renderers 0.1.8.dev53__tar.gz → 0.1.8.dev54__tar.gz

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.8.dev53
+Version: 0.1.8.dev54
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/renderers/__init__.py

@@ -54,6 +54,7 @@ from renderers.configs import (
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaXS2RendererConfig,
+    LagunaXS21RendererConfig,
     Llama3RendererConfig,
     MiniMaxM2RendererConfig,
     Nemotron3RendererConfig,
@@ -139,6 +140,7 @@ __all__ = [
     "KimiK2RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
+    "LagunaXS21RendererConfig",
     "Llama3Renderer",
     "Llama3RendererConfig",
     "MULTIMODAL_MODELS",

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/renderers/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.1.8.dev53'
-__version_tuple__ = version_tuple = (0, 1, 8, 'dev53')
+__version__ = version = '0.1.8.dev54'
+__version_tuple__ = version_tuple = (0, 1, 8, 'dev54')
 
 __commit_id__ = commit_id = None

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/renderers/base.py

@@ -1060,8 +1060,10 @@ 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.
+    # Poolside Laguna. XS-2.1's template is byte-identical to XS.2's minus
+    # the default system message; the config name selects the variant.
     "poolside/Laguna-XS.2": "laguna-xs.2",
+    "poolside/Laguna-XS-2.1": "laguna-xs-2.1",
     # GPT-OSS.
     "openai/gpt-oss-20b": "gpt-oss",
     "openai/gpt-oss-120b": "gpt-oss",
@@ -1327,6 +1329,7 @@ def _populate_registry():
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,
+            "laguna-xs-2.1": LagunaXS2Renderer,
             "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "nemotron-3-ultra": Nemotron3UltraRenderer,

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/renderers/configs.py

@@ -422,6 +422,27 @@ class LagunaXS2RendererConfig(BaseRendererConfig):
     chat template's ``render_assistant_messages_raw`` gate."""
 
 
+class LagunaXS21RendererConfig(BaseRendererConfig):
+    """Laguna XS-2.1 renderer config — distinct discriminator so auto
+    resolution gives XS-2.1 checkpoints the no-default-system-message
+    template variant.
+
+    XS-2.1's chat template is byte-identical to XS.2's except it ships
+    no default system message: when the caller provides none (and no
+    tools), the ``<system>`` block is omitted entirely. Shares
+    :class:`renderers.laguna_xs2.LagunaXS2Renderer`, which selects the
+    variant from ``config.name``.
+    """
+
+    name: Literal["laguna-xs-2.1"] = "laguna-xs-2.1"
+
+    enable_thinking: bool = False
+    """See :class:`LagunaXS2RendererConfig.enable_thinking`."""
+
+    render_assistant_messages_raw: bool = False
+    """See :class:`LagunaXS2RendererConfig.render_assistant_messages_raw`."""
+
+
 class Llama3RendererConfig(BaseRendererConfig):
     """Llama-3.x Instruct renderer config.
 
@@ -576,6 +597,7 @@ RendererConfig = Annotated[
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
+        LagunaXS21RendererConfig,
         Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
@@ -613,6 +635,7 @@ _CONFIG_BY_NAME: dict[str, type[BaseRendererConfig]] = {
     "kimi-k2": KimiK2RendererConfig,
     "kimi-k2.5": KimiK25RendererConfig,
     "laguna-xs.2": LagunaXS2RendererConfig,
+    "laguna-xs-2.1": LagunaXS21RendererConfig,
     "llama-3": Llama3RendererConfig,
     "minimax-m2": MiniMaxM2RendererConfig,
     "nemotron-3": Nemotron3RendererConfig,
@@ -659,6 +682,7 @@ __all__ = [
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaXS2RendererConfig",
+    "LagunaXS21RendererConfig",
     "Llama3RendererConfig",
     "MiniMaxM2RendererConfig",
     "Nemotron3RendererConfig",

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/renderers/laguna_xs2.py

@@ -1,4 +1,4 @@
-"""Laguna-XS.2 Renderer.
+"""Laguna-XS.2 / XS-2.1 Renderer.
 
