renderers 0.1.10.dev4__tar.gz → 0.1.10.dev10__tar.gz

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.10.dev4
+Version: 0.1.10.dev10
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE
@@ -56,7 +56,7 @@ next_prompt_ids = r.bridge_to_next_turn(
 )
 ```
 
-Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
+Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-s-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
 
 ## API
 

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/README.md

@@ -40,7 +40,7 @@ next_prompt_ids = r.bridge_to_next_turn(
 )
 ```
 
-Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
+Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-s-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
 
 ## API
 

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/docs/renderer-config.md

@@ -39,6 +39,7 @@ chat-template kwargs. Those fields are covered by parity tests against
 | Laguna XS.2 | `LagunaXS2RendererConfig` | `enable_thinking`, `render_assistant_messages_raw` | - |
 | Laguna M.1 | `LagunaM1RendererConfig` | `enable_thinking`, `render_assistant_messages_raw` | - |
 | Laguna XS-2.1 | `LagunaXS21RendererConfig` | `enable_thinking` | - |
+| Laguna S-2.1 | `LagunaS21RendererConfig` | `enable_thinking`, `preserve_thinking` | - |
 | Llama 3 | `Llama3RendererConfig` | `date_string`, `tools_in_user_message` | - |
 | MiniMax M2 | `MiniMaxM2RendererConfig` | `model_identity` | - |
 | Nemotron-3 Nano / Super | `Nemotron3RendererConfig` | `enable_thinking`, `truncate_history_thinking`, `low_effort` | - |
@@ -138,7 +139,7 @@ the knobs its template actually exposes:
 | Nemotron-3 / 3.5 | `truncate_history_thinking=False -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
 | DeepSeek R1 | `template` |
 | MiniMax M2 | `tool_cycle` |
-| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / M.1 / XS-2.1, Llama 3 | `all` |
+| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / M.1 / XS-2.1 / S-2.1, Llama 3 | `all` |
 | PrimeIntellect Qwen3 | `all` |
 
 Config construction raises when an explicit template knob directly contradicts

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/renderers/__init__.py

@@ -55,6 +55,7 @@ from renderers.configs import (
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaM1RendererConfig,
+    LagunaS21RendererConfig,
     LagunaXS2RendererConfig,
     LagunaXS21RendererConfig,
     Llama3RendererConfig,
@@ -93,6 +94,7 @@ _LAZY_RENDERERS: dict[str, str] = {
     "KimiK25Renderer": "renderers.kimi_k25",
     "KimiK2Renderer": "renderers.kimi_k2",
     "LagunaM1Renderer": "renderers.laguna_xs2",
+    "LagunaS21Renderer": "renderers.laguna_s21",
     "LagunaXS21Renderer": "renderers.laguna_xs2",
     "LagunaXS2Renderer": "renderers.laguna_xs2",
     "Llama3Renderer": "renderers.llama_3",
@@ -151,6 +153,8 @@ __all__ = [
     "KimiK2RendererConfig",
     "LagunaM1Renderer",
     "LagunaM1RendererConfig",
+    "LagunaS21Renderer",
+    "LagunaS21RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
     "LagunaXS21Renderer",

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/renderers/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.1.10.dev4'
-__version_tuple__ = version_tuple = (0, 1, 10, 'dev4')
+__version__ = version = '0.1.10.dev10'
+__version_tuple__ = version_tuple = (0, 1, 10, 'dev10')
 
