renderers 0.1.9.dev10__tar.gz → 0.1.10.dev4__tar.gz

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.9.dev10
+Version: 0.1.10.dev4
 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`, `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-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-0.1.9.dev10 → renderers-0.1.10.dev4}/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`, `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-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-0.1.9.dev10 → renderers-0.1.10.dev4}/docs/renderer-config.md

@@ -37,11 +37,13 @@ chat-template kwargs. Those fields are covered by parity tests against
 | Kimi K2 | `KimiK2RendererConfig` | - | `enable_thinking` |
 | Kimi K2.5 / 2.6 | `KimiK25RendererConfig` | `thinking` | `image_cache_max` |
 | 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` | - |
 | 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` | - | - |
 
@@ -133,10 +135,10 @@ 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 / XS-2.1, Llama 3 | `all` |
+| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / M.1 / XS-2.1, Llama 3 | `all` |
 | PrimeIntellect Qwen3 | `all` |
 
 Config construction raises when an explicit template knob directly contradicts

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/renderers/__init__.py

@@ -54,10 +54,12 @@ from renderers.configs import (
     Hy3RendererConfig,
     KimiK25RendererConfig,
     KimiK2RendererConfig,
+    LagunaM1RendererConfig,
     LagunaXS2RendererConfig,
     LagunaXS21RendererConfig,
     Llama3RendererConfig,
     MiniMaxM2RendererConfig,
+    Nemotron35RendererConfig,
     Nemotron3RendererConfig,
     Nemotron3UltraRendererConfig,
     PrimeQwen3RendererConfig,
@@ -90,10 +92,12 @@ _LAZY_RENDERERS: dict[str, str] = {
     "Hy3Renderer": "renderers.hy3",
     "KimiK25Renderer": "renderers.kimi_k25",
     "KimiK2Renderer": "renderers.kimi_k2",
+    "LagunaM1Renderer": "renderers.laguna_xs2",
     "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",
@@ -145,6 +149,8 @@ __all__ = [
     "KimiK25RendererConfig",
     "KimiK2Renderer",
     "KimiK2RendererConfig",
+    "LagunaM1Renderer",
+    "LagunaM1RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
     "LagunaXS21Renderer",
@@ -158,6 +164,8 @@ __all__ = [
     "MiniMaxM2RendererConfig",
     "MultiModalData",
     "MultimodalRenderer",
+    "Nemotron35Renderer",
+    "Nemotron35RendererConfig",
     "Nemotron3Renderer",
     "Nemotron3RendererConfig",
     "Nemotron3UltraRenderer",

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/renderers/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.1.9.dev10'
-__version_tuple__ = version_tuple = (0, 1, 9, 'dev10')
+__version__ = version = '0.1.10.dev4'
+__version_tuple__ = version_tuple = (0, 1, 10, 'dev4')
 
 __commit_id__ = commit_id = None

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/renderers/base.py

@@ -117,6 +117,7 @@ class Message(TypedDict, total=False):
     tool_calls: list[ToolCall]
     tool_call_id: str
     name: str
+    reasoning: str
     reasoning_content: str
 
 
@@ -1064,6 +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
@@ -1072,9 +1076,10 @@ 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. The two checkpoints ship different chat templates,
-    # each mirrored by its own renderer class.
+    # 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",
     # GPT-OSS.
     "openai/gpt-oss-20b": "gpt-oss",
@@ -1323,10 +1328,18 @@ def _populate_registry():
     from renderers.hy3 import Hy3Renderer
     from renderers.kimi_k2 import KimiK2Renderer
     from renderers.kimi_k25 import KimiK25Renderer
-    from renderers.laguna_xs2 import LagunaXS2Renderer, LagunaXS21Renderer
+    from renderers.laguna_xs2 import (
+        LagunaM1Renderer,
+        LagunaXS2Renderer,
+        LagunaXS21Renderer,
+    )
     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
@@ -1351,10 +1364,12 @@ def _populate_registry():
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,
+            "laguna-m.1": LagunaM1Renderer,
             "laguna-xs-2.1": LagunaXS21Renderer,
             "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "nemotron-3-ultra": Nemotron3UltraRenderer,
+            "nemotron-3.5": Nemotron35Renderer,
             "gpt-oss": GptOssRenderer,
         }
     )

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/renderers/configs.py

@@ -505,6 +505,28 @@ class LagunaXS2RendererConfig(BaseRendererConfig):
     chat template's ``render_assistant_messages_raw`` gate."""
 
