Misc. bug: llama-server assistant prefill only works when message content is a string (not a list of objects)

[find0x90]

Name and Version

.\llama-server --version
...
version: 5747 (0142961a)
built with clang version 18.1.8 for x86_64-pc-windows-msvc

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

llama-server

Command line

llama-server.exe --host 0.0.0.0 --port 1234 --flash-attn --no-warmup --model ~\llm\models\google\gemma-3-27b-it-qat-q4_0-gguf\gemma-3-27b-it-q4_0.gguf --mmproj ~\llm\models\google\gemma-3-27b-it-qat-q4_0-gguf\mmproj-model-f16-27B.gguf --gpu-layers 63 --temp 1.0 --repeat-penalty 1.0 --min-p 0.01 --top-k 64  --top-p 0.95 --cache-type-k q8_0 --cache-type-v q8_0 --ctx-size 16384

Problem description & steps to reproduce

See the below example where I send the same message in two different ways, first with "content" as a list of objects and then as a string. In the first case, the server does not continue the assistant message. In the second case, when the "content" field is a string, the server continues the assistant message as expected.

[user@88947fa9b7b9 ~]$ curl http://localhost:1234/v1/chat/completions -s --json '{
    "model": "Gemma3-27B",
    "messages": [
        {
            "role": "user",
            "content": "Hello!"
        },
        {
            "role": "assistant",
            "content": [
                {
                    "type": "text",
                    "text": "Ahoy-"
                }
            ]
        }
    ]
}' | jq '.choices[0].message.content'
"Hello to you too! 👋 \n\nIt's nice to meet you (virtually, of course!). How can I help you today? Do you have any questions, need some ideas, or just want to chat? \n\nLet me know what's on your mind! 😊\n"

[user@88947fa9b7b9 ~]$ curl http://localhost:1234/v1/chat/completions -s --json '{
    "model": "Gemma3-27B",
    "messages": [
        {
            "role": "user",
            "content": "Hello!"
        },
        {
            "role": "assistant",
            "content": "Ahoy-"
        }
    ]
}' | jq '.choices[0].message.content'
"hoy! 👋 \n\nHello to *you* too! How can I help you today? Are you looking to:\n\n* **Chat?** Just want someone to talk to?\n* **Brainstorm ideas?** \n* **Get information?** (I can try my best to answer your questions!)\n* **Write something?** (Stories, poems, code, etc.)\n* **Something else?**\n\nLet me know what's on your mind! 😊"

First Bad Commit

No response

Relevant log output