renderers 0.1.10.dev1__tar.gz → 0.1.10.dev10__tar.gz

{renderers-0.1.10.dev1 → renderers-0.1.10.dev10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.10.dev1
+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`, `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.dev1 → 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`, `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.dev1 → renderers-0.1.10.dev10}/docs/renderer-config.md

@@ -39,10 +39,12 @@ 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` | - |
 | Nemotron-3 Ultra | `Nemotron3UltraRendererConfig` | `enable_thinking`, `truncate_history_thinking`, `medium_effort` | - |
+| Nemotron-3.5 Lightning | `Nemotron35RendererConfig` | `enable_thinking`, `truncate_history_thinking` | - |
 | DeepSeek V3 | `DeepSeekV3RendererConfig` | - | - |
 | DeepSeek R1 | `DeepSeekR1RendererConfig` | - | - |
 
@@ -134,10 +136,10 @@ the knobs its template actually exposes:
 | gpt-oss | `auto_drop_analysis=False -> all`, else `tool_cycle` |
 | Hy3 | `preserved_thinking=True -> all`, else `tool_cycle` |
 | Kimi K2.5 / 2.6 | `thinking=False -> all`, else `tool_cycle` |
-| Nemotron-3 | `truncate_history_thinking=False -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
+| 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.dev1 → renderers-0.1.10.dev10}/renderers/__init__.py

@@ -55,10 +55,12 @@ from renderers.configs import (
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaM1RendererConfig,
+    LagunaS21RendererConfig,
     LagunaXS2RendererConfig,
     LagunaXS21RendererConfig,
     Llama3RendererConfig,
     MiniMaxM2RendererConfig,
+    Nemotron35RendererConfig,
     Nemotron3RendererConfig,
     Nemotron3UltraRendererConfig,
     PrimeQwen3RendererConfig,
@@ -92,10 +94,12 @@ _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",
     "MiniMaxM2Renderer": "renderers.minimax_m2",
+    "Nemotron35Renderer": "renderers.nemotron3",
     "Nemotron3Renderer": "renderers.nemotron3",
     "Nemotron3UltraRenderer": "renderers.nemotron3",
     "PrimeQwen3Renderer": "renderers.prime_qwen3",
@@ -149,6 +153,8 @@ __all__ = [
     "KimiK2RendererConfig",
     "LagunaM1Renderer",
     "LagunaM1RendererConfig",
+    "LagunaS21Renderer",
+    "LagunaS21RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
     "LagunaXS21Renderer",
@@ -162,6 +168,8 @@ __all__ = [
     "MiniMaxM2RendererConfig",
     "MultiModalData",
     "MultimodalRenderer",
+    "Nemotron35Renderer",
+    "Nemotron35RendererConfig",
     "Nemotron3Renderer",
     "Nemotron3RendererConfig",
     "Nemotron3UltraRenderer",

{renderers-0.1.10.dev1 → 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.dev1'
-__version_tuple__ = version_tuple = (0, 1, 10, 'dev1')
+__version__ = version = '0.1.10.dev10'
+__version_tuple__ = version_tuple = (0, 1, 10, 'dev10')
 
 __commit_id__ = commit_id = None

{renderers-0.1.10.dev1 → renderers-0.1.10.dev10}/renderers/base.py

@@ -1065,6 +1065,9 @@ MODEL_RENDERER_MAP: dict[str, str] = {
     "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16": "nemotron-3",
     "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16": "nemotron-3-ultra",
     "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-FP8": "nemotron-3-ultra",
+    # Nemotron 3.5 (Lightning). Its template is the Ultra variant's minus the
+    # effort kwarg (``nemotron-3.5``).
+    "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16": "nemotron-3.5",
     # Llama 3.2 (Instruct). Tested against the gated meta-llama repos and
     # the unrestricted unsloth/... mirror, which ships a byte-identical
     # chat template. ``Llama3Renderer`` defaults ``date_string`` to
@@ -1073,11 +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",
@@ -1325,6 +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,
@@ -1332,7 +1337,11 @@ def _populate_registry():
     )
     from renderers.llama_3 import Llama3Renderer
     from renderers.minimax_m2 import MiniMaxM2Renderer
-    from renderers.nemotron3 import Nemotron3Renderer, Nemotron3UltraRenderer
+    from renderers.nemotron3 import (
+        Nemotron3Renderer,
+        Nemotron3UltraRenderer,
+        Nemotron35Renderer,
+    )
     from renderers.prime_qwen3 import PrimeQwen3Renderer
     from renderers.qwen3 import Qwen3Renderer
     from renderers.qwen3_vl import Qwen3VLRenderer
@@ -1359,9 +1368,11 @@ 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,
+            "nemotron-3.5": Nemotron35Renderer,
             "gpt-oss": GptOssRenderer,
         }
     )

{renderers-0.1.10.dev1 → 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.
 
