renderers 0.1.8.dev56__tar.gz → 0.1.8.dev58__tar.gz

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.8.dev56
+Version: 0.1.8.dev58
 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`, `gpt-oss`, and `hy3`. 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`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
 
 ## API
 

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/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`, `gpt-oss`, and `hy3`. 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`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
 
 ## API
 

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/docs/renderer-config.md

@@ -26,6 +26,7 @@ chat-template kwargs. Those fields are covered by parity tests against
 | Renderer | Config class | Template fields | Renderer-only fields |
 | --- | --- | --- | --- |
 | Qwen3 | `Qwen3RendererConfig` | `enable_thinking` | - |
+| PrimeIntellect Qwen3 | `PrimeQwen3RendererConfig` | - | - |
 | Qwen3.5 | `Qwen35RendererConfig` | `enable_thinking`, `add_vision_id` | `image_cache_max` |
 | Qwen3.6 | `Qwen36RendererConfig` | `enable_thinking`, `add_vision_id`, `preserve_thinking` | `image_cache_max` |
 | Qwen3-VL | `Qwen3VLRendererConfig` | `add_vision_id` | `image_cache_max` |
@@ -136,6 +137,7 @@ the knobs its template actually exposes:
 | DeepSeek R1 | `template` |
 | MiniMax M2 | `tool_cycle` |
 | DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / XS-2.1, Llama 3 | `all` |
+| PrimeIntellect Qwen3 | `all` |
 
 Config construction raises when an explicit template knob directly contradicts
 an explicit generic bridge policy. For example:

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/renderers/__init__.py

@@ -60,6 +60,7 @@ from renderers.configs import (
     MiniMaxM2RendererConfig,
     Nemotron3RendererConfig,
     Nemotron3UltraRendererConfig,
+    PrimeQwen3RendererConfig,
     Qwen35RendererConfig,
     Qwen36RendererConfig,
     Qwen3RendererConfig,
@@ -95,6 +96,7 @@ _LAZY_RENDERERS: dict[str, str] = {
     "MiniMaxM2Renderer": "renderers.minimax_m2",
     "Nemotron3Renderer": "renderers.nemotron3",
     "Nemotron3UltraRenderer": "renderers.nemotron3",
+    "PrimeQwen3Renderer": "renderers.prime_qwen3",
     "Qwen35Renderer": "renderers.qwen35",
     "Qwen36Renderer": "renderers.qwen36",
     "Qwen3Renderer": "renderers.qwen3",
@@ -163,6 +165,8 @@ __all__ = [
     "ParsedResponse",
     "ParsedToolCall",
     "PlaceholderRange",
+    "PrimeQwen3Renderer",
+    "PrimeQwen3RendererConfig",
     "Qwen35Renderer",
     "Qwen35RendererConfig",
     "Qwen36Renderer",

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/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.dev56'
-__version_tuple__ = version_tuple = (0, 1, 8, 'dev56')
+__version__ = version = '0.1.8.dev58'
+__version_tuple__ = version_tuple = (0, 1, 8, 'dev58')
 
 __commit_id__ = commit_id = None

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/renderers/base.py

@@ -999,6 +999,10 @@ MODEL_RENDERER_MAP: dict[str, str] = {
     "Qwen/Qwen3-30B-A3B-Instruct-2507": "qwen3",
     "Qwen/Qwen3-30B-A3B-Thinking-2507": "qwen3",
     "Qwen/Qwen3-235B-A22B": "qwen3",
+    # PrimeIntellect Qwen3 — both sizes share the same Qwen3-Coder-style
+    # template with XML tool definitions and calls.
+    "PrimeIntellect/Qwen3-0.6B": "prime-qwen3",
+    "PrimeIntellect/Qwen3-1.7B": "prime-qwen3",
     # Qwen3.5. All seven sizes share the same renderer. The 4B / 9B /
     # 35B-A3B / 122B-A10B / 397B-A17B chat template defaults
     # ``enable_thinking=true`` (open ``<think>\n`` at the gen prompt);
@@ -1315,6 +1319,7 @@ def _populate_registry():
     from renderers.llama_3 import Llama3Renderer
     from renderers.minimax_m2 import MiniMaxM2Renderer
     from renderers.nemotron3 import Nemotron3Renderer, Nemotron3UltraRenderer
+    from renderers.prime_qwen3 import PrimeQwen3Renderer
     from renderers.qwen3 import Qwen3Renderer
     from renderers.qwen3_vl import Qwen3VLRenderer
     from renderers.qwen35 import Qwen35Renderer
@@ -1324,6 +1329,7 @@ def _populate_registry():
         {
             "default": DefaultRenderer,
             "qwen3": Qwen3Renderer,
+            "prime-qwen3": PrimeQwen3Renderer,
             "qwen3-vl": Qwen3VLRenderer,
             "qwen3.5": Qwen35Renderer,
             "qwen3.6": Qwen36Renderer,
@@ -1588,6 +1594,7 @@ def build_training_sample(
     role_to_mask: Callable[[Message], bool] | None = None,
     tools: list[ToolSpec] | None = None,
     content_sft_roles: "set[str] | frozenset[str] | None" = None,
+    ensure_final_stop: bool = False,
 ) -> RenderedTrainingSample:
     """Build a :class:`RenderedTrainingSample` for supervised training.
 
