Misc. bug: --override-kv for add_bos_token on Gemma4 is ignored

[personalmountains]

Name and Version

Running llama-server with --override-kv tokenizer.ggml.add_bos_token=bool:false still gives a double bos warning. Looks like #21500 hardcodes add_bos to true, regardless of the command line overrides.

Operating systems

Windows

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

llama-server

Command line

--override-kv tokenizer.ggml.add_bos_token=bool:false

Problem description & steps to reproduce

N/A

First Bad Commit

No response

Relevant log output

Logs

validate_override: Using metadata override ( bool) 'tokenizer.ggml.add_bos_token' = false
load: override 'tokenizer.ggml.add_bos_token' to 'true' for Gemma4
llm_tokenizer_bpe_session::check_double_bos_eos: Added a BOS token to the prompt as specified by the model but the prompt also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. Are you sure this is what you want?

[CISC]

Try 0 instead of false, but the correct value is true so you should not set it to false, it's the template that is incorrect, download the latest official one and use it with --chat-template-file.

[personalmountains]

@CISC Thanks for your help. add_bos_token is definitely a bool and I see my override being handled correctly up to the hardcoded value for gemma4.

you should not set it to false

The prompt I pass to llama-server already contains <bos> and so I do need to set it to false. It's the reason llama-server supports overriding these settings.