 
+class LagunaM1RendererConfig(BaseRendererConfig):
+    """Laguna M.1 renderer config.
+
+    Laguna M.1 shares Laguna XS.2's role and tool-call format, but its
+    official checkpoint has a distinct chat template: it does not inject
+    XS.2's fallback system message and it gives ``message.reasoning``
+    precedence over ``message.reasoning_content``. Served by
+    :class:`renderers.laguna_xs2.LagunaM1Renderer`.
+    """
+
+    name: Literal["laguna-m.1"] = "laguna-m.1"
+
+    enable_thinking: bool = False
+    """When ``True``, the generation prompt includes ``<think>``. Mirrors
+    the official template's ``enable_thinking`` kwarg and default."""
+
+    render_assistant_messages_raw: bool = False
+    """When ``True``, assistant messages use the official template's
+    verbatim passthrough branch. See
+    :class:`LagunaXS2RendererConfig.render_assistant_messages_raw`."""
+
+
 class LagunaXS21RendererConfig(BaseRendererConfig):
     """Laguna XS-2.1 renderer config.
 
@@ -635,6 +657,37 @@ 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).
 
@@ -680,11 +733,13 @@ RendererConfig = Annotated[
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
+        LagunaM1RendererConfig,
         LagunaXS21RendererConfig,
         Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
         Nemotron3UltraRendererConfig,
+        Nemotron35RendererConfig,
         DeepSeekV3RendererConfig,
         DeepSeekR1RendererConfig,
     ],
@@ -720,11 +775,13 @@ _CONFIG_BY_NAME: dict[str, type[BaseRendererConfig]] = {
     "kimi-k2": KimiK2RendererConfig,
     "kimi-k2.5": KimiK25RendererConfig,
     "laguna-xs.2": LagunaXS2RendererConfig,
+    "laguna-m.1": LagunaM1RendererConfig,
     "laguna-xs-2.1": LagunaXS21RendererConfig,
     "llama-3": Llama3RendererConfig,
     "minimax-m2": MiniMaxM2RendererConfig,
     "nemotron-3": Nemotron3RendererConfig,
     "nemotron-3-ultra": Nemotron3UltraRendererConfig,
+    "nemotron-3.5": Nemotron35RendererConfig,
     "deepseek-v3": DeepSeekV3RendererConfig,
     "deepseek-r1": DeepSeekR1RendererConfig,
 }
@@ -767,10 +824,12 @@ __all__ = [
     "Hy3RendererConfig",
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
+    "LagunaM1RendererConfig",
     "LagunaXS2RendererConfig",
     "LagunaXS21RendererConfig",
     "Llama3RendererConfig",
     "MiniMaxM2RendererConfig",
+    "Nemotron35RendererConfig",
     "Nemotron3RendererConfig",
     "Nemotron3UltraRendererConfig",
     "PrimeQwen3RendererConfig",

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/renderers/laguna_xs2.py

@@ -1,4 +1,4 @@
-"""Laguna-XS.2 / XS-2.1 Renderer.
+"""Poolside Laguna renderer family.
 
 Main properties:
 - Prefix is the single token ``〈|EOS|〉`` (also the EOS / stop token).
@@ -36,6 +36,11 @@ XS-2.1's template (upstream rev ``575f0f28``) is served by the
   section ends at ``</available_tools>``.
 - The ``<system>`` block is emitted whenever there is system content,
   tools, or ``enable_thinking`` — even if that leaves it empty.
+
+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``.
 """
 
 from __future__ import annotations
@@ -56,7 +61,11 @@ from renderers.base import (
     resolve_thinking_retention,
     should_rerender_for_thinking_retention,
 )
-from renderers.configs import LagunaXS2RendererConfig, LagunaXS21RendererConfig
+from renderers.configs import (
+    LagunaM1RendererConfig,
+    LagunaXS2RendererConfig,
+    LagunaXS21RendererConfig,
+)
 from renderers.parsing import parse_laguna_xs2
 