@@ -657,6 +688,37 @@ class Nemotron3UltraRendererConfig(BaseRendererConfig):
     user message. Mirrors the Ultra chat template's ``medium_effort`` kwarg."""
 
 
+class Nemotron35RendererConfig(BaseRendererConfig):
+    """Nemotron-3.5 (Lightning) renderer config.
+
+    Nemotron 3.5's chat template is the Ultra variant's minus the effort
+    kwarg: same reasoning-block glue (no ``\\n`` around ``</think>``, no
+    trim on truncated tool-call history), but its Jinja defines no
+    reasoning-effort variable at all. It shares the
+    :class:`renderers.nemotron3.Nemotron3Renderer` implementation via the
+    :class:`renderers.nemotron3.Nemotron35Renderer` subclass and is reached
+    via the ``nemotron-3.5`` discriminator.
+    """
+
+    name: Literal["nemotron-3.5"] = "nemotron-3.5"
+
+    enable_thinking: bool = True
+    """See :class:`Nemotron3RendererConfig.enable_thinking`."""
+
+    truncate_history_thinking: bool = True
+    """See :class:`Nemotron3RendererConfig.truncate_history_thinking`."""
+
+    @model_validator(mode="after")
+    def _check_thinking_retention(self):
+        _reject_thinking_retention_conflict(
+            self,
+            "truncate_history_thinking",
+            true_implies="tool_cycle",
+            false_implies="all",
+        )
+        return self
+
+
 class DeepSeekV3RendererConfig(BaseRendererConfig):
     """DeepSeek-V3 renderer config (non-reasoning).
 
@@ -704,10 +766,12 @@ RendererConfig = Annotated[
         LagunaXS2RendererConfig,
         LagunaM1RendererConfig,
         LagunaXS21RendererConfig,
+        LagunaS21RendererConfig,
         Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
         Nemotron3UltraRendererConfig,
+        Nemotron35RendererConfig,
         DeepSeekV3RendererConfig,
         DeepSeekR1RendererConfig,
     ],
@@ -745,10 +809,12 @@ _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,
     "nemotron-3-ultra": Nemotron3UltraRendererConfig,
+    "nemotron-3.5": Nemotron35RendererConfig,
     "deepseek-v3": DeepSeekV3RendererConfig,
     "deepseek-r1": DeepSeekR1RendererConfig,
 }
@@ -792,10 +858,12 @@ __all__ = [
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaM1RendererConfig",
+    "LagunaS21RendererConfig",
     "LagunaXS2RendererConfig",
     "LagunaXS21RendererConfig",
     "Llama3RendererConfig",
     "MiniMaxM2RendererConfig",
+    "Nemotron35RendererConfig",
     "Nemotron3RendererConfig",
     "Nemotron3UltraRendererConfig",
     "PrimeQwen3RendererConfig",

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.dev1 → 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.dev1 → renderers-0.1.10.dev10}/renderers/nemotron3.py

@@ -31,7 +31,11 @@ from renderers.base import (
     should_rerender_for_thinking_retention,
     trim_to_turn_close,
 )
-from renderers.configs import Nemotron3RendererConfig, Nemotron3UltraRendererConfig
+from renderers.configs import (
+    Nemotron3RendererConfig,
+    Nemotron3UltraRendererConfig,
+    Nemotron35RendererConfig,
+)
 from renderers.parsing import parse_qwen35
 
 # ---------------------------------------------------------------------------
@@ -76,10 +80,12 @@ def _render_extra_keys(obj: dict[str, Any], handled_keys: set[str]) -> list[str]
     return lines
 
 
-# The Nemotron-3 family ships two chat-template variants. Nano / Super share
+# The Nemotron family ships three chat-template variants. Nano / Super share
 # one (renderer ``Nemotron3Renderer`` / config ``name="nemotron-3"``); Ultra
 # differs in the reasoning-block glue — no ``\n`` around ``</think>`` — and is
-# the ``Nemotron3UltraRenderer`` subclass (``name="nemotron-3-ultra"``). Which
+# the ``Nemotron3UltraRenderer`` subclass (``name="nemotron-3-ultra"``);
+# Nemotron 3.5 (Lightning) uses the Ultra glue but defines no effort kwarg and
+# is the ``Nemotron35Renderer`` subclass (``name="nemotron-3.5"``). Which
 # variant a checkpoint uses is carried by ``MODEL_RENDERER_MAP``, so the right
 # renderer class is constructed and the variant is encoded by the class itself.
 
