Missing --jinja/--no-jinja flags in llama.cpp plugin after refactor

[AlexonOliveiraRH]

Description

After the refactor from YAML inference specs to Python plugins, the --jinja/--no-jinja flag logic was lost from the llama-server command builder. This causes:

• Text models (granite, phi4): missing --jinja + --chat-template-file, leading to template issues
• Multimodal models (smolvlm): receiving neither --no-jinja nor the mmproj-specific exclusion of --chat-template-file, causing tokenization errors

Steps to Reproduce

1. ramalama serve --port 8080 --name multi-modal -d smolvlm
2. Open the camera demo page
3. Submit an image for inference
4. Get error: Server error: 400 - Failed to tokenize prompt with number of bitmaps (1) does not match number of markers (0)

Root Cause

In ramalama/plugins/runtimes/inference/llama_cpp_commands.py, the _cmd_run() method lost the jinja logic that was previously defined in the YAML spec (inference-spec/engines/llama.cpp.yaml). The old YAML spec had:

• Text models: --jinja + --chat-template-file {path}
• Multimodal models (with mmproj): --no-jinja and NO --chat-template-file

Additionally, _get_mmproj_path() returns an empty string "" in dryrun mode (which is falsy in Python), so a double-check is needed: bool(mmproj_path) or bool(model._get_mmproj_path(False, False, False)).

Expected Behavior

The command builder should:

1. Detect if the model has a mmproj file (multimodal)
2. If multimodal: add --mmproj path + --no-jinja
3. If text-only: add --chat-template-file (if available) + --jinja

[olliewalsh]

After the refactor from YAML inference specs to Python plugins, the --jinja/--no-jinja flag logic was lost from the llama-server command builder.

No. The --no-jinja flag was a hack. I explicitly removed it in #2603 to fix qwen3.5.

This causes:
Text models (granite, phi4): missing --jinja + --chat-template-file, leading to template issues

No. --jinja is the default in llama.cpp.

Multimodal models (smolvlm): receiving neither --no-jinja nor the mmproj-specific exclusion of --chat-template-file, causing tokenization errors

It's an issue with the jinja chat templates used in the smolvlm quantized models which I raised upstream ggml-org/llama.cpp#21634

@AlexonOliveiraRH did AI tell you all of this? There is a template for bug report issues, please use it.

[AlexonOliveiraRH]

After the refactor from YAML inference specs to Python plugins, the --jinja/--no-jinja flag logic was lost from the llama-server command builder.

No. The --no-jinja flag was a hack. I explicitly removed it in #2603 to fix qwen3.5.

This causes:
Text models (granite, phi4): missing --jinja + --chat-template-file, leading to template issues

No. --jinja is the default in llama.cpp.

Multimodal models (smolvlm): receiving neither --no-jinja nor the mmproj-specific exclusion of --chat-template-file, causing tokenization errors

It's an issue with the jinja chat templates used in the smolvlm quantized models which I raised upstream ggml-org/llama.cpp#21634

@AlexonOliveiraRH did AI tell you all of this? There is a template for bug report issues, please use it.

I misunderstood the history here. Apologies for the noise, and thanks for the context @olliewalsh. I'll use the bug report template going forward. As we discussed in private, yes I used AI assistance just to open the Issue, PR, and text formatting, since these are very tedious and somewhat repetitive for me, not for the code itself, which is all on me. For the next ones, I'll use the Assisted-By flag anyways. Thanks.