 __commit_id__ = commit_id = None

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/renderers/base.py

@@ -1076,11 +1076,12 @@ MODEL_RENDERER_MAP: dict[str, str] = {
     # construction to pin a different date.
     "meta-llama/Llama-3.2-1B-Instruct": "llama-3",
     "meta-llama/Llama-3.2-3B-Instruct": "llama-3",
-    # Poolside Laguna. These checkpoints ship distinct chat templates,
-    # each mirrored by its own renderer class/config discriminator.
+    # Poolside Laguna. These checkpoints ship distinct chat templates, each
+    # mirrored by its own renderer class/config discriminator.
     "poolside/Laguna-XS.2": "laguna-xs.2",
     "poolside/Laguna-M.1": "laguna-m.1",
     "poolside/Laguna-XS-2.1": "laguna-xs-2.1",
+    "poolside/Laguna-S-2.1": "laguna-s-2.1",
     # GPT-OSS.
     "openai/gpt-oss-20b": "gpt-oss",
     "openai/gpt-oss-120b": "gpt-oss",
@@ -1328,6 +1329,7 @@ def _populate_registry():
     from renderers.hy3 import Hy3Renderer
     from renderers.kimi_k2 import KimiK2Renderer
     from renderers.kimi_k25 import KimiK25Renderer
+    from renderers.laguna_s21 import LagunaS21Renderer
     from renderers.laguna_xs2 import (
         LagunaM1Renderer,
         LagunaXS2Renderer,
@@ -1366,6 +1368,7 @@ def _populate_registry():
             "laguna-xs.2": LagunaXS2Renderer,
             "laguna-m.1": LagunaM1Renderer,
             "laguna-xs-2.1": LagunaXS21Renderer,
+            "laguna-s-2.1": LagunaS21Renderer,
             "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "nemotron-3-ultra": Nemotron3UltraRenderer,

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/renderers/configs.py

@@ -546,6 +546,37 @@ class LagunaXS21RendererConfig(BaseRendererConfig):
     upstream default."""
 
 
+class LagunaS21RendererConfig(BaseRendererConfig):
+    """Laguna S-2.1 renderer config.
+
+    S-2.1 is a larger sibling of XS-2.1 sharing its tokenizer (same vocab,
+    special tokens, and merges), but its chat template is *not* byte-identical:
+    ``enable_thinking`` defaults to ``True`` (XS-2.1 defaults ``False``), and a
+    new ``preserve_thinking`` kwarg widens the reasoning-display gate to
+    ``enable_thinking or preserve_thinking``. The token format is otherwise
+    identical, so this is served by
+    :class:`renderers.laguna_s21.LagunaS21Renderer`, a thin subclass of
+    ``LagunaXS21Renderer`` that only overrides that gate.
+    """
+
+    name: Literal["laguna-s-2.1"] = "laguna-s-2.1"
+
+    enable_thinking: bool = True
+    """When ``True``, the generation prompt ends with ``<think>`` and every
+    assistant turn renders ``<think>{reasoning}</think>``; when ``False``,
+    turns open with a bare ``</think>``. Mirrors the template's
+    ``enable_thinking`` kwarg — note S-2.1's upstream default is ``True``,
+    unlike XS-2.1's ``False``."""
+
+    preserve_thinking: bool = False
+    """When ``True``, assistant turns keep their ``<think>{reasoning}</think>``
+    block even while ``enable_thinking`` is ``False`` — the template gates
+    reasoning display on ``enable_thinking or preserve_thinking``. With the
+    default ``False`` the gate collapses to ``enable_thinking`` and the
+    renderer matches XS-2.1 turn-for-turn. Mirrors the template's
+    ``preserve_thinking`` kwarg and its upstream default."""
+
+
 class Llama3RendererConfig(BaseRendererConfig):
     """Llama-3.x Instruct renderer config.
 