 _DEFAULT_SYSTEM_MESSAGE = (
@@ -107,7 +116,12 @@ class LagunaXS2Renderer:
     def __init__(
         self,
         tokenizer: PreTrainedTokenizer,
-        config: LagunaXS2RendererConfig | LagunaXS21RendererConfig | None = None,
+        config: (
+            LagunaXS2RendererConfig
+            | LagunaM1RendererConfig
+            | LagunaXS21RendererConfig
+            | None
+        ) = None,
     ):
         self._tokenizer = tokenizer
         self.config = config or LagunaXS2RendererConfig()
@@ -481,12 +495,9 @@ class LagunaXS2Renderer:
         emit_text,
         emit_text_segments,
     ) -> None:
-        # Raw passthrough is an XS.2-only template gate; the XS-2.1
-        # config doesn't define it.
-        if (
-            isinstance(self.config, LagunaXS2RendererConfig)
-            and self.config.render_assistant_messages_raw
-        ):
+        # Raw passthrough is shared by XS.2 and M.1; XS-2.1's config does
+        # not expose this template gate.
+        if getattr(self.config, "render_assistant_messages_raw", False):
             self._render_assistant_raw(
                 msg_idx,
                 content,
@@ -495,16 +506,7 @@ class LagunaXS2Renderer:
             )
             return
 
-        reasoning_content = ""
-        if isinstance(msg.get("reasoning_content"), str):
-            reasoning_content = msg["reasoning_content"]
-        else:
-            # When the caller stores reasoning as a ``ThinkingPart`` inside
-            # a list-form ``content`` (e.g. after parse_response →
-            # reserialize), pull it out here so it survives the re-render.
-            part_thinking = self._thinking_text(msg.get("content"))
-            if part_thinking:
-                reasoning_content = part_thinking
+        reasoning_content, content = self._assistant_reasoning_and_content(msg, content)
 
         # ``<assistant>\n`` is template-injected scaffolding — the chat
         # template emits these as the generation prompt at inference and
@@ -567,6 +569,22 @@ class LagunaXS2Renderer:
         emit_special(self._assistant_end, msg_idx, is_sampled=True, is_content=True)
         emit_text("\n", msg_idx, is_sampled=False, is_content=False)
 
+    def _assistant_reasoning_and_content(
+        self, msg: Message, content: str
+    ) -> tuple[str, str]:
+        """Return the reasoning/body pair used by the XS.2 template."""
+        reasoning_content = ""
+        if isinstance(msg.get("reasoning_content"), str):
+            reasoning_content = msg["reasoning_content"]
+        else:
+            # When the caller stores reasoning as a ``ThinkingPart`` inside
+            # a list-form ``content`` (e.g. after parse_response →
+            # reserialize), pull it out here so it survives the re-render.
+            part_thinking = self._thinking_text(msg.get("content"))
+            if part_thinking:
+                reasoning_content = part_thinking
+        return reasoning_content, content
+
     def _render_assistant_raw(
         self,
         msg_idx: int,
@@ -620,6 +638,43 @@ class LagunaXS2Renderer:
         emit_text("\n", msg_idx, is_sampled=False, is_content=False)
 
 
+class LagunaM1Renderer(LagunaXS2Renderer):
+    """Renderer for the official ``poolside/Laguna-M.1`` checkpoint.
+
+    The token protocol is shared with XS.2, while the absent fallback
+    system prompt and assistant-reasoning precedence are M.1-specific.
+    """
+
+    def __init__(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        config: LagunaM1RendererConfig | None = None,
+    ):
+        super().__init__(tokenizer, config or LagunaM1RendererConfig())
+        self._default_system_message = ""
+
+    def _assistant_reasoning_and_content(
+        self, msg: Message, content: str
+    ) -> tuple[str, str]:
+        # Match the official Jinja exactly: ``reasoning`` wins whenever it
+        # is a string (including the empty string), then
+        # ``reasoning_content`` is considered. An inline </think> block is
+        # removed from content and becomes the fallback reasoning source.
+        reasoning_content = ""
+        if isinstance(msg.get("reasoning"), str):
+            reasoning_content = msg["reasoning"]
+        elif isinstance(msg.get("reasoning_content"), str):
+            reasoning_content = msg["reasoning_content"]
+
+        if "</think>" in content:
+            if not reasoning_content:
+                before_close = content.split("</think>", 1)[0].rstrip("\n")
+                reasoning_content = before_close.split("<think>")[-1].lstrip("\n")
+            content = content.rsplit("</think>", 1)[-1].lstrip("\n")
+
+        return reasoning_content, content
+
+
 class LagunaXS21Renderer(LagunaXS2Renderer):
     """Laguna-XS-2.1 — mirrors the ``poolside/Laguna-XS-2.1`` chat
     template (upstream rev ``575f0f28``); see the module docstring for

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/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-0.1.9.dev10 → renderers-0.1.10.dev4}/tests/conftest.py

@@ -36,7 +36,11 @@ 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"),

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/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.dev4/tests/test_laguna_m1.py