@@ -852,3 +858,16 @@ class Nemotron3UltraRenderer(Nemotron3Renderer):
 
     _config_cls = Nemotron3UltraRendererConfig
     _ultra = True
+
+
+class Nemotron35Renderer(Nemotron3Renderer):
+    """Renderer for Nemotron **3.5** (Lightning).
+
+    The 3.5 chat template is the Ultra variant's minus the effort kwarg: same
+    reasoning-block glue (carried by the ``_ultra`` class hook), but its Jinja
+    defines no reasoning-effort variable, so the config exposes none and the
+    effort hint is always empty.
+    """
+
+    _config_cls = Nemotron35RendererConfig
+    _ultra = True

{renderers-0.1.10.dev1 → renderers-0.1.10.dev10}/tests/conftest.py

@@ -36,11 +36,18 @@ RENDERER_MODELS = [
     # Ultra resolves to the `nemotron-3-ultra` config variant via the model
     # name (auto → MODEL_RENDERER_MAP → nemotron-3-ultra).
     ("nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "auto"),
+    # Nemotron 3.5 (Lightning): Ultra's template variant minus the effort
+    # kwarg (auto → MODEL_RENDERER_MAP → nemotron-3.5).
+    ("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "auto"),
     ("poolside/Laguna-XS.2", "auto"),
     ("poolside/Laguna-M.1", "auto"),
     # 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.dev1 → renderers-0.1.10.dev10}/tests/test_bridge.py

@@ -33,6 +33,7 @@ _BRIDGE_MODELS = [
     ("moonshotai/Kimi-K2-Instruct", "auto"),
     ("moonshotai/Kimi-K2.5", "auto"),
     ("nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "auto"),
+    ("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "auto"),
     ("tencent/Hy3", "auto"),
     ("unsloth/Llama-3.2-1B-Instruct", "llama-3"),
     ("openai/gpt-oss-20b", "gpt-oss"),

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.dev1 → renderers-0.1.10.dev10}/tests/test_nemotron3_parity.py

@@ -25,9 +25,10 @@ Every assertion compares ``renderer.render_ids(...)`` to
 renderer is byte-for-byte faithful for that case. Tokenizers are loaded from
 the local HF cache (offline); no network.
 
-The variants split across two configs: ``nemotron-3`` (Nano / Super, with
-``low_effort``) and ``nemotron-3-ultra`` (Ultra, with ``medium_effort``). The
-helper resolves the right config class per model from ``MODEL_RENDERER_MAP``.
+The variants split across three configs: ``nemotron-3`` (Nano / Super, with
+``low_effort``), ``nemotron-3-ultra`` (Ultra, with ``medium_effort``), and
+``nemotron-3.5`` (Lightning — Ultra's glue, no effort kwarg). The helper
+resolves the right config class per model from ``MODEL_RENDERER_MAP``.
 """
 
 from __future__ import annotations
@@ -44,7 +45,8 @@ from renderers.configs import _config_class_for
 NANO = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
 SUPER = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
 ULTRA = "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16"
-MODELS = [NANO, SUPER, ULTRA]
+LIGHTNING = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"
+MODELS = [NANO, SUPER, ULTRA, LIGHTNING]
 
 
 @lru_cache
@@ -107,7 +109,9 @@ def _assert_parity(
     )
 
 
-pytestmark = pytest.mark.parametrize("model", MODELS, ids=["nano", "super", "ultra"])
+pytestmark = pytest.mark.parametrize(
+    "model", MODELS, ids=["nano", "super", "ultra", "lightning"]
+)
 
