Vendored minja.hpp missing .replace() and other common Jinja string methods — chat templates from popular HF models fail to render

[samuel100]

Summary

The vendored copy of minja at shared/api/minja.hpp is missing handlers for several standard Jinja string methods that current upstream minja supports. The most impactful gap is .replace(), which causes OrtxApplyChatTemplate to throw Unknown method: replace on chat templates shipped with widely-used Hugging Face models.

Downstream report: microsoft/Foundry-Local#800 (smollm3-3b chat template fails on system_message.replace("/no_think", "").replace("/think", "").rstrip()).

Reproduction

Apply any chat template that uses .replace() via OrtxApplyChatTemplate. Minimal example from smollm3-3b's chat_template.jinja:

{%- set custom_instructions = system_message.replace("/no_think", "").replace("/think", "").rstrip() -%}

Result:

Microsoft.ML.OnnxRuntimeGenAI.OnnxRuntimeGenAIException:
Unknown method: replace at row 23, column 48:
{%- set custom_instructions = system_message.replace("/no_think", "").replace("/think", "").rstrip() -%}
                                               ^

The error is thrown by the catch-all throw std::runtime_error("Unknown method: " + method->get_name()) at the end of the string-method dispatch in shared/api/minja.hpp.

Root cause

Diffing shared/api/minja.hpp against current upstream google/minja:

String method	Upstream minja	Vendored in ort-extensions
.strip / .rstrip / .lstrip	✅	✅
.split	✅	✅
.replace	✅	❌ missing
.upper / .lower / .title	✅	❌ missing

The vendored header declares itself as "based on the minja project, with modifications to be more VS and mobile compiler friendly", but appears to have drifted behind upstream on string-method coverage. These methods are not optional — they are standard Jinja2 string methods that Hugging Face's transformers chat templates rely on heavily.

Impact

Any model whose chat_template.jinja uses .replace(), .upper(), .lower(), or .title() will fail to render via OrtxApplyChatTemplate. This affects ORT GenAI and every downstream consumer (Foundry Local, custom apps using ORT GenAI's Tokenizer::ApplyChatTemplate).

Confirmed-affected model: smollm3-3b-generic-gpu:1. Likely also affects other recent HF model templates that use the same idioms (Qwen, Llama-3.x derivatives, several SmolLM variants).

Suggested fix

Two options, in preference order:

1. Re-sync shared/api/minja.hpp with current upstream google/minja, preserving the VS / mobile compiler-friendly patches. This is the durable fix and would also pick up other upstream improvements since the last sync.
2. Targeted patch adding handlers for .replace, .upper, .lower, .title (and any other string methods that have landed upstream since the last sync) into the existing string-method dispatch in shared/api/minja.hpp.

Either way, please also consider adding a regression test under test/pp_api_test/ that renders the chat templates of a handful of popular HF models (e.g. SmolLM3, Qwen3, Llama-3.x, Phi-4) against OrtxApplyChatTemplate to catch future drift between vendored minja and what HF model authors actually ship.

Cross-references

• Downstream user report: Smollm3-3b chat completion fails: chat_template.jinja uses unsupported .replace() / .rstrip() methods foundry-local#800
• Related prior minja work in this repo: Fix Minja OOB & Parser Recursion Vulnerabilities #1070 (OOB / parser recursion), Add Qwen3 Chat Template Support #1001 (Qwen3 template), Refactor SPM-based BPE Tokenizer and Update Chat Template Implementation #965 (chat template impl refactor)

[samuel100]

Opened #1082 with a fix: adds .replace, .upper, .lower handlers to the string-method dispatch in shared/api/minja.hpp with semantics matching upstream google/minja and Python str, plus regression tests in test/pp_api_test/test_tokenizer_chat.cc. Verified locally that the exact failing line from smollm3-3b's chat_template.jinja (per microsoft/foundry-local#800) now renders correctly.

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.