@@ -1643,6 +1650,17 @@ def build_training_sample(
     or hand-coded renderers that haven't been wired up yet) ignore
     ``content_sft_roles`` silently — falling back to the original
     ``role_to_mask`` + ``sampled_mask`` behaviour.
+
+    ``ensure_final_stop`` appends the renderer's canonical stop token
+    when the sample ends with an assistant message that the template
+    leaves unterminated. Some templates close an assistant turn only
+    via the *next* message's role marker (e.g. GLM's ``<|user|>`` /
+    ``<|observation|>``), so a final assistant message renders with no
+    stop token at all. No-op when the template already closes the turn
+    in-message (ChatML ``<|im_end|>``, Llama ``<|eot_id|>``); where it
+    fires, the output intentionally diverges from ``apply_chat_template``.
+    Ignored for renderers without ``sampled_mask`` (``DefaultRenderer``) —
+    the close of an opaque template can't be located reliably.
     """
     rendered = renderer.render(messages, tools=tools)
     has_sampled_info = len(rendered.sampled_mask) == len(rendered.token_ids)
@@ -1683,6 +1701,25 @@ def build_training_sample(
         else:
             loss_mask.append(role_to_mask(msg))
 
+    token_ids = list(rendered.token_ids)
+    # Requires sampled_mask (opaque templates hide the assistant close)
+    # and a final assistant message the role filter trains.
+    if (
+        ensure_final_stop
+        and has_sampled_info
+        and messages[-1].get("role") == "assistant"
+        and (role_to_mask is None or role_to_mask(messages[-1]))
+    ):
+        stop_ids = set(renderer.get_stop_token_ids())
+        last_trainable = next(
+            (k for k in range(len(loss_mask) - 1, -1, -1) if loss_mask[k]), None
+        )
+        if last_trainable is None or token_ids[last_trainable] not in stop_ids:
+            token_ids.append(renderer.get_stop_token_ids()[0])
+            # loss_mask=True marks the token as trainable — the appended
+            # stop is a training target, like any sampled token.
+            loss_mask.append(True)
+
     # Surface the multimodal payload for VLM renderers. ``None`` for text
     # renderers and for text-only samples (empty media) so downstream
     # ``multi_modal_data is not None`` is a reliable "has media" check.
@@ -1690,12 +1727,12 @@ def build_training_sample(
     if mm is not None and mm.is_empty():
         mm = None
     mm_token_type_ids = (
-        _build_mm_token_type_ids(mm.mm_placeholders, len(rendered.token_ids))
+        _build_mm_token_type_ids(mm.mm_placeholders, len(token_ids))
         if mm is not None and mm.mm_placeholders
         else None
     )
     return RenderedTrainingSample(
-        token_ids=rendered.token_ids,
+        token_ids=token_ids,
         loss_mask=loss_mask,
         multi_modal_data=mm,
         mm_token_type_ids=mm_token_type_ids,

{renderers-0.1.8.dev56 → renderers-0.1.8.dev58}/renderers/configs.py

@@ -181,6 +181,12 @@ class Qwen3RendererConfig(BaseRendererConfig):
     ``enable_thinking`` kwarg."""
 
 
+class PrimeQwen3RendererConfig(BaseRendererConfig):
+    """PrimeIntellect Qwen3 renderer config."""
+
+    name: Literal["prime-qwen3"] = "prime-qwen3"
+
+
 class Qwen35RendererConfig(BaseRendererConfig):
     """Qwen3.5 renderer config."""
 
@@ -649,6 +655,7 @@ RendererConfig = Annotated[
         AutoRendererConfig,
         DefaultRendererConfig,
         Qwen3RendererConfig,
+        PrimeQwen3RendererConfig,
         Qwen35RendererConfig,
         Qwen36RendererConfig,
         Qwen3VLRendererConfig,
@@ -688,6 +695,7 @@ _CONFIG_BY_NAME: dict[str, type[BaseRendererConfig]] = {
     "auto": AutoRendererConfig,
     "default": DefaultRendererConfig,
     "qwen3": Qwen3RendererConfig,
+    "prime-qwen3": PrimeQwen3RendererConfig,
     "qwen3.5": Qwen35RendererConfig,
     "qwen3.6": Qwen36RendererConfig,
     "qwen3-vl": Qwen3VLRendererConfig,
@@ -752,6 +760,7 @@ __all__ = [
     "MiniMaxM2RendererConfig",
     "Nemotron3RendererConfig",
     "Nemotron3UltraRendererConfig",
+    "PrimeQwen3RendererConfig",
     "Qwen35RendererConfig",
     "Qwen36RendererConfig",
     "Qwen3RendererConfig",