template error?
I have been testing the new qwen3.8 27b q4 - q8 (unsloth version) but with your template.
Run a tool-eval-bench for all the quants.
Noticed this is llama.cpp logs on some tests:
llama-server | 2026-08-14T22:38:04.340105932Z [41237] parse: error parsing grammar: expecting newline or end at ::= | " " | "\n"{1,2} [ \t]{0,20}
llama-server | 2026-08-14T22:38:04.340123068Z [41237]
llama-server | 2026-08-14T22:38:04.340124348Z [41237]
llama-server | 2026-08-14T22:38:04.340125452Z [41237] root ::=
llama-server | 2026-08-14T22:38:04.340126588Z [41237] space ::= | " " | "\n"{1,2} [ \t]{0,20}
Could it be template related?
No, this is not related to the chat template.
That error comes from llama-server itself when it tries to parse GBNF grammars. When you run tool-eval-bench, the benchmark passes JSON tool schemas in its requests. llama-server tries to convert those schemas into sampling grammar rules, and when a schema has certain optional fields, llama-server accidentally generates invalid grammar syntax (the leading pipe in space ::= | " ").
Chat templates only format prompt text and have no control over llama-server's internal grammar converter.
To resolve it, grab a recent pre-built release of llama.cpp, as newer builds have fixes for schema-to-grammar conversion. You can also check if tool-eval-bench has an option to disable grammar constraints and evaluate raw tool outputs instead.
Ahh... thank you for the enlightenment... and thank you for your chat template.