llama-server ignores --chat-template-file when --mmproj is provided.

[carlosjln]

Name and Version

llama-server version b9524

Operating systems

Windows

GGML backends

CUDA

Hardware

RTX 2080 ti

Models

gemma-4-12b-it-UD-Q4_K_XL.gguf

Problem description & steps to reproduce

llama-server ignores --chat-template-file when --mmproj is provided.

Expected:
--chat-template-file should be validated and override GGUF metadata template regardless of --mmproj.

Actual:
When --mmproj is present, llama-server uses the GGUF embedded tokenizer.chat_template and silently ignores --chat-template-file. Passing a nonexistent template file path produces no error. Removing --mmproj makes the same bad path error, and a valid custom template is loaded correctly.

Repro:

1. Run llama-server with -m model.gguf --jinja --chat-template-file nonexistent.jinja
   Result: error, as expected.
2. Run llama-server with -m model.gguf --mmproj mmproj.gguf --jinja --chat-template-file nonexistent.jinja
   Result: no error, server starts and uses GGUF template.

First Bad Commit

No response

Relevant log output

n/a

[CISC]

That's not reproducible (and should not be possible as the template file is read during argument parsing), I get the following error:

error while handling argument "--chat-template-file": error: failed to open file 'nonexistent.jinja'

[carlosjln]

This is the exact params I'm using and the template file is only loaded and displayed in the console and the webui model info if I remove -mm "%MMPROJ%"

when -mm "%MMPROJ%" is present, it doesn't even throw the error

"%EXEC_PATH%" ^
	--host %HOST% ^
	--port %PORT% ^
	-m "%MODEL_FILE%" ^
	-mm "%MMPROJ%"
	-a "%MODEL_ALIAS%" ^
	-c %CONTEXT_SIZE% ^
	-t %THREADS% ^
	-tb %THREADS_BATCH% ^
	--threads-http %THREADS_HTTP% ^
	-ngl %GPU_LAYERS% ^
	-fa 1 ^
	-b 1024 ^
	-ub 512 ^
	--cont-batching ^
	--cache-prompt ^
	--cache-reuse 256 ^
	--cache-ram 8192 ^
	--cache-idle-slots ^
	--ctx-checkpoints 1024 ^
	--kv-unified ^
	--cache-type-k q4_0 ^
	--cache-type-v q4_0 ^
	--context-shift ^
	--mlock ^
	--no-mmap ^
	--metrics ^
	--temp 0.9 ^
	--top-k 65 ^
	--top-p 0.95 ^
	--min-p 0.05 ^
	--repeat-penalty 1.0 ^
	--reasoning on ^
	--reasoning-budget 4096 ^
	--jinja ^
	--tools all ^
	--chat-template-file "%CHAT_TEMPLATE%"

When I remove it, it behaves as expected

[carlosjln]

wait a minute...

-m "%MODEL_FILE%" ^
-mm "%MMPROJ%"
-a "%MODEL_ALIAS%" ^

I have about 2h scratching my head and searching the internet... the issue is right there, the missing " ^" ignores the rest of the argument list.

Sorry to bother 🤦