renderers 0.1.8.dev39__tar.gz → 0.1.8.dev41__tar.gz

{renderers-0.1.8.dev39 → renderers-0.1.8.dev41}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.8.dev39
+Version: 0.1.8.dev41
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE

{renderers-0.1.8.dev39 → renderers-0.1.8.dev41}/renderers/__init__.py

@@ -52,6 +52,7 @@ from renderers.configs import (
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaXS2RendererConfig,
+    Llama3RendererConfig,
     MiniMaxM2RendererConfig,
     Nemotron3RendererConfig,
     Qwen35RendererConfig,
@@ -82,6 +83,7 @@ _LAZY_RENDERERS: dict[str, str] = {
     "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 @@ __all__ = [
     "KimiK2RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
+    "Llama3Renderer",
+    "Llama3RendererConfig",
     "MULTIMODAL_MODELS",
     "Message",
     "MiniMaxM2Renderer",

{renderers-0.1.8.dev39 → renderers-0.1.8.dev41}/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.dev39'
-__version_tuple__ = version_tuple = (0, 1, 8, 'dev39')
+__version__ = version = '0.1.8.dev41'
+__version_tuple__ = version_tuple = (0, 1, 8, 'dev41')
 
 __commit_id__ = commit_id = None

{renderers-0.1.8.dev39 → renderers-0.1.8.dev41}/renderers/base.py

@@ -1037,9 +1037,22 @@ MODEL_RENDERER_MAP: dict[str, str] = {
     "moonshotai/Kimi-K2-Instruct": "kimi-k2",
     "moonshotai/Kimi-K2.5": "kimi-k2.5",
     "moonshotai/Kimi-K2.6": "kimi-k2.5",
-    # Nemotron 3.
+    # Nemotron 3. Nano / Super share one chat-template variant; the Ultra
+    # checkpoints use the Ultra variant — the renderer auto-selects it from
+    # the model name (see ``nemotron3._ULTRA_DEFAULTS``). BF16 and FP8 share the
+    # same tokenizer and template.
     "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": "nemotron-3",
     "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16": "nemotron-3",
+    "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16": "nemotron-3",
+    "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-FP8": "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 +1342,7 @@ 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 +1365,7 @@ def _populate_registry():
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,
+            "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "gpt-oss": GptOssRenderer,
         }

{renderers-0.1.8.dev39 → renderers-0.1.8.dev41}/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."""
 
@@ -337,6 +362,26 @@ class Nemotron3RendererConfig(BaseRendererConfig):
     """When ``True``, the generation prompt includes ``<think>``. Mirrors
     the chat template's ``enable_thinking`` kwarg."""
 
+    ultra: bool | None = None
+    """Select the Nemotron-3 **Ultra** chat-template variant.
+
+    ``None`` (default) auto-detects from the model name (see
+    ``renderers.nemotron3._ULTRA_DEFAULTS``): the Ultra checkpoints resolve
+    to ``True``; Nano / Super and unknown checkpoints to ``False``. Set
+    explicitly to force a variant — e.g. an Ultra fine-tune or a
+    locally-pathed checkpoint whose ``name_or_path`` isn't in the table.
+
+    Ultra's template differs from Nano/Super: the reasoning block is glued
+    as ``<think>\\n{reasoning}</think>{content}`` (no ``\\n`` around
+    ``</think>``), truncated historical turns collapse to
+    ``<think></think>{content}`` (no ``\\n``), and the thinking-truncation
+    boundary follows the template's ``loop.index0 < last_user_idx`` rule
+    (drop thinking on every assistant turn before the last user message).
+
+    Not a chat-template kwarg — it picks which template the renderer
+    mirrors, not a variable passed into one — so it's listed in
+    ``_internal_fields`` and excluded from ``template_field_names()``."""
+
     truncate_history_thinking: bool = True
     """When ``False``, keep ``<think>{reasoning}</think>`` on past-cycle
     assistant turns instead of dropping them. Mirrors the chat
@@ -344,6 +389,15 @@ class Nemotron3RendererConfig(BaseRendererConfig):
     ``preserve_all_thinking`` / ``preserve_thinking_between_tool_calls``
     — see :class:`BaseRendererConfig` for the contract."""
 
+    # ``ultra`` is a template-variant SELECTOR — it picks which template the
+    # renderer mirrors (Ultra vs Nano/Super), not a variable passed into one;
+    # there is no ``ultra`` Jinja variable. Marked internal so the parity
+    # matrix doesn't cross it as a template field. Same ``_internal_fields``
+    # mechanism DeepSeek-V3 uses for its no-op ``enable_thinking``, for a
+    # different underlying reason (theirs is an ignored kwarg, this is a
+    # variant switch).
+    _internal_fields = frozenset({"ultra"})
+
 
 class DeepSeekV3RendererConfig(BaseRendererConfig):
     """DeepSeek V3 renderer config.
@@ -381,6 +435,7 @@ RendererConfig = Annotated[
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
+        Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
         DeepSeekV3RendererConfig,
@@ -415,6 +470,7 @@ _CONFIG_BY_NAME: dict[str, type[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 +513,7 @@ __all__ = [
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaXS2RendererConfig",
+    "Llama3RendererConfig",
     "MiniMaxM2RendererConfig",
     "Nemotron3RendererConfig",
     "Qwen35RendererConfig",