
Template::Jinja2
Complete Jinja2 template engine for Raku, targeting compatibility with the Python reference implementation (Jinja2 3.1). Designed for HuggingFace chat templates and general-purpose templating.
Synopsis
use Template::Jinja2;
my $env = Template::Jinja2.new;
my $tmpl = $env.from-string('Hello {{ name }}!');
say $tmpl.render(name => 'World'); # Hello World!
HuggingFace Chat Templates
use Template::Jinja2;
use JSON::Fast;
my $config = from-json('tokenizer_config.json'.IO.slurp);
my $chat-template = $config<chat_template>;
my $env = Template::Jinja2.new;
my $result = $env.from-string($chat-template).render(
messages => [
{ role => 'system', content => 'You are helpful.' },
{ role => 'user', content => 'Hello!' },
],
bos_token => '<s>',
add_generation_prompt => True,
);
Produces byte-identical output to Python Jinja2 for ChatML, Llama 3, Mistral, Gemma 2, Zephyr, and Cohere Command A templates.
Features
All 15 Jinja2 tag types plus loop controls:
{% if %} / {% elif %} / {% else %} / {% endif %}
{% for %} / {% else %} / {% endfor %} with full loop context
{% set %} / {% endset %} with filter chains and namespace support
{% with %} / {% endwith %}
{% block %} / {% endblock %} with scoped blocks
{% extends %} / {% include %}
{% import %} / {% from ... import %}
{% macro %} / {% endmacro %} with defaults, varargs, kwargs
{% call %} / {% endcall %} with parameterized caller
{% filter %} / {% endfilter %}
{% do %} / {% raw %} / {# comment #}
{% autoescape %} / {% endautoescape %}
{% break %} / {% continue %}
Expressions
Arithmetic: +, -, *, /, //, %, **
Comparison: ==, !=, <, >, <=, >=>, chained
Logical: and, or, not
Membership: in, not in
Identity: is, is not
Ternary: expr if test else expr
String concat: ~
Filters: value | filter(args)
Subscript: items[0], items[-1], items[1:3], items[::-1]
Dot access: obj.attr, list.0
Literals: strings, integers, floats, hex/octal/binary, scientific notation, lists, dicts, tuples
Python Methods
Jinja2 has no method syntax of its own: value.method() resolves the name on the underlying Python object and calls it. Templates in the wild — HuggingFace chat templates especially — rely on that constantly, so the Python str and dict methods are implemented with Python semantics, not those of Raku's similarly-named methods.
{%- for m in messages -%}
{#- strip() as both a guard and as output -#}
{%- if m.content.strip() -%}
<|{{ m.role }}|>{{ m.content.strip() }}
{%- endif -%}
{#- reasoning split, straight out of the GLM chat template -#}
{%- if '</think>' in m.content -%}
{%- set reasoning = m.content.split('</think>')[0].split('<think>')[-1] -%}
{%- endif -%}
{#- get() with a default, straight out of the Kimi chat template -#}
{%- set name = m.get('name') or m['role'] -%}
{%- set rc = m.get('reasoning', m.get('reasoning_content', '')) -%}
{%- endfor -%}
{#- dict iteration for tool schemas -#}
{%- for k, v in tool.items() -%}"{{ k }}": {{ v | tojson }}{%- endfor -%}
String methods:
strip(chars?), lstrip(chars?), rstrip(chars?) — no argument strips whitespace; with an argument it strips any of the given characters, so "abcHIcba".strip("abc") is "HI", not a prefix trim
startswith(prefix), endswith(suffix) — the argument may be a tuple (or list) of candidates, any of which may match: name.startswith(('sys_', 'usr_'))
replace(old, new, count?) — replaces every occurrence unless count is given; an empty old inserts between every character, as in Python
split(sep?, maxsplit?) — with no separator, splits on runs of whitespace and drops empty fields ("a b c".split() is ['a', 'b', 'c'], however many blanks separate the words); with a separator it is a literal split that keeps them ("a||b".split("|") is ['a', '', 'b']). maxsplit caps the number of splits and leaves the remainder verbatim
lower(), upper() — Unicode-aware case folding
title() — Python's str.title(): the first letter after every non-letter is uppercased, so "it's a test" becomes "It'S A Test". Note this is deliberately not the title filter, which follows Jinja2's gentler word splitting and yields "It's A Test"
find(needle, start?, end?) — the first index, or -1 when absent (never an undefined value); negative start/end count from the end
Dict methods:
items() — (key, value) tuples, ready for {% for k, v in d.items() %}
keys(), values()
get(key), get(key, default) — a missing key yields None (falsy, renders as None) with no default, so m.get('name') or m['role'] works exactly as it does in Python
Method results chain, feed filters and tests, and bind through {% set %}:
{{ msg.content.strip().split('\n')[0] }}
{{ msg.content.strip() | tojson }}
{% set parts = ref.split('|') %}{{ parts[0] }}/{{ parts[-1] }}
Two deliberate divergences from CPython, both consequences of a Raku Hash having no insertion order to preserve:
items(), keys() and values() iterate in sorted-key order, which is what {% for k in d %}, |items and |dictsort already do in this engine. Output is deterministic, but a template that rebuilds JSON by iterating a dict emits its keys sorted rather than in the order they were written.