@@ -735,6 +766,7 @@ RendererConfig = Annotated[
         LagunaXS2RendererConfig,
         LagunaM1RendererConfig,
         LagunaXS21RendererConfig,
+        LagunaS21RendererConfig,
         Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
@@ -777,6 +809,7 @@ _CONFIG_BY_NAME: dict[str, type[BaseRendererConfig]] = {
     "laguna-xs.2": LagunaXS2RendererConfig,
     "laguna-m.1": LagunaM1RendererConfig,
     "laguna-xs-2.1": LagunaXS21RendererConfig,
+    "laguna-s-2.1": LagunaS21RendererConfig,
     "llama-3": Llama3RendererConfig,
     "minimax-m2": MiniMaxM2RendererConfig,
     "nemotron-3": Nemotron3RendererConfig,
@@ -825,6 +858,7 @@ __all__ = [
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaM1RendererConfig",
+    "LagunaS21RendererConfig",
     "LagunaXS2RendererConfig",
     "LagunaXS21RendererConfig",
     "Llama3RendererConfig",

renderers-0.1.10.dev10/renderers/laguna_s21.py

@@ -0,0 +1,33 @@
+"""Laguna S-2.1 Renderer — a larger sibling of Laguna XS-2.1.
+
+S-2.1 shares XS-2.1's tokenizer and token format, so this is a thin subclass of
+:class:`renderers.laguna_xs2.LagunaXS21Renderer`; see that module for the shared
+format. The template delta is two thinking kwargs: ``enable_thinking`` defaults
+to ``True`` (XS-2.1 defaults ``False``), and ``preserve_thinking`` widens the
+reasoning-display gate to ``enable_thinking or preserve_thinking``.
+"""
+
+from __future__ import annotations
+
+from transformers.tokenization_utils import PreTrainedTokenizer
+
+from renderers.configs import LagunaS21RendererConfig
+from renderers.laguna_xs2 import LagunaXS21Renderer
+
+
+class LagunaS21Renderer(LagunaXS21Renderer):
+    """Mirrors the ``poolside/Laguna-S-2.1`` chat template."""
+
+    # Narrows the inherited attribute so the S-2.1-only ``preserve_thinking``
+    # resolves; ``__init__`` always stores an S-2.1 config.
+    config: LagunaS21RendererConfig
+
+    def __init__(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        config: LagunaS21RendererConfig | None = None,
+    ):
+        super().__init__(tokenizer, config or LagunaS21RendererConfig())
+
+    def _render_history_reasoning(self) -> bool:
+        return self.config.enable_thinking or self.config.preserve_thinking

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/renderers/laguna_xs2.py

@@ -41,6 +41,9 @@ Laguna M.1's official template (revision ``2bf8a4ab``) is served by
 :class:`LagunaM1Renderer`. It shares XS.2's byte layout, tool syntax, and
 generation prompt, but has no fallback system prompt and reads assistant
 reasoning from ``reasoning`` before falling back to ``reasoning_content``.
+
+S-2.1 subclasses :class:`LagunaXS21Renderer` from :mod:`renderers.laguna_s21`,
+overriding only the ``_render_history_reasoning`` seam.
 """
 
 from __future__ import annotations
@@ -63,6 +66,7 @@ from renderers.base import (
 )
 from renderers.configs import (
     LagunaM1RendererConfig,
+    LagunaS21RendererConfig,
     LagunaXS2RendererConfig,
     LagunaXS21RendererConfig,
 )
@@ -120,6 +124,7 @@ class LagunaXS2Renderer:
             LagunaXS2RendererConfig
             | LagunaM1RendererConfig
             | LagunaXS21RendererConfig
+            | LagunaS21RendererConfig
             | None
         ) = None,
     ):
@@ -688,10 +693,17 @@ class LagunaXS21Renderer(LagunaXS2Renderer):
     def __init__(
         self,
         tokenizer: PreTrainedTokenizer,
-        config: LagunaXS21RendererConfig | None = None,
+        config: LagunaXS21RendererConfig | LagunaS21RendererConfig | None = None,
     ):
         super().__init__(tokenizer, config or LagunaXS21RendererConfig())
 
+    def _render_history_reasoning(self) -> bool:
+        """Whether an assistant turn renders its ``<think>{reasoning}</think>``
+        block (vs opening with a bare ``</think>``). Mirrors the template's
+        reasoning-display gate; XS-2.1 gates this on ``enable_thinking`` alone.
+        """
+        return self.config.enable_thinking
+
     def render(
         self,
         messages: list[Message],
@@ -975,7 +987,7 @@ class LagunaXS21Renderer(LagunaXS2Renderer):
         # scaffolding the model never samples.
         emit_special(self._assistant, msg_idx, is_sampled=False, is_content=False)
 
-        if self.config.enable_thinking:
+        if self._render_history_reasoning():
             # ``<think>{reasoning}</think>`` renders verbatim, even when
             # the reasoning is empty; the opener is the gen-prompt
             # prefill, the rest the model sampled.

