model : add Ling 3.0 VL support

[aetherbird]

Overview

Adds support for Ling-3.0-flash-VL, the vision-language variant of Ling 3.0 Flash (124B total / 5.1B active, hybrid KDA + gated MLA, 512-expert MoE). The text backbone is identical to bailingmoe3; the release adds a 27-block vision tower (qwen3_moe_vit family), a norm-only 2x2 merger, a two-layer projector, and uses M-RoPE with sections [8, 12, 12] shared between text and vision positions.

Additional information

• test-llama-archs --arch bailingmoe3 passes (MoE fixture round-trip).
• Converted the full model to BF16 GGUF (917 tensors, 248.9 GB) and mmproj (334 tensors) from the released safetensors.
• End-to-end image + video inference on llama-server successful.

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: Yes, I used AI to help write and check the code. I personally, thoroughly reviewed the CONTRIBUTING.md file.

[aetherbird]

Reworked per review in 7fcea1b. "bailingmoe3vl" has been removed, and VL now converts under bailingmoe3 in conversion/bailingmoe3.py (text model subclasses BailingMoeV3Model, vision tower subclasses Qwen3VLVisionModel).

• mrope_section is written as rope.dimension_sections, M-RoPE is selected via use_mrope() when sections are present, and text-only files will keep NORM rope.
• Projector type renamed to ling3vl.
• The mtmd graph uses build_vit with a rope callback instead of a hand-written ViT loop.
• num_shared_experts fallback has been moved into BailingMoeV3Model and the num_nextn_predict_layers default has been removed.

Validation: test-llama-archs --arch bailingmoe3 passes on the M-RoPE fixture. End-to-end: a converted GGUF produces byte-identical greedy output (text and image prompts) before and after the fold.

[CISC]

Fix mrope CI failure. :)

[ngxson]

hmm, did your agent accidentally changed the title of this PR?

[aetherbird]

hmm, did your agent accidentally changed the title of this PR?

No. I re-titled this PR to match the intention of the rework that you and CISC requested.

[aetherbird]

BailingMoeV3VLModel was missing the model_arch declaration that TextModel.__init_subclass__ requires. This has been fixed in 0d9cb8e.

The converter has now been validated end-to-end on a synthetic VL checkpoint.

[ngxson]

"fold" sounds weird. original title is already clear enough

many PRs in the past already doing exactly what you are doing here. you should have a look at them first

[aetherbird]

"fold" sounds weird. original title is already clear enough

many PRs in the past already doing exactly what you are doing here. you should have a look at them first

Changed.

[aetherbird]

• Re-ran the converter against the released inclusionAI/Ling-3.0-flash-VL checkpoint.
• Re-validated video input decode (the model described video input correctly).
• Re-validated that a 76k token prompt processes and generates cleanly.

[CISC]

@aetherbird re #29151 (review), see f.ex. https://github.com/ggml-org/llama.cpp/actions/runs/35756040134/job/106940559614?pr=29151#step:3:1262

[aetherbird]

@aetherbird re #29151 (review), see f.ex. https://github.com/ggml-org/llama.cpp/actions/runs/35756040134/job/106940559614?pr=29151#step:3:1262

Fixed in 8ef43a4 by giving bailingmoe3 its own case after the list's return LLAMA_ROPE_TYPE_NORM.

The bailingmoe3 mrope gate was appended to the shared NORM fall-through list in llama_model_rope_type(), so every arch in that list inherited the use_mrope() check. deepseek32 got M-RoPE positions and its ggml_rope_ext call hit the assert.

Verified on a build of the PR merged with current master: test-generate-models generates all fixtures with no asserts and test-llama-archs --arch bailingmoe3 still passes.