@@ -0,0 +1,232 @@
+"""Exact chat-template parity for the official Laguna M.1 checkpoint."""
+
+from __future__ import annotations
+
+from functools import lru_cache
+
+import pytest
+from pydantic import TypeAdapter
+
+from renderers import (
+    LagunaM1Renderer,
+    LagunaM1RendererConfig,
+    RendererConfig,
+    create_renderer,
+)
+from renderers.base import load_tokenizer
+
+_MODEL = "poolside/Laguna-M.1"
+
+TOOLS = [
+    {
+        "type": "function",
+        "function": {
+            "name": "shell",
+            "description": "Run a shell command",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "cmd": {"type": "string"},
+                    "timeout": {"type": "integer"},
+                },
+                "required": ["cmd"],
+            },
+        },
+    }
+]
+
+
+@lru_cache(maxsize=None)
+def _tok():
+    return load_tokenizer(_MODEL)
+
+
+def _renderer(**kwargs) -> LagunaM1Renderer:
+    renderer = create_renderer(_tok(), LagunaM1RendererConfig(**kwargs))
+    assert isinstance(renderer, LagunaM1Renderer)
+    return renderer
+
+
+def _expected(messages, *, tools=None, add_generation_prompt=False, **kwargs):
+    return list(
+        _tok().apply_chat_template(
+            messages,
+            tools=tools,
+            add_generation_prompt=add_generation_prompt,
+            tokenize=True,
+            return_dict=False,
+            **kwargs,
+        )
+    )
+
+
+def test_auto_selection_and_typed_config():
+    renderer = create_renderer(_tok())
+    assert isinstance(renderer, LagunaM1Renderer)
+    assert isinstance(renderer.config, LagunaM1RendererConfig)
+
+    parsed = TypeAdapter(RendererConfig).validate_python(
+        {"name": "laguna-m.1", "enable_thinking": True}
+    )
+    assert isinstance(parsed, LagunaM1RendererConfig)
+    assert parsed.enable_thinking is True
+
+
+@pytest.mark.parametrize("enable_thinking", [False, True])
+def test_no_system_and_generation_prompt_parity(enable_thinking):
+    messages = [{"role": "user", "content": "  preserve me  "}]
+    renderer = _renderer(enable_thinking=enable_thinking)
+    got = renderer.render_ids(messages, add_generation_prompt=True)
+    assert got == _expected(
+        messages,
+        add_generation_prompt=True,
+        enable_thinking=enable_thinking,
+    )
+
+    text = _tok().decode(got)
+    assert text.startswith("〈|EOS|〉<user>\n  preserve me  \n</user>\n<assistant>\n")
+    assert "<system>" not in text
+    assert text.endswith("<think>" if enable_thinking else "</think>")
+
+
+@pytest.mark.parametrize("enable_thinking", [False, True])
+def test_reasoning_content_history_parity_in_both_modes(enable_thinking):
+    messages = [
+        {"role": "user", "content": "Compute."},
+        {
+            "role": "assistant",
+            "reasoning_content": "\n  two steps  \n",
+            "content": "\n  result  \n",
+        },
+    ]
+    got = _renderer(enable_thinking=enable_thinking).render_ids(messages)
+    assert got == _expected(messages, enable_thinking=enable_thinking)
+    assert "<think>\ntwo steps\n</think>\nresult\n</assistant>\n" in _tok().decode(got)
+
+
+def test_reasoning_field_precedence_and_inline_think_extraction():
+    precedence = [
+        {"role": "user", "content": "Compute."},
+        {
+            "role": "assistant",
+            "reasoning": "preferred",
+            "reasoning_content": "ignored",
+            "content": "answer",
+        },
+    ]
+    got = _renderer(enable_thinking=True).render_ids(precedence)
+    assert got == _expected(precedence, enable_thinking=True)
+    text = _tok().decode(got)
+    assert "preferred" in text
+    assert "ignored" not in text
+
+    empty_reasoning_wins = [
+        {"role": "user", "content": "Compute."},
+        {
+            "role": "assistant",
+            "reasoning": "",
+            "reasoning_content": "also ignored",
+            "content": "answer",
+        },
+    ]
+    got = _renderer(enable_thinking=True).render_ids(empty_reasoning_wins)
+    assert got == _expected(empty_reasoning_wins, enable_thinking=True)
+    assert "also ignored" not in _tok().decode(got)
+
+    inline = [
+        {"role": "user", "content": "Compute."},
+        {
+            "role": "assistant",
+            "content": "<think>\ninline reason\n</think>\nvisible answer",