 Main properties:
 - Prefix is the single token ``〈|EOS|〉`` (also the EOS / stop token).
@@ -12,10 +12,13 @@ Main properties:
 - Tool calls: ``<tool_call>`` / ``</tool_call>`` ARE single tokens, but the
   inner ``<arg_key>`` / ``</arg_key>`` / ``<arg_value>`` / ``</arg_value>``
   markers are plain text — parsed via regex on the decoded inner block.
-- The template bakes in a default system prompt when ``messages[0]`` is not
-  a system message. The system block also contains the tools section (under
-  a ``### Tools`` header with an ``<available_tools>`` listing and prose
-  format instructions that vary on ``enable_thinking``).
+- XS.2's template bakes in a default system prompt when ``messages[0]`` is
+  not a system message; XS-2.1's (otherwise byte-identical) template
+  removed it, so no system message and no tools means no ``<system>``
+  block at all. The config's ``name`` selects the variant. The system
+  block also contains the tools section (under a ``### Tools`` header
+  with an ``<available_tools>`` listing and prose format instructions
+  that vary on ``enable_thinking``).
 - Reasoning is rendered for every assistant message — no last-user-index
   gating. ``thinking_retention`` is accepted for protocol uniformity but
   is effectively a no-op since past reasoning is preserved by default.
@@ -39,7 +42,7 @@ from renderers.base import (
     resolve_thinking_retention,
     should_rerender_for_thinking_retention,
 )
-from renderers.configs import LagunaXS2RendererConfig
+from renderers.configs import LagunaXS2RendererConfig, LagunaXS21RendererConfig
 from renderers.parsing import parse_laguna_xs2
 
 _DEFAULT_SYSTEM_MESSAGE = (
@@ -81,7 +84,7 @@ class LagunaXS2Renderer:
     def __init__(
         self,
         tokenizer: PreTrainedTokenizer,
-        config: LagunaXS2RendererConfig | None = None,
+        config: LagunaXS2RendererConfig | LagunaXS21RendererConfig | None = None,
     ):
         self._tokenizer = tokenizer
         self.config = config or LagunaXS2RendererConfig()
@@ -89,6 +92,11 @@ class LagunaXS2Renderer:
             self.config,
             "all",
         )
+        # XS.2's template bakes in a default system prompt; XS-2.1's
+        # (otherwise byte-identical) template removed it.
+        self._default_system_message = (
+            _DEFAULT_SYSTEM_MESSAGE if self.config.name == "laguna-xs.2" else ""
+        )
 
         self._eos = self._token_id("〈|EOS|〉")
         self._think = self._token_id("<think>")
@@ -184,7 +192,7 @@ class LagunaXS2Renderer:
         emit_special(self._eos, -1, is_sampled=False, is_content=False)
 
         # ── System header (absorbs messages[0] if it's a system message) ──
-        system_content = _DEFAULT_SYSTEM_MESSAGE
+        system_content = self._default_system_message
         system_msg_idx = -1
         caller_has_system = bool(messages and messages[0].get("role") == "system")
         if caller_has_system:
@@ -196,16 +204,12 @@ class LagunaXS2Renderer:
         # gate: when the caller passes an empty system message and no tools,
         # the whole ``<system>...</system>`` block is omitted.
         if has_sys_content or tools:
-            # The template emits ``<system>\n`` then conditionally a second
-            # ``\n``. Bundle those into one emit so BPE merges ``\n\n`` into
-            # its single-token form (rather than two ``\n`` atoms).
-            emit_text(
-                "<system>\n\n" if has_sys_content else "<system>\n",
-                -1,
-                is_sampled=False,
-                is_content=False,
-            )
             if has_sys_content:
+                # The template emits ``<system>\n`` then a second ``\n``
+                # before the system body. Bundle those into one emit so BPE
+                # merges ``\n\n`` into its single-token form (rather than
+                # two ``\n`` atoms).
+                emit_text("<system>\n\n", -1, is_sampled=False, is_content=False)
                 # If the caller provided system content, it's body bytes;
                 # otherwise this is the default system prompt (scaffold).
                 sys_is_content = caller_has_system
@@ -224,6 +228,11 @@ class LagunaXS2Renderer:
                     if self.config.enable_thinking
                     else _TOOLS_FOOTER_NO_THINKING
                 )
+                if not has_sys_content:
+                    # No system body: ``<system>\n`` runs straight into the
+                    # tools header's ``\n\n`` — encode them together so BPE
+                    # merges the ``\n\n\n`` seam as the template does.
+                    tool_text = "<system>\n" + tool_text
                 emit_text(tool_text, -1, is_sampled=False, is_content=False)
             emit_text("\n</system>\n", -1, is_sampled=False, is_content=False)
 

{renderers-0.1.8.dev53 → renderers-0.1.8.dev54}/tests/conftest.py

@@ -37,6 +37,9 @@ RENDERER_MODELS = [
     # name (auto → MODEL_RENDERER_MAP → nemotron-3-ultra).
     ("nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "auto"),
     ("poolside/Laguna-XS.2", "auto"),
+    # XS-2.1 resolves to the `laguna-xs-2.1` config variant (no default
+    # system message) via the model name.
+    ("poolside/Laguna-XS-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