renderers

Programmable chat templates for LLM training and inference. A renderer turns a model's chat template into a Python object that can render messages → token ids, parse completion ids → structured assistant messages, and extend a multi-turn rollout without re-rendering model-sampled history.

Standalone on PyPI, and portable across training and inference stacks (transformers, vLLM, SGLang, Tinker). Initially developed for RL training with verifiers and prime-rl at Prime Intellect.

Install

uv add renderers

At a glance

from transformers import AutoTokenizer
from renderers import create_renderer

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
r = create_renderer(tok)                            # → Qwen3Renderer (auto-resolved)

prompt_ids = r.render_ids(
    [{"role": "user", "content": "hi"}],
    add_generation_prompt=True,
)
# Feed prompt_ids to a Token-In, Token-Out endpoint.
# It returns completion_ids sampled by the model.

parsed = r.parse_response(completion_ids)
# ParsedResponse(content=..., reasoning_content=..., tool_calls=...)

For the next turn, extend the previous sampled stream instead of re-rendering history:

next_prompt_ids = r.bridge_to_next_turn(
    previous_prompt_ids=prompt_ids,
    previous_completion_ids=completion_ids,
    new_messages=[{"role": "tool", "content": "..."}],
)

Hand-coded renderers ship for qwen3, qwen3-vl, qwen3.5, qwen3.6, glm-5, glm-5.1, glm-4.5, minimax-m2, deepseek-v3, kimi-k2, kimi-k2.5, nemotron-3, gpt-oss. Anything else falls back to DefaultRenderer, a generic apply_chat_template wrapper.

API

class Renderer(Protocol):
    def render(messages, *, tools=None, add_generation_prompt=False) -> RenderedTokens: ...
    def render_ids(messages, *, tools=None, add_generation_prompt=False) -> list[int]: ...
    def parse_response(token_ids) -> ParsedResponse: ...
    def get_stop_token_ids() -> list[int]: ...
    def bridge_to_next_turn(prev_prompt_ids, prev_completion_ids, new_messages, *, tools=None) -> list[int] | None: ...

• RenderedTokens carries token_ids and message_indices — one entry per token attributing each to its source message (-1 for structural scaffolding). Lets build_training_sample build a per-token loss mask in one render.
• ParsedResponse is (content, reasoning_content, tool_calls). It scans token ids for special-token boundaries (e.g. id 151657 for <tool_call> on Qwen3) — a literal "<tool_call>" in user content tokenizes to ordinary text ids and never matches.
• Round-trip: rendering [user, assistant(content, reasoning, tool_calls)], slicing the assistant completion, and feeding it through parse_response returns an equivalent structured message. Tested per-renderer in tests/test_roundtrip.py.

bridge_to_next_turn (the core contract)

Given (prev_prompt_ids, prev_completion_ids) and new environment messages, return ids for the next turn's prompt such that the result starts with prev_prompt_ids + prev_completion_ids byte-for-byte and continues with the new messages plus the next assistant opener. If that cannot be proven safe, return None and the caller falls back to a full render.

Each hand-coded bridge:

1. Anchors at the previous turn's canonical close token. On clean stops it's already in prev_completion_ids. On truncation, the renderer synthesizes the close as non-loss prompt context.
2. Refuses assistant content in new_messages — re-rendering sampled tokens would replace them with canonical template bytes.
3. Renders only the new messages in the framing the model family expects.

DefaultRenderer.bridge_to_next_turn returns None unconditionally — the template's close is unknown, so the contract can't be proven.

Picking a renderer

r = create_renderer(tok)                # AutoRendererConfig is the implicit default

Auto-detect matches tokenizer.name_or_path against MODEL_RENDERER_MAP by exact match. Prefix matching is intentionally off — same architecture can ship different chat templates (base vs instruct, fine-tune renames). Fine-tunes must pass an explicit typed config (e.g. Qwen3RendererConfig()); unknown names fall back to DefaultRenderer.

Pools

from renderers import create_renderer_pool

pool = create_renderer_pool("Qwen/Qwen3-8B", size=16)
with pool.checkout() as r:
    ids = r.render_ids(messages)

Each slot owns its own tokenizer copy. Construction fans out across a thread pool so a 32-slot pool doesn't serially eat ~10–15s of from_pretrained calls at startup.