+        },
+    ]
+    assert _renderer().render_ids(inline) == _expected(inline)
+
+
+def test_tools_calls_responses_and_generation_prompt_parity():
+    messages = [
+        {"role": "system", "content": "Use tools carefully."},
+        {"role": "user", "content": "Inspect the machine."},
+        {
+            "role": "assistant",
+            "reasoning": "Need a command.",
+            "content": "Running it.",
+            "tool_calls": [
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {
+                        "name": "shell",
+                        "arguments": {"cmd": "uname -a", "timeout": 5},
+                    },
+                }
+            ],
+        },
+        {
+            "role": "tool",
+            "tool_call_id": "call_1",
+            "content": '{"stdout":"Linux"}',
+        },
+    ]
+    renderer = _renderer(enable_thinking=True)
+    got = renderer.render_ids(messages, tools=TOOLS, add_generation_prompt=True)
+    assert got == _expected(
+        messages,
+        tools=TOOLS,
+        add_generation_prompt=True,
+        enable_thinking=True,
+    )
+
+    text = _tok().decode(got)
+    assert "<available_tools>\n" in text
+    assert (
+        "<tool_call>shell\n<arg_key>cmd</arg_key>\n"
+        "<arg_value>uname -a</arg_value>\n<arg_key>timeout</arg_key>\n"
+        "<arg_value>5</arg_value>\n</tool_call>\n"
+    ) in text
+    assert '<tool_response>\n{"stdout":"Linux"}\n</tool_response>\n' in text
+    assert text.endswith("<assistant>\n<think>")
+
+
+def test_reasoning_content_and_tool_call_round_trip():
+    renderer = _renderer(enable_thinking=True)
+    prompt = [{"role": "user", "content": "Inspect the machine."}]
+    assistant = {
+        "role": "assistant",
+        "reasoning": "Need a command.",
+        "content": "Running it.",
+        "tool_calls": [
+            {
+                "function": {
+                    "name": "shell",
+                    "arguments": {"cmd": "uname -a", "timeout": 5},
+                }
+            }
+        ],
+    }
+    prompt_ids = renderer.render_ids(prompt, add_generation_prompt=True)
+    full_ids = renderer.render_ids([*prompt, assistant])
+    completion_ids = full_ids[len(prompt_ids) :]
+    parsed = renderer.parse_response(completion_ids, tools=TOOLS)
+
+    assert parsed.reasoning_content == "Need a command."
+    assert parsed.content == "Running it."
+    assert len(parsed.tool_calls) == 1
+    assert parsed.tool_calls[0].name == "shell"
+    assert parsed.tool_calls[0].arguments == {"cmd": "uname -a", "timeout": 5}
+
+
+def test_raw_assistant_round_trip_parity():
+    messages = [
+        {"role": "user", "content": "Continue."},
+        {
+            "role": "assistant",
+            "content": "<think>raw reason</think>raw body</assistant>",
+        },
+    ]
+    renderer = _renderer(enable_thinking=True, render_assistant_messages_raw=True)
+    assert renderer.render_ids(messages) == _expected(
+        messages,
+        enable_thinking=True,
+        render_assistant_messages_raw=True,
+    )

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/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-0.1.9.dev10 → renderers-0.1.10.dev4}/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-0.1.9.dev10 → renderers-0.1.10.dev4}/tests/test_renderer_config_parity.py

@@ -63,7 +63,11 @@ _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"),
     ("tencent/Hy3", "auto"),
     ("openai/gpt-oss-20b", "gpt-oss"),
@@ -119,7 +123,7 @@ _KWARG_VALUES: dict[str, list[Any]] = {
         "You are a helpful assistant. Your name is MiniMax-M2.5 and is built by MiniMax.",
         "You are CustomBot, a research assistant.",
     ],
-    # Laguna-XS.2 — switches assistant rendering to a verbatim
+    # Laguna-XS.2 / Laguna-M.1 — switches assistant rendering to a verbatim
     # passthrough mode. The renderer paths diverge significantly under
     # this flag, so both values are exercised.
     "render_assistant_messages_raw": [True, False],

{renderers-0.1.9.dev10 → renderers-0.1.10.dev4}/tests/test_roundtrip.py

@@ -46,7 +46,10 @@ _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
     # ``enable_thinking=False`` its template drops assistant reasoning at
     # render time, so the reasoning round-trip can't hold by design.