renderers 0.1.8.dev54__tar.gz → 0.1.8.dev56__tar.gz

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.8.dev54
+Version: 0.1.8.dev56
 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`, and `gpt-oss`. 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`, `nemotron-3`, `nemotron-3-ultra`, `llama-3`, `gpt-oss`, and `hy3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
 
 ## API
 

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/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`, and `gpt-oss`. 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`, `nemotron-3`, `nemotron-3-ultra`, `llama-3`, `gpt-oss`, and `hy3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
 
 ## API
 

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/docs/renderer-config.md

@@ -32,9 +32,11 @@ chat-template kwargs. Those fields are covered by parity tests against
 | GLM-5 / 5.1 | `GLM5RendererConfig` / `GLM51RendererConfig` | `enable_thinking`, `clear_thinking` | - |
 | GLM-4.5 | `GLM45RendererConfig` | `enable_thinking` | - |
 | gpt-oss | `GptOssRendererConfig` | `reasoning_effort`, `conversation_start_date` | `use_system_prompt`, `knowledge_cutoff`, `model_identity`, `auto_drop_analysis` |
+| Hy3 | `Hy3RendererConfig` | `reasoning_effort`, `preserved_thinking`, `is_training`, `raw_last_assistant`, `fallback_strategy` | - |
 | 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 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` | - |
@@ -128,11 +130,12 @@ the knobs its template actually exposes:
 | GLM-5 / 5.1 | `clear_thinking=False -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
 | GLM-4.5 | `enable_thinking=False -> all`, else `tool_cycle` |
 | 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` |
 | DeepSeek R1 | `template` |
 | MiniMax M2 | `tool_cycle` |
-| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2, Llama 3 | `all` |
+| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / XS-2.1, Llama 3 | `all` |
 
 Config construction raises when an explicit template knob directly contradicts
 an explicit generic bridge policy. For example:

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/renderers/__init__.py

@@ -51,6 +51,7 @@ from renderers.configs import (
     GLM51RendererConfig,
     GLM5RendererConfig,
     GptOssRendererConfig,
+    Hy3RendererConfig,
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaXS2RendererConfig,
@@ -85,8 +86,10 @@ _LAZY_RENDERERS: dict[str, str] = {
     "GLM51Renderer": "renderers.glm5",
     "GLM5Renderer": "renderers.glm5",
     "GptOssRenderer": "renderers.gpt_oss",
+    "Hy3Renderer": "renderers.hy3",
     "KimiK25Renderer": "renderers.kimi_k25",
     "KimiK2Renderer": "renderers.kimi_k2",
+    "LagunaXS21Renderer": "renderers.laguna_xs2",
     "LagunaXS2Renderer": "renderers.laguna_xs2",
     "Llama3Renderer": "renderers.llama_3",
     "MiniMaxM2Renderer": "renderers.minimax_m2",
@@ -133,6 +136,8 @@ __all__ = [
     "GLM5RendererConfig",
     "GptOssRenderer",
     "GptOssRendererConfig",
+    "Hy3Renderer",
+    "Hy3RendererConfig",
     "ImagePart",
     "KimiK25Renderer",
     "KimiK25RendererConfig",
@@ -140,6 +145,7 @@ __all__ = [
     "KimiK2RendererConfig",
     "LagunaXS2Renderer",
     "LagunaXS2RendererConfig",
+    "LagunaXS21Renderer",
     "LagunaXS21RendererConfig",
     "Llama3Renderer",
     "Llama3RendererConfig",

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/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.dev54'
-__version_tuple__ = version_tuple = (0, 1, 8, 'dev54')
+__version__ = version = '0.1.8.dev56'
+__version_tuple__ = version_tuple = (0, 1, 8, 'dev56')
 
 __commit_id__ = commit_id = None

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/renderers/base.py

@@ -1060,13 +1060,19 @@ 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. 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. The two checkpoints ship different chat templates,
+    # each mirrored by its own renderer class.
     "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",
+    # Tencent Hunyuan Hy3 (295B-A21B MoE). The FP8 checkpoint shares the same
+    # tokenizer and chat template. Hy3-preview is deliberately unmapped: it
+    # ships an older, incompatible template (un-suffixed special tokens,
+    # ``interleaved_thinking`` instead of ``preserved_thinking``).
+    "tencent/Hy3": "hy3",
+    "tencent/Hy3-FP8": "hy3",
 }
 