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/tests/conftest.py

@@ -44,6 +44,10 @@ RENDERER_MODELS = [
     # XS-2.1 resolves to `LagunaXS21Renderer` via the model name — its
     # chat template differs from XS.2's (see renderers/laguna_xs2.py).
     ("poolside/Laguna-XS-2.1", "auto"),
+    # S-2.1 is a larger sibling of XS-2.1 sharing its tokenizer but NOT its
+    # chat template (S-2.1 defaults enable_thinking=True and adds a
+    # preserve_thinking gate), so it resolves to LagunaS21Renderer.
+    ("poolside/Laguna-S-2.1", "auto"),
     # DeepSeek-V3/R1 are intentionally NOT in this shared barrage: their
     # chat templates can't render the barrage's tool-call fixtures (the
     # templates require ``tool['type']`` and a string-serialized

renderers-0.1.10.dev10/tests/test_laguna_s21.py

@@ -0,0 +1,128 @@
+"""Laguna-S-2.1 focused tests.
+
+S-2.1 shares XS-2.1's tokenizer and token format, so the shared matrices
+(conftest / config-parity) already assert byte parity on the common shapes
+via :class:`LagunaS21Renderer`. This file pins the two behaviours that make
+S-2.1 *differ* from XS-2.1 — and that the shared barrage can't reach because
+it never sets the relevant kwargs:
+
+- ``enable_thinking`` defaults to ``True`` (XS-2.1 defaults ``False``), so the
+  auto-resolved renderer renders reasoning by default, and
+- the new ``preserve_thinking`` kwarg keeps historical ``<think>`` blocks even
+  while ``enable_thinking`` is ``False`` — the template gates reasoning display
+  on ``enable_thinking or preserve_thinking``.
+"""
+
+from __future__ import annotations
+
+from functools import lru_cache
+from itertools import product
+
+from renderers import create_renderer
+from renderers.base import load_tokenizer
+from renderers.configs import LagunaS21RendererConfig
+from renderers.laguna_s21 import LagunaS21Renderer
+
+_MODEL = "poolside/Laguna-S-2.1"
+
+
+@lru_cache(maxsize=None)
+def _tok():
+    return load_tokenizer(_MODEL)
+
+
+def _renderer(**config_kwargs) -> LagunaS21Renderer:
+    renderer = create_renderer(_tok(), LagunaS21RendererConfig(**config_kwargs))
+    assert isinstance(renderer, LagunaS21Renderer)
+    return renderer
+
+
+def _expected(msgs, *, add_generation_prompt=False, **template_kwargs):
+    return list(
+        _tok().apply_chat_template(
+            msgs,
+            add_generation_prompt=add_generation_prompt,
+            tokenize=True,
+            return_dict=False,
+            **template_kwargs,
+        )
+    )
+
+
+def test_auto_resolves_to_s21_renderer_with_thinking_on_by_default():
+    """``poolside/Laguna-S-2.1`` resolves to :class:`LagunaS21Renderer`, and
+    its config defaults ``enable_thinking=True`` / ``preserve_thinking=False``
+    — matching S-2.1's template defaults (unlike XS-2.1's thinking-off)."""
+    r = create_renderer(_tok())  # auto-resolve via MODEL_RENDERER_MAP
+    assert isinstance(r, LagunaS21Renderer)
+    assert r.config.name == "laguna-s-2.1"
+    assert r.config.enable_thinking is True
+    assert r.config.preserve_thinking is False
+
+
+def test_default_renders_empty_think_wrapper():
+    """With thinking on by default, a reasoning-free assistant turn still
+    opens with the empty ``<think></think>`` wrapper (this is exactly the