{{ d.keys() }} renders as a plain list (['a', 'b']) rather than Python's dict_keys(['a', 'b']) view repr. Iterating, indexing and filtering are unaffected.
Anything else raises a TemplateRuntimeError naming the method and the receiver's type — 'str' object has no method 'casefold' — rather than reaching a same-named Raku method or leaking an internal error. Calling a method on an undefined value raises UndefinedError identifying the expression that produced it.
Filters (52)
abs, attr, batch, capitalize, center, count, d/default, dictsort, e/escape, filesizeformat, first, float, forceescape, format, groupby, indent, int, items, join, last, length, list, lower, lstrip, map, max, min, pprint, random, reject, rejectattr, replace, reverse, round, rstrip, safe, select, selectattr, slice, sort, string, striptags, sum, title, tojson, trim, truncate, unique, upper, urlencode, urlize, wordcount, wordwrap, xmlattr
Tests (20+)
defined, undefined, none, boolean, integer, float, number, string, sequence, mapping, iterable, callable, sameas, eq/equalto, ne, lt/lessthan, le, gt/greaterthan, ge, even, odd, divisibleby, in, true, false, lower, upper
Template Inheritance
{# base.html #}
<html>{% block content %}default{% endblock %}</html>
{# child.html #}
{% extends "base.html" %}
{% block content %}{{ super() }} + override{% endblock %}
Supports multi-level inheritance with super() chaining.
Whitespace Control
{%- trim left -%} {# trim both sides #}
{%+ no_lstrip +%} {# disable lstrip/trim_blocks #}
my $env = Template::Jinja2.new(:trim-blocks, :lstrip-blocks);
Custom Delimiters
# ERB-style
my $env = Template::Jinja2.new(
:block-start('<%'), :block-end('%>'),
:variable-start('<%='), :variable-end('%>'),
:comment-start('<%#'), :comment-end('%>'));
# PHP-style
my $env = Template::Jinja2.new(
:block-start('<?'), :block-end('?>'),
:variable-start('<?='), :variable-end('?>'),
:comment-start('<!--'), :comment-end('-->'));
Loaders
use Template::Jinja2::Loader;
# File system
my $env = Template::Jinja2.new(
loader => FileSystemLoader.new(searchpath => 'templates/'));
# In-memory
my $env = Template::Jinja2.new(
loader => DictLoader.new(templates => {
'base.html' => '...',
'child.html' => '...',
}));
Configuration
my $env = Template::Jinja2.new(
:autoescape, # Auto-escape {{ }} output
:trim-blocks, # Strip newline after block tags
:lstrip-blocks, # Strip leading whitespace before block tags
:!keep-trailing-newline, # Strip trailing newline from output
:line-statement-prefix('#'), # Line-based block syntax
:line-comment-prefix('##'), # Line-based comments
globals => { site => 'My Site' },
filters => { custom => sub ($v) { ... } },
tests => { custom => sub ($v) { ... } },
);
Architecture
Template String -> [Lexer] -> Tokens -> [Parser] -> AST -> [Renderer] + Context -> Output
Lexer — Raku grammar tokenizer (default delimiters) or regex chunker (custom delimiters)
Parser — Pratt-style expression parser with block tag dispatch
AST — 30+ node types covering all Jinja2 constructs
Renderer — AST walker with expression evaluation, scope chain, filter/test/method dispatch
Methods — Python str/dict method semantics for value.method() calls
Context — Scope stack with Undefined sentinel and Namespace support
Testing
644 tests across 24 test files, including:
Ported tests from the Python Jinja2 reference test suite
Real-world HuggingFace chat template validation
Byte-identical output verification against Python Jinja2 3.1.6
Python str/dict method semantics, expectation-for-expectation against CPython
# Run all tests
prove6 -I lib t/
# Run a single test
raku -I lib t/09-huggingface.rakutest
Author
Matt Doughty
License
Artistic-2.0