 
@@ -1302,9 +1308,10 @@ def _populate_registry():
     from renderers.glm5 import GLM5Renderer, GLM51Renderer
     from renderers.glm45 import GLM45Renderer
     from renderers.gpt_oss import GptOssRenderer
+    from renderers.hy3 import Hy3Renderer
     from renderers.kimi_k2 import KimiK2Renderer
     from renderers.kimi_k25 import KimiK25Renderer
-    from renderers.laguna_xs2 import LagunaXS2Renderer
+    from renderers.laguna_xs2 import LagunaXS2Renderer, LagunaXS21Renderer
     from renderers.llama_3 import Llama3Renderer
     from renderers.minimax_m2 import MiniMaxM2Renderer
     from renderers.nemotron3 import Nemotron3Renderer, Nemotron3UltraRenderer
@@ -1326,10 +1333,11 @@ def _populate_registry():
             "minimax-m2": MiniMaxM2Renderer,
             "deepseek-v3": DeepSeekV3Renderer,
             "deepseek-r1": DeepSeekR1Renderer,
+            "hy3": Hy3Renderer,
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,
-            "laguna-xs-2.1": LagunaXS2Renderer,
+            "laguna-xs-2.1": LagunaXS21Renderer,
             "llama-3": Llama3Renderer,
             "nemotron-3": Nemotron3Renderer,
             "nemotron-3-ultra": Nemotron3UltraRenderer,
@@ -1764,6 +1772,8 @@ def _get_offset_tokenizer(tokenizer):
 def attribute_text_segments(
     tokenizer,
     segments: "list[tuple[str, bool]]",
+    *,
+    overlap_is_content: bool = False,
 ) -> "list[tuple[int, bool]]":
     """Tokenize concatenated segments as a single BPE pass and return
     ``(token_id, is_content)`` pairs.
@@ -1782,6 +1792,16 @@ def attribute_text_segments(
     tokens (rare; usually pre-tokenizer artefacts) are attributed to
     the most recently entered segment.
 
+    ``overlap_is_content=True`` widens the content bit: a token counts
+    as content when *any* of its source characters fall in a content
+    segment, not just its first. Templates whose wrap glues directly
+    onto the body with no whitespace (e.g. ``<user>{content}</user>``)
+    can merge wrap and body bytes into one token; under the first-char
+    policy such a token would land on the wrap side and the body would
+    no longer be recoverable from the content run. Over-inclusion keeps
+    every body byte inside the ``is_content=True`` run at the cost of a
+    few adjacent wrap bytes.
+
     Requires a HuggingFace fast tokenizer with offset tracking. Every
     model in ``MODEL_RENDERER_MAP`` ships one, so the offset lookup
     always succeeds for tokenizers obtained via :func:`load_tokenizer`.
@@ -1818,13 +1838,25 @@ def attribute_text_segments(
 
     out: list[tuple[int, bool]] = []
     last_is_content = spans[-1][2] if spans else False
-    for tok_id, (start, _end) in zip(token_ids, offsets):
+    for tok_id, (start, end) in zip(token_ids, offsets):
         if start >= total_len:
             # Token's character offset is past every segment (shouldn't
             # normally happen for add_special_tokens=False, but defensive
             # against tokenizer-specific edge cases).
             out.append((tok_id, last_is_content))
             continue
+        if overlap_is_content and end > start:
+            out.append(
+                (
+                    tok_id,
+                    any(
+                        seg_is_content
+                        for seg_start, seg_end, seg_is_content in spans
+                        if seg_start < end and start < seg_end
+                    ),
+                )
+            )
+            continue
         # Find the segment that contains `start`. Segments are
         # contiguous and ordered, so a linear scan is fine — the inner
         # loop runs at most len(segments) times per token and segments

{renderers-0.1.8.dev54 → renderers-0.1.8.dev56}/renderers/configs.py

@@ -310,6 +310,70 @@ class GLM45RendererConfig(BaseRendererConfig):
     the chat template's ``enable_thinking`` kwarg."""
 
