renderers 0.1.8.dev40__tar.gz → 0.1.8.dev42__tar.gz

{renderers-0.1.8.dev40 → renderers-0.1.8.dev42}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.8.dev40
+Version: 0.1.8.dev42
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE

{renderers-0.1.8.dev40 → renderers-0.1.8.dev42}/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.dev40 → renderers-0.1.8.dev42}/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.dev40'
-__version_tuple__ = version_tuple = (0, 1, 8, 'dev40')
+__version__ = version = '0.1.8.dev42'
+__version_tuple__ = version_tuple = (0, 1, 8, 'dev42')
 
 __commit_id__ = commit_id = None

{renderers-0.1.8.dev40 → renderers-0.1.8.dev42}/renderers/base.py

@@ -1045,6 +1045,14 @@ 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",
     "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.
@@ -1334,6 +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
@@ -1356,6 +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.dev40 → renderers-0.1.8.dev42}/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."""
 
@@ -410,6 +435,7 @@ RendererConfig = Annotated[
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
+        Llama3RendererConfig,
         MiniMaxM2RendererConfig,
         Nemotron3RendererConfig,
         DeepSeekV3RendererConfig,
@@ -444,6 +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,
@@ -486,6 +513,7 @@ __all__ = [
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaXS2RendererConfig",
+    "Llama3RendererConfig",
     "MiniMaxM2RendererConfig",
     "Nemotron3RendererConfig",
     "Qwen35RendererConfig",

{renderers-0.1.8.dev40 → renderers-0.1.8.dev42}/renderers/kimi_k25.py

@@ -42,7 +42,7 @@ from renderers.base import (
     trim_to_turn_close,
 )
 from renderers.configs import KimiK25RendererConfig
-from renderers.parsing import parse_kimi_k2_section
+from renderers.parsing import _reasoning_end_token_index, parse_kimi_k2_section
 from renderers.qwen3_vl import (
     _image_hash,
     _is_image_part,
@@ -452,6 +452,13 @@ def _parse_kimi_k2_response(
             ids = ids[:i]
             break
 
+    # Reasoning first: a tool-call section the model drafts *inside* its
+    # <think> trace must not be parsed as a real call (regression #78 — cf.
+    # parse_qwen3). K2.5 renders </think> as text, so locate the boundary by
+    # decoding; the section scan then starts past it. content_ids still begins
+    # at 0, so the </think> text-split below recovers reasoning unchanged.
+    reasoning_end = _reasoning_end_token_index(tokenizer, ids)
+
     # Token-ID path — produces spans. Only run if every relevant special
     # token resolved at init (i.e. is in the tokenizer's vocab).
     tool_calls: list[ParsedToolCall] = []
@@ -471,6 +478,7 @@ def _parse_kimi_k2_response(
             tool_call_begin_id=tool_call_begin_id,
             tool_call_argument_begin_id=tool_call_argument_begin_id,
             tool_call_end_id=tool_call_end_id,
+            scan_start=reasoning_end,
         )
         text = (
             tokenizer.decode(content_ids, skip_special_tokens=False)
@@ -481,9 +489,13 @@ def _parse_kimi_k2_response(
         text = tokenizer.decode(ids, skip_special_tokens=False) if ids else ""
 
     # Fallback path: model emitted literal-text section delimiters (singular
-    # variant) rather than special tokens. Spans unavailable here.
+    # variant) rather than special tokens. Spans unavailable here. Start the
+    # search past the first </think> so a literal section drafted inside the
+    # reasoning trace isn't matched as a real call (regression #78).
     if not tool_calls:
-        tc_match = _TOOL_CALLS_SECTION_RE.search(text)
+        think_close = text.find("</think>")
+        search_from = think_close + len("</think>") if think_close != -1 else 0
+        tc_match = _TOOL_CALLS_SECTION_RE.search(text, search_from)
         if tc_match:
             text = text[: tc_match.start()]
             tool_section = (