 
 TOOLS = [
@@ -648,8 +652,8 @@ _EFFORT_SHAPES = [
 def test_low_effort_kwarg(model, flag, shape, extra):
     """``low_effort`` appends ``\\n\\n{reasoning effort: low}`` to the last user
     message on **Super**; it's a no-op on **Nano** (its template never defines
-    it). Ultra's config has no such field, so it's skipped."""
-    if model == ULTRA:
+    it). The Ultra / Lightning configs have no such field, so they're skipped."""
+    if model in (ULTRA, LIGHTNING):
         pytest.skip("low_effort is a nemotron-3 (Nano/Super) kwarg")
     _assert_parity(model, shape, low_effort=flag, **extra)
 
@@ -660,7 +664,7 @@ def test_low_effort_kwarg(model, flag, shape, extra):
 )
 def test_medium_effort_kwarg(model, flag, shape, extra):
     """``medium_effort`` appends ``\\n\\n{reasoning effort: efficient}`` on
-    **Ultra**. Nano/Super configs have no such field, so they're skipped."""
+    **Ultra**. The other configs have no such field, so they're skipped."""
     if model != ULTRA:
         pytest.skip("medium_effort is a nemotron-3-ultra kwarg")
     _assert_parity(model, shape, medium_effort=flag, **extra)
@@ -668,9 +672,12 @@ def test_medium_effort_kwarg(model, flag, shape, extra):
 
 def test_effort_kwarg_lives_on_the_right_variant(model):
     """Each effort kwarg is declared only on the variant whose template defines
-    it — the discriminated union rejects the wrong combination at config load."""
+    it — the discriminated union rejects the wrong combination at config load.
+    Lightning's template defines no effort variable, so its config has neither."""
     fields = _config_cls(model).template_field_names()
     if model == ULTRA:
         assert "medium_effort" in fields and "low_effort" not in fields
+    elif model == LIGHTNING:
+        assert "low_effort" not in fields and "medium_effort" not in fields
     else:
         assert "low_effort" in fields and "medium_effort" not in fields

{renderers-0.1.10.dev1 → renderers-0.1.10.dev10}/tests/test_nemotron3_ultra.py

@@ -1,4 +1,4 @@
-"""Offline wiring tests for the Nemotron-3 variant split.
+"""Offline wiring tests for the Nemotron variant split.
 
 Assert the model→renderer mapping, the per-variant typed-config surface, and
 the name-based ``low_effort`` gating WITHOUT loading any tokenizer (no
@@ -6,15 +6,16 @@ network). This pins the wiring the parity matrix can't reach — in particular
 the FP8 Ultra entry, which no test loads a tokenizer for — so it can't
 silently rot.
 
-The two variants:
+The three variants:
 
 * ``nemotron-3`` — Nano / Super, shared template. Config exposes ``low_effort``
   (honoured on Super, a no-op on Nano).
 * ``nemotron-3-ultra`` — Ultra, distinct ``</think>`` glue. Config exposes
   ``medium_effort``.
+* ``nemotron-3.5`` — Lightning, Ultra's glue but no effort kwarg at all.
 
-Both route to the one ``Nemotron3Renderer`` class, which selects the variant
-from ``config.name``.
+All route to the one ``Nemotron3Renderer`` implementation via per-variant
+subclasses.
 """
 
 from types import SimpleNamespace
@@ -23,9 +24,15 @@ from renderers.base import MODEL_RENDERER_MAP, RENDERER_REGISTRY, _populate_regi
 from renderers.configs import (
     Nemotron3RendererConfig,
     Nemotron3UltraRendererConfig,
+    Nemotron35RendererConfig,
     _config_class_for,
 )
-from renderers.nemotron3 import Nemotron3Renderer, Nemotron3UltraRenderer, _is_super
+from renderers.nemotron3 import (
+    Nemotron3Renderer,
+    Nemotron3UltraRenderer,
+    Nemotron35Renderer,
+    _is_super,
+)
 
 _ULTRA_REPOS = [
     "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16",
@@ -35,6 +42,9 @@ _NANO_SUPER_REPOS = [
     "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
     "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
 ]
+_LIGHTNING_REPOS = [
+    "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+]
 
 
 def _fake_tok(name):
@@ -46,18 +56,23 @@ def test_models_map_to_their_variant():
         assert MODEL_RENDERER_MAP.get(repo) == "nemotron-3-ultra", repo
     for repo in _NANO_SUPER_REPOS:
         assert MODEL_RENDERER_MAP.get(repo) == "nemotron-3", repo
+    for repo in _LIGHTNING_REPOS:
+        assert MODEL_RENDERER_MAP.get(repo) == "nemotron-3.5", repo
 