 
+class Hy3RendererConfig(BaseRendererConfig):
+    """Tencent Hunyuan Hy3 renderer config.
+
+    Hy3 reasons via a ``reasoning_effort`` gate rather than a boolean
+    ``enable_thinking``. ``"no_think"`` (the template default) prefills an
+    empty ``<think></think>`` at the generation prompt so the model answers
+    directly; ``"low"`` / ``"high"`` prefill only the ``<think>`` opener so
+    the model streams reasoning up to a ``</think>`` it emits itself.
+
+    ``preserved_thinking`` mirrors the template kwarg of the same name:
+    ``True`` keeps ``<think>{reasoning}</think>`` on every historical
+    assistant turn; ``False`` collapses past-cycle reasoning to
+    ``<think></think>``, keeping it only on the in-flight turn (after the
+    last user query). ``None`` (default) follows the template's own
+    default — ``True`` when ``tools`` are supplied at render time, ``False``
+    otherwise. Bridge policy tracks the same resolution: ``"all"`` whenever
+    ``preserved_thinking`` resolves to ``True`` for the tools at hand,
+    ``"tool_cycle"`` otherwise.
+    """
+
+    name: Literal["hy3"] = "hy3"
+
+    reasoning_effort: Literal["no_think", "low", "high"] = "no_think"
+    """Reasoning gate. Mirrors the chat template's ``reasoning_effort`` kwarg.
+    ``"no_think"`` prefills ``<think></think>`` at the generation prompt;
+    ``"low"`` / ``"high"`` prefill just ``<think>``."""
+
+    preserved_thinking: bool | None = None
+    """Keep historical assistant reasoning. Mirrors the template's
+    ``preserved_thinking`` kwarg. ``None`` defers to the template default
+    (``True`` with tools, ``False`` without). Bridge policy is ``"all"``
+    whenever this resolves to ``True`` for the tools at hand."""
+
+    is_training: bool = False
+    """Mirrors the template's ``is_training`` kwarg. ``True`` renders SFT
+    targets: reasoning is kept on every assistant turn (regardless of
+    ``preserved_thinking`` / position) and the final assistant is terminated
+    with ``<｜hy_eos｜>``. Leave ``False`` for inference-faithful renders;
+    the training loss mask is normally derived via ``build_training_sample``
+    rather than this flag."""
+
+    raw_last_assistant: bool = False
+    """Mirrors the template's ``raw_last_assistant`` kwarg. When ``True`` a
+    trailing non-tool assistant message is emitted as raw visible content —
+    no ``<think>`` wrap, no ``<｜hy_eos｜>`` — for prefill / continuation."""
+
+    fallback_strategy: Literal["reasoning_toolcall_retry"] | None = None
+    """Mirrors the template's ``fallback_strategy`` kwarg. The sole active
+    value, ``"reasoning_toolcall_retry"``, forces ``reasoning_effort="high"``
+    and suppresses the generation prompt (``add_generation_prompt=False``)."""
+
+    @model_validator(mode="after")
+    def _check_thinking_retention(self):
+        if self.preserved_thinking is not None and self.thinking_retention is not None:
+            implied = "all" if self.preserved_thinking else "tool_cycle"
+            if self.thinking_retention != implied:
+                raise ValueError(
+                    f"preserved_thinking={self.preserved_thinking!r} implies "
+                    f"thinking_retention={implied!r}, which conflicts with "
+                    f"explicit thinking_retention={self.thinking_retention!r}."
+                )
+        return self
+
+
 class GptOssRendererConfig(BaseRendererConfig):
     """OpenAI gpt-oss (harmony) renderer config.
 
@@ -423,24 +487,22 @@ class LagunaXS2RendererConfig(BaseRendererConfig):
 
 
 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``.
+    """Laguna XS-2.1 renderer config.
+
+    XS-2.1's chat template reads a single kwarg, ``enable_thinking``,
+    which gates both the generation prompt (``<think>`` vs ``</think>``)
+    and whether assistant reasoning is rendered into the history at all.
+    Served by :class:`renderers.laguna_xs2.LagunaXS21Renderer`.
     """
 
     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`."""
+    """When ``True``, the generation prompt ends with ``<think>`` and
+    every assistant turn renders ``<think>{reasoning}</think>``; when
+    ``False``, turns open with a bare ``</think>`` and reasoning is not
+    rendered. Mirrors the template's ``enable_thinking`` kwarg and its
+    upstream default."""
 
 
 class Llama3RendererConfig(BaseRendererConfig):
@@ -594,6 +656,7 @@ RendererConfig = Annotated[
         GLM51RendererConfig,
         GLM45RendererConfig,
         GptOssRendererConfig,
+        Hy3RendererConfig,
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
@@ -632,6 +695,7 @@ _CONFIG_BY_NAME: dict[str, type[BaseRendererConfig]] = {
     "glm-5.1": GLM51RendererConfig,
     "glm-4.5": GLM45RendererConfig,
     "gpt-oss": GptOssRendererConfig,
+    "hy3": Hy3RendererConfig,
     "kimi-k2": KimiK2RendererConfig,
     "kimi-k2.5": KimiK25RendererConfig,
     "laguna-xs.2": LagunaXS2RendererConfig,
@@ -679,6 +743,7 @@ __all__ = [
     "GLM51RendererConfig",
     "GLM5RendererConfig",
     "GptOssRendererConfig",
+    "Hy3RendererConfig",
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaXS2RendererConfig",