convert: Zyphra Zamba2-VL (text model and mmproj)

[bong-water-water-bong]

Converts Zyphra's Zamba2-VL: a text GGUF and an mmproj. It runs through mtmd with no C++ changes.

Zamba2-VL is a Zamba2 language model under language_model., with tensor names identical to Zamba2ForCausalLM's, plus Qwen2.5-VL's vision tower and merger under vision_tower..

Text GGUF

• The Zamba2 converter on language_model.*, routed by the top-level architecture.
• The vocab is padded to vocab_size: tokenizer.json has 32,005 tokens and the embedding has 32,064 rows.
• The GGUF gets a chat template that also takes string content with media markers, which is what llama.cpp passes. It renders exactly the same prompt as Zyphra's template, with images before the user turn. Checked with jinja2 for both forms.

mmproj

• The Qwen2.5-VL vision converter on vision_tower.*.
• temporal_patch_size is 1, so the second patch conv (which clip's Qwen2-VL graph adds for a still image) is written as zeros. No runtime change.

Why no C++ changes: mtmd takes positions from the text model's rope type, so Zamba2 gets plain sequential positions. It wraps images in <|vision_start|>/<|vision_end|>, which Zamba2-VL's vocab already has.

Checked on Strix Halo with Zamba2-VL-2.7B (text Q8_0, mmproj F16), llama-server --jinja --mmproj, and a 448×336 test image (red square, blue circle, the text "HELLO 42"):

• Vision embeddings against transformers' Qwen2_5_VisionTransformerPretrainedModel in FP32, with Zamba2-VL's vision weights and Qwen2VLImageProcessor from its preprocessor config, 192 tokens × 2560:
  • CPU: mean per-token cosine 0.99989 (min 0.9979), relative L2 1.3%
  • Vulkan: mean 0.99944 (min 0.9903), relative L2 3.2%
• Answers: "A red square and a blue circle with the text "HElo 42" underneath." The shape and colour question is answered correctly. The text is misread the same way on CPU and Vulkan, so the misread is the model's.
• Decode about 40 tok/s on Vulkan0.

🤖 Generated with Claude Code