Why use a renderer

For RL the trainer must see the exact token ids the sampler saw. The standard alternative — let the inference engine apply the chat template, parse tool calls, parse reasoning, and re-render full history every turn — silently breaks token identity. These are the failure modes a renderer's bridge_to_next_turn sidesteps by never re-rendering prior turns:

• Boolean round-trip. Engine emits false; client parses to Python bool(False); apply_chat_template re-renders via str(False) → "False". Capital F. Reproducible on Qwen3.5-35B-A3B + mini-swe-agent-plus at ~50% break rate per rollout.
• BPE retokenization drift. The same substring tokenizes differently depending on neighbouring bytes. json + p + enderer (3 tokens) vs jsonp + enderer (2 tokens) when whitespace shifts by one character. Every subsequent token is shifted from there on.
• Tool-call XML drift. The engine emits a no-arg call with a stylistic empty </parameter>; the Jinja re-render of the reconstructed dict drops it. Extension property broken at every such call.
• Thinking stripped from non-latest assistants. Some templates strip <think>…</think> blocks from prior assistant turns when re-rendering. The recorded stream has the thinking; the next prompt does not.
• Max-seq-len truncation zeroing the anchor. Client-side max_seq_len enforcement zeros completion_ids when prompt_len > max_seq_len. The bridge anchor is empty, falling back to full re-render — triggering every mode above.
• Scaffold-level history rewriting. Some agent scaffolds (e.g. opencode's experimental_repairToolCall) rewrite tool calls before sending them back as history. The next turn's prompt contains a tool call the model never emitted. A renderer cannot fix this — the drift happens before rendering.

Empirical delta on Qwen3.5-35B-A3B + mini-swe-agent-plus, step 0:

client path	breaks	training samples from 64 rollouts
apply_chat_template (full re-render)	32	77
renderers bridge_to_next_turn	0	64

Each break fragments a rollout into multiple training samples — every fragment re-encodes its prefix, inflating compute roughly linearly with the number of breaks.

Typed renderer configs

Each renderer accepts a typed pydantic config that pins its template-control kwargs at construction. create_renderer and create_renderer_pool take one positional config argument:

from renderers import (
    create_renderer,
    AutoRendererConfig,
    Qwen3RendererConfig,
    GLM5RendererConfig,
    DefaultRendererConfig,
)

# Auto-resolve renderer from the tokenizer's model name. Carries the
# shared preserve_* flags; template kwargs require an explicit choice.
renderer = create_renderer(tokenizer)
renderer = create_renderer(tokenizer, AutoRendererConfig(preserve_all_thinking=True))

# Explicit choice — the typed config exposes exactly the fields that
# renderer's chat template honours.
renderer = create_renderer(tokenizer, Qwen3RendererConfig(enable_thinking=False))
renderer = create_renderer(tokenizer, GLM5RendererConfig(clear_thinking=False))

# Default renderer (apply_chat_template fallback) — extra fields are
# captured via pydantic ``extra="allow"`` and forwarded to the Jinja
# template; tool / reasoning parsers are typed.
renderer = create_renderer(
    tokenizer,
    DefaultRendererConfig(tool_parser="qwen3", reasoning_parser="think"),
)

Discriminated union: every per-renderer config is a variant of RendererConfig, dispatched on the name field. Bogus combinations (e.g. add_vision_id under name="qwen3") error at construction with a pydantic.ValidationError. Downstream pydantic configs (prime-rl orchestrator, verifiers ClientConfig) hold a single field typed as RendererConfig and inherit the same strict-per-variant validation.

Two shared behaviour flags live on every variant via _BaseRendererConfig:

• preserve_all_thinking=True — every past assistant's reasoning_content is kept, even when the chat template would drop it.
• preserve_thinking_between_tool_calls=True — reasoning is kept on assistants in the in-flight tool cycle (post-last-user A-T-…-A block when it contains a tool response). A new user turn closes the block and drops its thinking.

These OR-compose with template-level toggles (e.g. GLM-5 clear_thinking, Nemotron-3 truncate_history_thinking): either flag saying "keep" wins. preserve_* can only ever extend retention — never override a template kwarg into a "drop" decision. The canonical use case is compaction: injecting a user turn like "summarize the work so far" puts every prior assistant in a past cycle, and preserve_all_thinking=True keeps reasoning visible end-to-end.

DefaultRenderer

Fallback for unsupported models. Wraps apply_chat_template and accepts tool_parser / reasoning_parser (vLLM convention) plus arbitrary Jinja kwargs via DefaultRendererConfig's extra="allow". bridge_to_next_turn returns None because the template's close is unknown, so multi-turn rollouts fall back to full re-render. Implementing a hand-coded renderer is a few hundred lines of Python (render_ids + parse_response + bridge_to_next_turn) and is the only path that closes the failure modes above by construction.

Roadmap

• VLM support. ContentPart is text-only today; Qwen3VLRenderer ships only because Qwen3-VL's text-only chat template differs from Qwen3's. Plan: add ImagePart / VideoPart, multimodal bridges, validate against a Qwen3-VL RL run.
• Patched chat templates. Some shipped templates re-tokenize history, normalize JSON, or auto-strip thinking — each breaks the extension property. Plan: a use_patched opt-in per renderer that renders the same surface form while avoiding known-bad patterns.

Testing

uv sync --group dev
uv run pytest

Round-trip parity (render → parse → original) and token-level parity against apply_chat_template are tested per renderer. End-to-end validation runs against Reverse-Text, Wordle, OpenCode-Math, and RLM-SWE environments.

License

Licensed under the Apache License, Version 2.0.