fix(vl): strip model.visual.* in qwen2_vl/qwen3_vl/qwen3_vl_moe sanitize

[Jonathangadeaharder]

Problem

HuggingFace ForConditionalGeneration VL checkpoints (Qwen3_5MoeForConditionalGeneration, Qwen3VLForConditionalGeneration, Qwen2VLForConditionalGeneration) nest the language model and vision tower under a top-level model dict:

model.language_model.*   (language model weights)
model.visual.*           (vision tower weights)

The existing sanitize() helpers only popped bare visual / vision_tower top-level keys — which matches the mlx-vlm conversion layout but not the HF layout. On HF checkpoints:

• qwen3_vl_moe.sanitize crashes with KeyError: 'model' because weights["language_model"]["model"] doesn't exist (the language model is nested under model.language_model.model.*, so after tree_unflatten it lives at weights["model"]["language_model"]["model"], not weights["language_model"]["model"]).
• qwen3_vl and qwen2_vl leak model.visual.* into the sanitized weights, re-prefixed as language_model.model.visual.* — causing "parameters not in model" errors downstream.

Reproducer

from mlx.utils import tree_unflatten
flat = [
    ("model.visual.pos_embed.weight", None),
    ("model.language_model.model.layers.0.self_attn.q_proj.weight", None),
    ("model.language_model.lm_head.weight", None),
]
tree = tree_unflatten(flat)
# top keys: ["model"] — no bare "visual" or "language_model"
tree.pop("visual", None)  # no-op
tree["language_model"]["model"]  # KeyError: "model"

Fix

After tree_unflatten, pop the top-level model dict, re-root its language_model subtree to the top level, and drop its visual / vision_tower subtrees. Three model classes changed identically:

• mlx_lm/models/qwen3_vl_moe.py
• mlx_lm/models/qwen3_vl.py
• mlx_lm/models/qwen2_vl.py

Tests

TestVLSanitize in tests/test_models.py covers all three model classes against the HF ForConditionalGeneration layout (9-key fixture mirroring Ornith-1.0's real weight_map). Each test asserts:

1. sanitize() does not crash,
2. zero visual / vision_tower keys remain,
3. the language model tensors survive, re-rooted under language_model.*.

All three fail on main (Red), pass after the fix (Green). Full tests/test_models.py suite: 75 passed, 1 skipped (pre-existing test_ssm failure unchanged, unrelated).

Verification

Verified end-to-end on deepreinforce-ai/Ornith-1.0-35B (333 model.visual.* tensors, full vision_config): with this fix, the model loads cleanly where it previously crashed with KeyError: 'model' during sanitize().