hf-chat-template

Render a Hugging Face chat_template into a prompt string, byte-for-byte identical to Python's transformers.apply_chat_template. The template is the Jinja2 string stored in a model's tokenizer_config.json.

[dependencies]
hf-chat-template = "0.1"

Example

use hf_chat_template::{ChatTemplate, Message};

let tmpl = ChatTemplate::from_str(
    "{% for m in messages %}<|im_start|>{{ m.role }}\n{{ m.content }}<|im_end|>\n{% endfor %}\
     {% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}",
)?;

let prompt = tmpl.render_messages(&[Message::user("Hello!")], true)?;
assert_eq!(prompt, "<|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n");
# Ok::<(), hf_chat_template::Error>(())

Load a real model's config to inject its special tokens and resolve named templates:

use hf_chat_template::{ChatTemplate, Message, TokenizerConfig};

let json = std::fs::read_to_string("tokenizer_config.json")?;
let cfg: TokenizerConfig = serde_json::from_str(&json)?;
let tmpl = ChatTemplate::from_tokenizer_config(&cfg)?;

let prompt = tmpl.render_messages(&[Message::user("Hi")], true)?;
# Ok::<(), Box<dyn std::error::Error>>(())

For tools, documents, or model-specific kwargs, build a RenderInput. For an arbitrary context, call render_value with a minijinja::Value.

What it does

The Jinja engine is minijinja. This crate adds the transformers compatibility layer on top of it, plus a corpus that checks byte-identical output against real models on every commit.

It installs the globals that templates use: raise_exception, strftime_now, and a Python-compatible tojson that matches the separators and key order of json.dumps(..., ensure_ascii=False). Python string, list, and dict methods come from pycompat.

It handles the three chat_template shapes (single string, named list, dict), special-token injection, and the string-or-parts multimodal content.

It emits a prompt string. Turning that into token IDs stays the caller's job (tokenizers, tiktoken-rs).

Verified compatibility

These models render byte-identical to transformers in CI. See COMPATIBILITY.md and the corpus.

Loading from the Hub

The hub feature adds from_hub, which fetches a model's tokenizer_config.json and compiles its template in one call. It uses the synchronous hf-hub client with rustls, so there is no system OpenSSL dependency. Authentication follows hf-hub: the HF_TOKEN env var or the token from huggingface-cli login, which gated repos need.

hf-chat-template = { version = "0.1", features = ["hub"] }

use hf_chat_template::{ChatTemplate, Message};

let tmpl = ChatTemplate::from_hub("Qwen/Qwen2.5-0.5B-Instruct")?;
let prompt = tmpl.render_messages(&[Message::user("Hi")], true)?;
# Ok::<(), hf_chat_template::Error>(())

Tokenizing

The tokenizers feature adds render_and_encode, which renders the prompt and encodes it to token IDs in one step. It encodes with add_special_tokens = false, because the template already emits the model's special tokens. This is what transformers.apply_chat_template(..., tokenize=True) does, and it avoids a doubled BOS.

hf-chat-template = { version = "0.1", features = ["tokenizers"] }

use hf_chat_template::{ChatTemplate, Message, RenderInput};
use hf_chat_template::tokenizers::Tokenizer;

let tmpl = ChatTemplate::from_str("{{ messages[0].content }}")?;
let tok = Tokenizer::from_file("tokenizer.json").unwrap();
let input = RenderInput { messages: vec![Message::user("hi")], ..Default::default() };
let (prompt, ids) = tmpl.render_and_encode(&input, &tok)?;
# Ok::<(), hf_chat_template::Error>(())

Feature flags

pycompat is on by default. It adds Python methods on values (.strip(), .split(), | items) through minijinja-contrib. Disable it to drop that dependency when your templates do not use those methods.

hub is off by default. It adds from_hub and from_hub_revision, pulling in hf-hub and a TLS stack that the core string-rendering path does not need.

tokenizers is off by default. It adds render_and_encode and re-exports tokenizers, pulling in that crate and its onig regex backend.

Caveats

This crate does not add or strip a BOS token. If a template emits {{ bos_token }}, set add_special_tokens = false at encode time so the tokenizer does not add BOS a second time. It renders what the template says.

strftime_now defaults to UTC, while transformers uses local time. Inject a FixedClock or your own Clock to match a specific reference.

License

Dual-licensed under MIT or Apache-2.0, at your option.

Files under tests/corpus/ are trimmed excerpts of upstream model configs, redistributed under each model's own license. See tests/corpus/README.md.