JinjaParser crashes server on unsupported chat_template statement keywords (e.g. {% macro %}) #273

Summary

Found while running OllamaBenchmarks against Bonsai-27B: JinjaParser (the chat-template Jinja2-subset interpreter, per docs/TOKENIZERS.md) only handles if/for/set statement keywords. Any other keyword — e.g. {% macro %}, which Qwen3-style chat templates use — throws an unhandled exception during parsing.

Impact

ServerStartup.LoadModel does not catch this exception, so the entire server process crashes before the model loads for any GGUF whose declared tokenizer.chat_template uses a macro (or any other statement type JinjaParser doesn't support) — not a graceful per-request failure, a full startup crash.

Repro

Serve prism-ml/Ternary-Bonsai-27B-gguf (or any GGUF with a {% macro %}-based chat template) via dotllm serve. Server crashes on load.

Workaround used (not a fix)

The benchmarking session served a byte-length-identical patched copy of the GGUF with a reduced if/for/set-only ChatML template, bypassing the actual template.

Suggested fix direction

Two independent improvements, likely both worth doing:

1. JinjaParser should support (or at minimum gracefully skip/no-op) {% macro %} blocks rather than throwing on an unrecognized keyword — real-world Qwen/Qwen3-family templates commonly use macros.
2. ServerStartup.LoadModel should catch template-parsing failures and either fail loudly with a clear, actionable error message (rather than an unhandled crash) or fall back to a safe default template, rather than taking down the whole process.

References

• docs/TOKENIZERS.md (chat template / Jinja2-subset interpreter docs)