heads) and an MTP head. So the 2.0 port cannot load it.

Proposal:

1. Port 3.5 to mlx-vlm: extend nemotron_parse or add nemotron_parse_v3_5. The MTP head can be optional at first (plain greedy decoding).
2. Publish converted weights (bf16, 8-bit, 4-bit) with fidelity numbers against the Transformers reference.
3. Add VlmEngineType.MLX to the preset's supported_engines, with a pinned repo and revision for the MLX weights, and reuse the existing mapper nemotron_parse_v3_5_utils.py unchanged.
4. Add the auto-inline engine selection: MLX on Apple Silicon, Transformers on CUDA.

We are interested in doing the port and can contribute it.

Alternatives

• Transformers engine with the MPS backend: not tested by NVIDIA (the model card lists Linux only), and it needs trust_remote_code.
• API engine against a remote GPU server: works, but local Apple Silicon users then need a remote GPU.