fix(load): wire MTP sanitize-preservation patch into VLM inference path

[sangemaru]

Problem

MTP heads loaded onto VLM-path checkpoints (any model with vision_config → loads via VLMBatchedEngine, which includes all Qwen3.6 PARO models) are silently stripped at load, leaving the MTPModule at random init → 0% accept rate.

The model still generates correct output (the base path is fine), but speculative decoding is dead weight: the MTP head runs, produces garbage drafts from random-init weights, and every draft is rejected.

Root cause

mlx-vlm's stock Model.sanitize() (mlx_vlm/models/qwen3_5/qwen3_5.py) unconditionally drops every mtp.* key:

# ignore mtp weights
weights = {key: value for key, value in weights.items() if "mtp." not in key}

oMLX already has a sanitize-preservation override for exactly this — apply_mlx_vlm_mtp_patch() in omlx/patches/mlx_vlm_mtp/qwen35_vlm_model.py, which keeps mtp.* keys and remaps them to language_model.mtp.*. But it was only wired into the oQ-quantization path, never the inference load path.

maybe_apply_pre_load_patches() in omlx/utils/model_loading.py called only apply_mlx_vlm_mtp_runtime_patch() (which attaches the MTPModule to LanguageModel.__init__) — so the head module was created but then left unfilled because the stock sanitize stripped its weights before load_weights().

LLM-path checkpoints (no vision_config → BatchedEngine) were unaffected, because the mlx-lm sanitize patch (omlx/patches/mlx_lm_mtp/qwen35_model.py) already preserves MTP on that path. The bug only bit VLM-arch checkpoints that ship MTP heads.

Fix

Call apply_mlx_vlm_mtp_patch() alongside apply_mlx_vlm_mtp_runtime_patch() in the mtp_enabled VLM branch of maybe_apply_pre_load_patches(). Two-import + one-call addition; no behavior change for any path that wasn't already broken.

Verification

Tested on z-lab/Qwen3.6-27B-PARO (declares text_config.mtp_num_hidden_layers=1) with an injected BF16 MTP head (mtp.safetensors from guru87/Qwen3.6-27B-MTP, lifted verbatim from Qwen/Qwen3.6-27B):

Instrumented paroquant.inference.backends.mlx.load.load right before model.load_weights(...):

Metric	Before fix	After fix
mtp.* keys in the load dict	0	15
loaded MTP params vs source weights	random init (fc.weight std≈0.006, input_layernorm all-ones)	bit-exact (maxabsdiff=0.000000)
MTP accept rate (smoke)	0%	68.8%
32K warm accept rate	—	61.5%
64K warm accept rate	—	75.2%

Log line added on success:

mlx-vlm MTP sanitize patch applied for /path/to/Qwen3.6-27B-PARO

Note (not part of this fix)

On M1 Max, MTP remains net-negative vs no-MTP for steady-state decode regardless of this fix, because the sampler runs serially per accepted token (tracked separately in #1311). This PR is strictly about correctness — making the MTP head actually load on the VLM path. Whether MTP is a win is a separate, hardware/runtime question.

Environment: oMLX 0.3.9.dev2, M1 Max 64 GB, macOS 25.4, mlx-lm 0.31.3, paroquant[mlx] 0.1.15.