feat: add Llama-3 renderer for Llama-3.2-1B/3B-Instruct

renderers/__init__.py

@@ -52,6 +52,7 @@
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaXS2RendererConfig,
+    Llama3RendererConfig,
     MiniMaxM2RendererConfig,
     Nemotron3RendererConfig,
     Qwen35RendererConfig,
@@ -82,6 +83,7 @@
     "KimiK25Renderer": "renderers.kimi_k25",
     "KimiK2Renderer": "renderers.kimi_k2",
     "LagunaXS2Renderer": "renderers.laguna_xs2",
+    "Llama3Renderer": "renderers.llama_3",
     "MiniMaxM2Renderer": "renderers.minimax_m2",
     "Nemotron3Renderer": "renderers.nemotron3",
     "Qwen35Renderer": "renderers.qwen35",
@@ -130,6 +132,8 @@ def __dir__() -> list[str]:
     "KimiK2RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
+    "Llama3Renderer",
+    "Llama3RendererConfig",
     "MULTIMODAL_MODELS",
     "Message",
     "MiniMaxM2Renderer",

renderers/base.py

@@ -1040,6 +1040,14 @@ def bridge_to_next_turn(self, *args: Any, **kwargs: Any) -> "RenderedTokens | No
     # Nemotron 3.
     "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": "nemotron-3",
     "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16": "nemotron-3",
+    # 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
+    # "26 Jul 2024" — matching the chat template's strftime fallback —
+    # so the renderer is reproducible. Pass ``date_string=...`` at
+    # 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": "laguna-xs.2",
     # GPT-OSS.
@@ -1329,6 +1337,7 @@ def _populate_registry():
     from renderers.kimi_k2 import KimiK2Renderer
     from renderers.kimi_k25 import KimiK25Renderer
     from renderers.laguna_xs2 import LagunaXS2Renderer
+    from renderers.llama_3 import Llama3Renderer
     from renderers.minimax_m2 import MiniMaxM2Renderer
     from renderers.nemotron3 import Nemotron3Renderer
     from renderers.qwen3 import Qwen3Renderer
@@ -1351,6 +1360,7 @@ def _populate_registry():
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,
+            "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "gpt-oss": GptOssRenderer,
         }

renderers/configs.py

@@ -318,6 +318,31 @@ class LagunaXS2RendererConfig(BaseRendererConfig):
     chat template's ``render_assistant_messages_raw`` gate."""
 
 
+class Llama3RendererConfig(BaseRendererConfig):
+    """Llama-3.x Instruct renderer config.
+
+    Llama-3 ships no reasoning channel, so the base ``preserve_*_thinking``
+    flags don't apply: ``Llama3Renderer`` raises ``NotImplementedError``
+    if either is set (matching ``DefaultRenderer``'s contract for the
+    same case). Both fields below mirror real ``apply_chat_template``
+    kwargs.
+    """
+
+    name: Literal["llama-3"] = "llama-3"
+
+    date_string: str = "26 Jul 2024"
+    """``Today Date`` value injected into the system preamble. Pinned to
+    the chat template's ``strftime`` fallback by default so output stays
+    deterministic; override per instance for production runs that want
+    today's date. Mirrors the chat template's ``date_string`` kwarg."""
+
+    tools_in_user_message: bool = True
+    """When ``True`` (default), tool descriptions + JSON signatures inject
+    into the first user message; ``False`` routes them into the system
+    block instead. Mirrors the chat template's ``tools_in_user_message``
+    kwarg."""
+
+
 class MiniMaxM2RendererConfig(BaseRendererConfig):
     """MiniMax M2 / M2.5 renderer config."""
 
@@ -381,6 +406,7 @@ class DeepSeekV3RendererConfig(BaseRendererConfig):
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
+        Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
         DeepSeekV3RendererConfig,
@@ -415,6 +441,7 @@ class DeepSeekV3RendererConfig(BaseRendererConfig):
     "kimi-k2": KimiK2RendererConfig,
     "kimi-k2.5": KimiK25RendererConfig,
     "laguna-xs.2": LagunaXS2RendererConfig,
+    "llama-3": Llama3RendererConfig,
     "minimax-m2": MiniMaxM2RendererConfig,
     "nemotron-3": Nemotron3RendererConfig,
     "deepseek-v3": DeepSeekV3RendererConfig,
@@ -457,6 +484,7 @@ def config_from_name(name: str) -> BaseRendererConfig | None:
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaXS2RendererConfig",
+    "Llama3RendererConfig",
     "MiniMaxM2RendererConfig",
     "Nemotron3RendererConfig",
     "Qwen35RendererConfig",