+    render_ids divergence from the XS-2.1 renderer)."""
+    msgs = [
+        {"role": "user", "content": "What is 2+2?"},
+        {"role": "assistant", "content": "4"},
+    ]
+    r = _renderer()  # defaults: enable_thinking=True
+    ours = r.render_ids(msgs)
+    assert ours == _expected(msgs)
+    assert "<assistant><think></think>4</assistant>" in _tok().decode(ours)
+
+
+def test_preserve_thinking_keeps_history_when_thinking_off():
+    """``enable_thinking=False`` alone drops reasoning, but with
+    ``preserve_thinking=True`` the historical ``<think>{reasoning}</think>``
+    survives — matching the template's ``enable_thinking or preserve_thinking``
+    gate."""
+    msgs = [
+        {"role": "user", "content": "What is 2+2?"},
+        {"role": "assistant", "reasoning_content": "Simple arithmetic", "content": "4"},
+    ]
+    r = _renderer(enable_thinking=False, preserve_thinking=True)
+    ours = r.render_ids(msgs)
+    assert ours == _expected(msgs, enable_thinking=False, preserve_thinking=True)
+    assert "<think>Simple arithmetic</think>" in _tok().decode(ours)
+
+
+def test_no_preserve_thinking_drops_history_when_thinking_off():
+    """With both flags off the gate collapses to ``enable_thinking`` and the
+    turn opens with a bare ``</think>``, dropping the reasoning — identical to
+    the XS-2.1 renderer's default."""
+    msgs = [
+        {"role": "user", "content": "What is 2+2?"},
+        {"role": "assistant", "reasoning_content": "Simple arithmetic", "content": "4"},
+    ]
+    r = _renderer(enable_thinking=False, preserve_thinking=False)
+    ours = r.render_ids(msgs)
+    assert ours == _expected(msgs, enable_thinking=False, preserve_thinking=False)
+    assert "Simple arithmetic" not in _tok().decode(ours)
+
+
+def test_all_thinking_flag_combos_match_template():
+    """Byte parity against ``apply_chat_template`` for every
+    ``enable_thinking`` × ``preserve_thinking`` combination, over a multi-turn
+    conversation carrying historical reasoning."""
+    msgs = [
+        {"role": "user", "content": "Q1"},
+        {"role": "assistant", "reasoning_content": "deduce", "content": "A1"},
+        {"role": "user", "content": "Q2"},
+        {"role": "assistant", "reasoning_content": "more", "content": "A2"},
+    ]
+    for enable_thinking, preserve_thinking in product((False, True), repeat=2):
+        r = _renderer(
+            enable_thinking=enable_thinking, preserve_thinking=preserve_thinking
+        )
+        for add_generation_prompt in (False, True):
+            assert r.render_ids(
+                msgs, add_generation_prompt=add_generation_prompt
+            ) == _expected(
+                msgs,
+                add_generation_prompt=add_generation_prompt,
+                enable_thinking=enable_thinking,
+                preserve_thinking=preserve_thinking,
+            ), (enable_thinking, preserve_thinking, add_generation_prompt)

{renderers-0.1.10.dev4 → renderers-0.1.10.dev10}/tests/test_renderer_config_parity.py

@@ -69,6 +69,7 @@ _RENDERER_MODELS = [
     ("poolside/Laguna-XS.2", "auto"),
     ("poolside/Laguna-M.1", "auto"),
     ("poolside/Laguna-XS-2.1", "auto"),
+    ("poolside/Laguna-S-2.1", "auto"),
     ("tencent/Hy3", "auto"),
     ("openai/gpt-oss-20b", "gpt-oss"),
 ]