feat(sft): require renderer-based tokenization

configs/ci/integration/reverse_text_sft/resume.toml

@@ -6,6 +6,10 @@ resume_step = 5
 [model]
 name = "PrimeIntellect/Qwen3-0.6B"
 
+# PrimeIntellect/Qwen3-0.6B ships its own chat template (distinct from Qwen3's)
+[renderer]
+name = "prime-qwen3"
+
 [data]
 name = "PrimeIntellect/Reverse-Text-SFT"
 batch_size = 4

configs/ci/integration/reverse_text_sft/start.toml

@@ -5,6 +5,10 @@ max_steps = 5
 [model]
 name = "PrimeIntellect/Qwen3-0.6B"
 
+# PrimeIntellect/Qwen3-0.6B ships its own chat template (distinct from Qwen3's)
+[renderer]
+name = "prime-qwen3"
+
 [data]
 name = "PrimeIntellect/Reverse-Text-SFT"
 batch_size = 4

configs/ci/integration/reverse_text_sft_lora/resume.toml

@@ -9,6 +9,10 @@ save_adapter_separately = true
 [model]
 name = "PrimeIntellect/Qwen3-0.6B"
 
+# PrimeIntellect/Qwen3-0.6B ships its own chat template (distinct from Qwen3's)
+[renderer]
+name = "prime-qwen3"
+
 [model.lora]
 rank = 8
 target_modules = [

configs/ci/integration/reverse_text_sft_lora/start.toml

@@ -8,6 +8,10 @@ save_adapter_separately = true
 [model]
 name = "PrimeIntellect/Qwen3-0.6B"
 
+# PrimeIntellect/Qwen3-0.6B ships its own chat template (distinct from Qwen3's)
+[renderer]
+name = "prime-qwen3"
+
 [model.lora]
 rank = 8
 target_modules = [

docs/training.md

@@ -144,12 +144,20 @@ Two accepted layouts:
 
 If both columns are present, `messages` takes precedence.
 
-**Tool definitions.** For tool-use SFT, add a `tools` column (OpenAI function-calling format) or `tool_defs` ([`verifiers`](https://github.com/PrimeIntellect-ai/verifiers) rollout format). Each row's value can be either a list of dicts or a JSON-encoded string of a list — both are accepted, and `tool_defs` rows are auto-converted to OAI shape before being passed into the chat template's `tools=...` argument. The `chat_template_kwargs` column, if present, is forwarded verbatim into `apply_chat_template`.
+**Tool definitions and renderer controls.** For tool-use SFT, add a `tools` column (OpenAI function-calling format) or `tool_defs` ([`verifiers`](https://github.com/PrimeIntellect-ai/verifiers) rollout format). Each row's value can be either a list of dicts or a JSON-encoded string of a list — both are accepted, and `tool_defs` rows are auto-converted to OAI shape before being passed into the renderer.
 
-**Position-dependent chat templates.** Multi-turn SFT under the default tokenization path (`build_incremental_token_mask`) requires that tokenizing the first _k_ turns of a conversation be a strict prefix of tokenizing all _n ≥ k_ turns. Qwen3's upstream template _violates_ this — it strips past `<think>` blocks across user turns, silently corrupting the loss mask. Two fixes:
+Renderer-backed SFT reads template controls from the typed `[renderer]` config in the SFT TOML. For example:
+
+```toml
+[renderer]
+name = "qwen3"
+enable_thinking = false
+```
+
+If a model needs another template control, add it to that model's renderer config in `renderers` (for example a new field on the relevant `*RendererConfig`) and consume it in the renderer implementation.
+
+**Renderer-backed tokenization.** SFT tokenization is renderer-only. The [`renderers`](algorithms.md#renderers) package owns message-to-token conversion and loss attribution end-to-end, so position-dependent chat templates (for example templates that strip past `—` blocks across user turns) do not corrupt the loss mask. `[renderer]` defaults to `name = "auto"`; set a typed renderer config only when you need model-specific template controls. Hand-coded renderers ship for Qwen3, Qwen3.5, GLM-5, GLM-4.5, Kimi K2/K2.5, MiniMax M2, DeepSeek V3, Nemotron 3, GPT-OSS.
 
-- **Enable the renderer** (set a typed `[renderer]` config, e.g. `name = "qwen3"`, recommended; defaults to `"auto"` for RL). The [`renderers`](algorithms.md#renderers) package owns tokenization end-to-end and is robust to position-dependent templates. Hand-coded renderers ship for Qwen3, Qwen3.5, GLM-5, GLM-4.5, Kimi K2/K2.5, MiniMax M2, DeepSeek V3, Nemotron 3, GPT-OSS. Not supported for VLMs.
-- **Patched chat template** — the prime-rl–patched checkpoints (e.g. `PrimeIntellect/Qwen3-0.6B`, used in `examples/reverse_text/sft.toml`) ship a chat template that preserves thinking. Or supply your own.
 
