[void-mckenzie]

Part of #671 (the Qwen3-VL item in #794). Serves image input on qwen3vl (dense) and qwen3vlmoe GGUFs from their qwen3vl_merger mmproj on the #803 base: the one projector with DeepStack merger tensors. The Qwen3.5 / 3.6 / 3.8 / Flash-Next towers sharing this code still refuse a DeepStack mmproj.

What it does

• gguf_mm_config derives deepstack_visual_indexes from the v.deepstack.{block}.* tensors; merger k = tap rank, as the converter writes it (deepstack_merger_list.{k} -> v.deepstack.{deepstack_visual_indexes[k]}). Checked against the mmproj's is_deepstack_layers KV and the text GGUF's {arch}.n_deepstack_layers; a tap past the tower depth or an incomplete merger refuses.
• With a tower the text runs interleaved mrope, sections from rope.dimension_sections (every published family config is [24, 20, 20] interleaved). Without one, the feat(gguf): serve qwen3vl and qwen3vlmoe gguf checkpoints text only #807 text-only path is unchanged.
• iter_mmproj_weights maps the mergers to visual.deepstack_merger_list.{k}.* and stacks the patch embedding's two temporal slices; the rest is HF layout. Tower runs bf16 via --mmproj or folder discovery.
• Both GGUF registry rows gain mm_processor / encoders.

Tower parity

model.visual.* range-fetched from the HF safetensors: unsloth mmproj-BF16 is bit-exact against Qwen3-VL-4B-Thinking (315/315) and Qwen3-VL-30B-A3B-Thinking (351/351), the 18 DeepStack merger tensors per model included. Taps [5, 11, 17] / [8, 16, 24], matching the KV and n_deepstack_layers = 3.

Check sweep

fetch_headers --search Qwen3-VL (17.7k shards) plus a header stub of every *mmproj*.gguf (100 files; the tool skips them, so fetch() was driven directly):

result	files
pass, vision exercised (41 repos)	477
pass, repo ships no mmproj	119
IQ1_M kernel gap	29
broken uploads (fail identically on main)	12
ik_llama custom types	10

645 trunks of qwen3vl / qwen3vlmoe, 596 pass. No subfolder-layout repos in the corpus, so every published mmproj is folder-discoverable.

Serving

RTX 5090 32 GB, driver 580.95.05, Ryzen 9 7950X3D, 187 GiB RAM.

• 4B UD-Q4_K_XL + auto-discovered mmproj-BF16: reads OCR text verbatim, /v1/stats reports ["text", "image"].
• 30B-A3B UD-Q3_K_XL (MoE offload): two images answered in order; a repeated image prompt hits the radix through the span (#cached-token 0 -> 222); 13.4k-token prefill ~11.1k tok/s, decode 212-263 tok/s.

Not covered

• video (the rows register the vision encoder only, as the HF checkpoints)
• Qwen3-ASR / Omni audio towers; their text files parse (n_deepstack_layers = 0 agrees with a tap-less tower)
• the 80 repos that ship no mmproj

No per-family GGUF test file, per the #721 review direction; tests/models/test_qwen3_vl.py's live-registry encoder test covers the rows.

[jason-fxz]

Cross-checked at 54bb59f on an RTX PRO 6000 Blackwell (96 GB) against a local implementation of the same change:

• Tower: unsloth BF16 / F32 mmproj vs Qwen/Qwen3-VL-8B-Instruct and Qwen/Qwen3-VL-30B-A3B-Instruct, 351/351 bit-exact each; F16 max 3e-8, Qwen's own Q8_0 max 2.4e-3. vision_config, image processor (12 images) and image token ids equal HF.
• Qwen3.5 / 3.6 / 3.8: derived configs, refusals, tower weights and auto-discovery byte-identical to main on 13 pairings.
• Header stubs of 120 Qwen3-VL repos: in all 259 qwen3vl_merger mmproj files is_deepstack_layers matches the v.deepstack.* tensors and every merger is complete; every qwen3vl / qwen3vlmoe text file declares n_deepstack_layers = 3. gguf_check over the 380 text x mmproj pairs gives the same result per file as our version.
• e2e, 10 setups x 17 probes (8B unsloth Q8_0 and Qwen Q4_K_M, 30B-A3B Q4_K_M; auto-discovered BF16, F16 / Q8_0 via --mmproj, --mm-disable vision, no mmproj): prompt tokens equal FreeToken serving the HF checkpoint 17/17, greedy answers identical to our implementation, the two text-only setups reject images with a 400.
• tests/models tests/mm tests/tokenizer tests/server: the same 7 failures as main (test_quant_config.py cases that read the local HF cache).

LGTM.