ENH: Update parameters reference - #181
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request adds the mmproj_device parameter to the params_reference.md documentation. The review feedback suggests organizing this parameter under the 'Multimodal' section and updating its documented default value from nullptr to "" to avoid confusion for Python users, both of which require updates to the documentation generation script.
| | `load_mode` | llama_load_mode | `LLAMA_LOAD_MODE_AUTO` | R/W | how to load the model. | | ||
| | `mcp_servers_config` | str | `` | R/W | path to JSON file with MCP server definitions | | ||
| | `mcp_servers_json` | str | `` | R/W | inline JSON with MCP server definitions | | ||
| | `mmproj_device` | str | `nullptr` | R/W | GPU device for the multimodal projector (empty means auto, 'none' disables offload). | |
There was a problem hiding this comment.
For better organization and discoverability, the mmproj_device parameter should be documented under the 'Multimodal' section along with other related parameters like mmproj and mmproj_use_gpu.
To fix this, please add 'mmproj_device' to the Multimodal group list within COMMON_PARAMS_GROUPS in scripts/generate_params_doc.py and regenerate this documentation file.
| | `load_mode` | llama_load_mode | `LLAMA_LOAD_MODE_AUTO` | R/W | how to load the model. | | ||
| | `mcp_servers_config` | str | `` | R/W | path to JSON file with MCP server definitions | | ||
| | `mcp_servers_json` | str | `` | R/W | inline JSON with MCP server definitions | | ||
| | `mmproj_device` | str | `nullptr` | R/W | GPU device for the multimodal projector (empty means auto, 'none' disables offload). | |
There was a problem hiding this comment.
The default value for mmproj_device is documented as nullptr, which is the C++ default. However, the Python wrapper translates this to an empty string ("") for the 'auto' behavior. To avoid confusion for Python users, the documentation should reflect the effective default value, which is "".
This likely requires an adjustment in scripts/generate_params_doc.py to handle this case, or a special case for this parameter's default value.
1 participant
No description provided.