 See [Algorithms § Multi-Turn Trajectories](algorithms.md#multi-turn-trajectories) for the full picture.
 

examples/reverse_text/sft.toml

@@ -5,6 +5,10 @@ max_steps = 100
 [model]
 name = "PrimeIntellect/Qwen3-0.6B"
 
+# PrimeIntellect/Qwen3-0.6B ships its own chat template (distinct from Qwen3's)
+[renderer]
+name = "prime-qwen3"
+
 [data]
 name = "willcb/R1-reverse-wikipedia-paragraphs-v1-1000"
 seq_len = 4096

examples/wordle/rl.toml

@@ -30,9 +30,6 @@ max_completion_tokens = 1024
 [inference.parallel]
 dp = 6
 
-# Qwen3 finetune with the standard PI template patch (byte-identical to
-# PrimeIntellect/Qwen3-0.6B base); always re-emits prior <think> blocks.
-# Match that with renderer thinking_retention.
+# Fine-tune inherits the PrimeIntellect Qwen3 template byte-for-byte.
 [orchestrator.renderer]
-name = "qwen3"
-thinking_retention = "all"
+name = "prime-qwen3"

examples/wordle/sft.toml

@@ -5,6 +5,10 @@ max_steps = 20
 [model]
 name = "PrimeIntellect/Qwen3-1.7B"
 
+# PrimeIntellect/Qwen3-1.7B ships its own chat template (distinct from Qwen3's)
+[renderer]
+name = "prime-qwen3"
+
 [data]
 name = "willcb/V3-wordle"
 seq_len = 1024

packages/prime-rl-configs/src/prime_rl/configs/sft.py

@@ -3,7 +3,8 @@
 from typing import Annotated, Literal, TypeAlias
 
 from pydantic import Field, model_validator
-from renderers import RendererConfig
+from renderers import AutoRendererConfig, DefaultRendererConfig, RendererConfig
+from renderers.base import MODEL_RENDERER_MAP
 
 from prime_rl.configs.shared import (
     EnvVars,
@@ -30,7 +31,7 @@ class BaseDataConfig(BaseConfig):
     batch_size: int = Field(128, ge=1)
     """Global batch size."""
 
-    seq_len: int = Field(128, ge=1)
+    seq_len: int = Field(256, ge=1)
     """Sequence length."""
 
     pack_function: Literal["cat", "stack"] = "cat"
@@ -51,6 +52,12 @@ def validate_batch_size(self):
 class FakeDataConfig(BaseDataConfig):
     type: Literal["fake"] = "fake"
 
+    seq_len: int = Field(128, ge=1)
+    """Sequence length."""
+
+    pack_function: Literal["cat", "stack"] = "cat"
+    """Sample packing strategy."""
+
     length: Literal["fixed", "variable"] = "fixed"
     """Use fixed-length samples or variable-length samples."""
 
@@ -175,13 +182,8 @@ class SFTConfig(BaseConfig):
 
     tokenizer: TokenizerConfig = TokenizerConfig()
 
-    renderer: RendererConfig | None = None
-    """Typed renderer config (``renderers.RendererConfig`` discriminated
-    union). When set, SFT tokenizes samples through the ``renderers``
-    library (single ``render()`` + ``message_indices`` mask) instead of
-    the default ``build_incremental_token_mask`` path. Required for chat
-    templates that render position-dependently (e.g. Qwen3, Qwen3.5).
-    ``None`` (default) uses the legacy tokenization path."""
+    renderer: RendererConfig = AutoRendererConfig()
+    """Renderer config. Defaults to auto-selecting from the tokenizer model name."""
 
     data: DataConfig = SFTDataConfig()
 
@@ -225,8 +227,8 @@ class SFTConfig(BaseConfig):
     dist_timeout_seconds: int = 3600
     """Timeout in seconds for torch distributed ops."""
 
-    loss_impl: Literal["liger", "torch", "liger_fused", "quack_fused"] = "torch"
-    """Cross-entropy loss implementation. ``liger_fused`` fuses the lm_head projection with the CE loss to avoid materializing full logits. ``quack_fused`` uses quack-kernels for chunked linear + CE with CuTe DSL CUDA kernels."""
+    loss_impl: Literal["liger", "torch", "liger_fused", "quack_fused"] = "liger_fused"
+    """Cross-entropy loss implementation. Defaults to fused Liger loss to avoid materializing full logits."""
 
     heartbeat: HeartbeatConfig | None = None
     """BetterStack heartbeat configuration for monitoring training progress."""
@@ -271,6 +273,28 @@ def validate_deployment(self):
             raise ValueError("Must use SLURM for multi-node deployment.")
         return self
 
+    @model_validator(mode="after")
+    def validate_typed_renderer(self):
+        """Require a typed renderer whenever SFT renders real samples."""
+        if self.data.type == "fake" and self.val is None:
+            return self
+
+        model_id = self.tokenizer.name or self.model.name
+        if isinstance(self.renderer, AutoRendererConfig):
+            if model_id in MODEL_RENDERER_MAP:
+                return self
+            reason = f"no typed renderer is registered for {model_id!r}"
+        elif isinstance(self.renderer, DefaultRendererConfig):
+            reason = "renderer.name='default' selects DefaultRenderer"
+        else:
+            return self
+
+        raise ValueError(
+            f"SFT requires a typed renderer with sampled-token and content attribution, but {reason}. "
+            "Implement and register the renderer in the renderers package, or explicitly select an existing "
+            "typed renderer only when its template is verified to match."
+        )
+
     @model_validator(mode="after")
     def validate_pack_function(self):
         if self.model.cp > 1:
@@ -319,9 +343,9 @@ def dont_do_massive_traces(self):
 
     @model_validator(mode="after")
     def validate_renderer_vs_vlm(self):
-        if self.renderer is not None and self.model.vlm is not None:
+        if self.model.vlm is not None:
             raise ValueError(
-                "renderer is not supported for VLMs in SFT. The renderer tokenizes "
+                "renderer-only SFT does not support VLMs yet. The renderer tokenizes "
                 "text-only message dicts client-side and cannot handle image inputs."
             )
         return self