 
 def test_each_discriminator_maps_to_its_config_and_renderer_class():
     # Config discriminator → config class.
     assert _config_class_for("nemotron-3") is Nemotron3RendererConfig
     assert _config_class_for("nemotron-3-ultra") is Nemotron3UltraRendererConfig
-    # Registry → renderer class (Ultra is a sibling subclass, matching the
-    # GLM-5/5.1 and Qwen3.5/3.6 house style — not one class under two names).
+    assert _config_class_for("nemotron-3.5") is Nemotron35RendererConfig
+    # Registry → renderer class (Ultra / 3.5 are sibling subclasses, matching
+    # the GLM-5/5.1 and Qwen3.5/3.6 house style — not one class under N names).
     _populate_registry()
     assert RENDERER_REGISTRY["nemotron-3"] is Nemotron3Renderer
     assert RENDERER_REGISTRY["nemotron-3-ultra"] is Nemotron3UltraRenderer
+    assert RENDERER_REGISTRY["nemotron-3.5"] is Nemotron35Renderer
     assert issubclass(Nemotron3UltraRenderer, Nemotron3Renderer)
+    assert issubclass(Nemotron35Renderer, Nemotron3Renderer)
 
 
 def test_variant_is_encoded_by_the_class():
@@ -66,8 +81,10 @@ def test_variant_is_encoded_by_the_class():
     # config.name → class). Default config also follows the class.
     assert Nemotron3Renderer._ultra is False
     assert Nemotron3UltraRenderer._ultra is True
+    assert Nemotron35Renderer._ultra is True
     assert Nemotron3Renderer._config_cls is Nemotron3RendererConfig
     assert Nemotron3UltraRenderer._config_cls is Nemotron3UltraRendererConfig
+    assert Nemotron35Renderer._config_cls is Nemotron35RendererConfig
 
 
 def test_template_fields_per_variant():
@@ -80,6 +97,10 @@ def test_template_fields_per_variant():
     assert Nemotron3UltraRendererConfig.template_field_names() == frozenset(
         {"enable_thinking", "truncate_history_thinking", "medium_effort"}
     )
+    # Lightning's template defines no effort variable at all.
+    assert Nemotron35RendererConfig.template_field_names() == frozenset(
+        {"enable_thinking", "truncate_history_thinking"}
+    )
 
 
 def test_configs_reject_the_other_variants_effort_kwarg():
@@ -91,6 +112,10 @@ def test_configs_reject_the_other_variants_effort_kwarg():
         Nemotron3RendererConfig(medium_effort=True)  # type: ignore[call-arg]
     with pytest.raises(ValidationError):
         Nemotron3UltraRendererConfig(low_effort=True)  # type: ignore[call-arg]
+    with pytest.raises(ValidationError):
+        Nemotron35RendererConfig(low_effort=True)  # type: ignore[call-arg]
+    with pytest.raises(ValidationError):
+        Nemotron35RendererConfig(medium_effort=True)  # type: ignore[call-arg]
     # And the removed ``ultra`` selector is gone entirely.
     with pytest.raises(ValidationError):
         Nemotron3RendererConfig(ultra=True)  # type: ignore[call-arg]

{renderers-0.1.10.dev1 → renderers-0.1.10.dev10}/tests/test_renderer_config_parity.py

@@ -63,9 +63,13 @@ _RENDERER_MODELS = [
     # Ultra: auto-resolves to the ``nemotron-3-ultra`` config via the model
     # name; parity asserted against the Ultra apply_chat_template (``medium_effort``).
     ("nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "auto"),
+    # Nemotron 3.5 (Lightning): Ultra's template minus the effort kwarg — its
+    # config declares only ``enable_thinking`` / ``truncate_history_thinking``.
+    ("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "auto"),
     ("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"),
 ]

{renderers-0.1.10.dev1 → renderers-0.1.10.dev10}/tests/test_roundtrip.py

@@ -46,6 +46,8 @@ _ROUNDTRIP_MODELS = [
     # Ultra: parse must recover content after a </think> glued directly to it
     # (no separating newline) — the Ultra-specific glue stresses the round-trip.
     ("nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "auto"),
+    # Nemotron 3.5 (Lightning): same glue as Ultra, distinct renderer class.
+    ("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", "auto"),
     ("poolside/Laguna-XS.2", "auto"),
     ("poolside/Laguna-M.1", "auto"),
     # Laguna-XS-2.1 is deliberately absent: under the default