[MiaAI-Lab]

What this changes

The Qwen3.5 vision tower on CUDA read only the MLX layout (vision_tower.*, patch weight [hidden, T, P, P, C]). It now also reads Hugging Face names: if model.visual.patch_embed.proj.weight exists it loads model.visual.* with the patch weight [hidden, C, T, P, P] uploaded as stored; otherwise it opens the MLX prefix exactly as before. Needed by the EXL3 checkpoint in #1150 (https://huggingface.co/Mia-AiLab/Ornith-1.5-122B-EXL3-EAQv1-4-bpw, converted straight from the HF model), but independent of it. Tracking issue: #1149.

Receipt

• Environment: TensorFold d1018b3 plus this commit; GB10, nvcr.io/nvidia/pytorch:26.07-py3, Zig 0.17.0; checkpoint Mia-AiLab/Ornith-1.5-122B-EXL3-EAQv1-4-bpw @ 9c5c913b3de1eef901520d6fedaf450044ea0423.
• Exactness: with this commit and qwen3_5 cuda: serve ExLlamaV3 (EXL3) Qwen3.5-MoE checkpoints #1150, one image request (a 672x224 PNG, temperature 0, --vision, 64 tokens) returned a coherent description identical to that from a build carrying my unsubmitted patches. The MLX-named path was not re-run on a GPU.
• Decode speed / prompt speed: n/a, the change only affects tower load.
• Tests run: zig build test 211/211 steps succeeded (with qwen3_5 cuda: serve ExLlamaV3 (EXL3) Qwen3.5-MoE checkpoints #1150 applied: 703 of 709 tests passed, 6 skipped); zig fmt --check clean; lean_check.py 444 problems before and after (the file is 598 lines).
• Not run, and why: no new unit test, because the naming choice needs a checkpoint fixture; no MLX-named vision checkpoint run on this branch.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted/resumed/concurrent equality: n/a, load-time only.
• No precision traded for speed.
• No GPU kernel added or changed.
• Prompt processing is no slower than the last release.
• Platforms: CUDA on GB10 only; Metal not run.
• New tests: none, see above.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[ashhart]

Landed on main as 167cb96 with a maintainer fix-up (23d8e05) that proves tower names and tensor bytes without a device, thank you @MiaAI-Lab. The Qwen vision CUDA tower now reads under Hugging Face's tensor names too, and the host suites pass on current main.

- Synthema, Ash's AI Assistant.

[ashhart]

This shipped in TensorFold 1.1.4: https://github.com/ashhart/TensorFold/releases/tag/v1.1.4

Thank you @MiaAI-Lab for the contribution.

- Synthema, Ash's AI Assistant.