Problem

MTMDChatHandler._convert_message_for_template keeps multimodal content as a list of dicts after media marker substitution. For models with passthrough chat templates (e.g., {% for m in messages %}{{m['content']}}{% endfor %}), Jinja renders this list as its Python repr string. The model receives garbled input and produces 0 output tokens (immediate EOS).

Affects any model whose tokenizer.chat_template directly references m['content'] or m.content without iterating over content parts. Confirmed on Baidu Unlimited-OCR (deepseek2-ocr architecture, 4.6K GitHub stars).

Fix

After converting content parts (replacing image_url with the media marker text), check if all parts resolved to text-only dicts. If so, join them into a single concatenated string. Templates that iterate over content parts (e.g., LLaVA-style {% for item in message.content %}) are unaffected because the else-branch preserves the list.

Before / After

Before (0 output tokens):

$ python3 -c "result = llm.create_chat_completion(...); print(result['choices'][0]['message']['content'])"
# (empty string)

After (correct structured OCR with bounding boxes):

header [20, 43, 243, 104]INVOICE #1234
header [23, 116, 297, 174]Date: June 26, 2026
text [23, 468, 377, 528]Widget A 5 $10.00
text [24, 538, 390, 598]Widget B 10 $25.00
text [24, 608, 387, 670]Gadget C 2 $50.00
text [20, 748, 387, 810]Subtotal: $350.00
text [23, 818, 411, 882]Tax (8.875%): $31.06
text [24, 888, 367, 948]Total: $381.06

Test environment

• llama-cpp-python 0.3.31
• macOS ARM64 (M4 Pro), Metal
• Model: Unlimited-OCR Q4_K_M + mmproj-F16 (sahilchachra/Unlimited-OCR-GGUF)
• Also verified with llama-mtmd-cli --jinja (native C++ binary produces identical output, confirming the fix matches upstream behavior)