UPSTREAM PR #19495: jinja: add missing tojson filter for undefined type

[loci-dev]

Note

Source pull request: ggml-org/llama.cpp#19495

jinja: add missing tojson filter for undefined type

value_undefined_t::get_builtins() was the only type missing the
tojson filter, causing "unknown filter 'tojson' for type Undefined"
when templates reference undefined variables (e.g. tools | tojson
with no tools defined).

[loci-review]

Overview

Analysis of llama.cpp across 115,762 functions (33 modified, 0 new, 0 removed, 115,729 unchanged) reveals negligible performance impact. All changes occur in non-critical paths—template engine initialization, HTTP library utilities, and STL functions—with zero modifications to core inference operations (matrix multiplication, attention mechanisms, KV cache management).

Power Consumption Changes:

• build.bin.llama-tts: 361,809.82 nJ → 361,580.66 nJ (-0.063%)
• build.bin.llama-cvector-generator: 356,206.11 nJ → 356,127.02 nJ (-0.022%)
• build.bin.libllama.so, build.bin.libmtmd.so, build.bin.llama-bench, build.bin.llama-tokenize, build.bin.llama-quantize, build.bin.llama-qwen2vl-cli, build.bin.libggml-base.so, build.bin.libggml-cpu.so, build.bin.libggml.so, build.bin.llama-gguf-split, build.bin.llama-llava-cli, build.bin.llama-minicpmv-cli, build.bin.llama-gemma3-cli: 0.000% change

Function Analysis

Jinja Template Engine (value_undefined_t::get_builtins): Response time increased 790-868ns (+2.8-3.1%) due to adding tojson filter (commit 488b598). This one-time startup cost enables JSON serialization of undefined values, improving template robustness. Throughput time increased 7.5-8.2ns (+1.2-1.3%).

HTTP Library (httplib::Client::Post, std::function::operator=): Response time changes of -86ns to +8ns with security improvements (overflow prevention, Expect: 100-continue support). Throughput time increased 26ns (+28.7%) in Post wrapper, offset by 86ns improvement (-52.4%) in operator=.

STL Functions: std::__adjacent_find improved 48ns (-8.6% response, -26.9% throughput), std::__detail::_Executor::_M_handle_subexpr_begin improved 37ns (-13.3% response, -27.8% throughput). Other analyzed functions saw negligible changes from compiler optimizations.

All changes represent correctness improvements (template filter support, HTTP security) with acceptable performance trade-offs in non-critical initialization paths.

🔎 Full breakdown: Loci Inspector.
💬 Questions? Tag @loci-dev.