chat: fix llama-server image placeholder issue for PaddleOCR-VL

[megemini]

Fix chat template of PaddleOCR-VL for llama-server command, which requires media_marker instead of image (see #18825 (comment))

This should be able to handle these case:

• llama-server without --chat-template-file argument, and clean prompt without <__media__>
• llama-cli without --chat-template-file argument, and clean prompt without <__media__>

Tested:

• llama-server with command

./build/bin/llama-server -m /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-1.5-GGUF.gguf \
  --mmproj /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-1.5-GGUF-mmproj.gguf \
  --port 8111 --host 0.0.0.0 --ctx-size 131072 -n 4096 --temp 0 --jinja

• llama-cli with command

./build/bin/llama-cli -m /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-1.5-GGUF.gguf \
  --mmproj /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-1.5-GGUF-mmproj.gguf \
  --color on\
  --image /home/shun/Pictures/display_formula_7.png \
  --prompt "Formula Recognition:" --temp 0

@ngxson is this a good way to fix this issue?

[ngxson]

@ngxson is this a good way to fix this issue?

Are you being affiliated with Paddle team?

If yes, ask them to fix it themself.

[megemini]

@zhang-prog PaddleOCR-VL need update template to fix this issue.

[pwilkin]

I'd prefer not to add any code that patches templates on the fly, that's a very bad practice. Either fix the template at the source or, if it's not possible, just provide a corrected template.

[megemini]

I'd prefer not to add any code that patches templates on the fly, that's a very bad practice.

Agree ~~~

I think there may be three ways to fix this issue:

• update template: depends on PaddleOCR team
• update llama.cpp jinja parse code: a hard work
• hot fix (what this pr did) : temporary solution

whatever, I upload the fixed templates:

https://modelscope.cn/models/megemini/PaddleOCR-VL-1.5-GGUF/files
https://modelscope.cn/models/megemini/PaddleOCR-VL-GGUF/files