[samsja]

Summary

Adds a deepseek-v4.1 renderer for DeepSeek-V4.1-Flash, moved here from prime-rl's example custom renderer (PrimeIntellect-ai/prime-rl#4057) so configs can select it by name (including on hosted training, where a repo-relative import_path does not resolve).

For plain chats, V4.1's reference encoder (encoding/encoding.py in the checkpoint) is V4's with three changes, applied on top of DeepSeekV4Renderer:

• a <｜System｜> token leads the conversation whenever it opens with a reasoning-effort prefix or a system message;
• the reasoning effort is a numeric budget, Reasoning Effort: {budget} (range 1-100, ...), thinking mode only, default high (75); low / high / max map to 50 / 75 / 100;
• thinking mode is the default.

A leading inline <think>...</think> in an assistant's content moves into reasoning_content, as other renderers do.

Not supported yet: tool calls (V4.1 renamed the DSML tags) and mid-conversation system messages; both raise NotImplementedError. bridge_to_next_turn returns None (full re-render) since V4's extension path would emit V4's tags.

Selected by name only (name = "deepseek-v4.1"): it is not in MODEL_RENDERER_MAP, because auto-mapped renderers must pass the shared parity suites, which need tool support.

Test plan

• tests/test_deepseek_v41.py: token-for-token parity with the checkpoint's shipped encoding/encoding.py (downloaded from the Hub) over 4 conversations × thinking on/off × effort low/high/max/30; config discriminator and effort range; inline-reasoning split; tool and mid-conversation system rejection.
• pytest tests/test_deepseek_v41.py tests/test_deepseek_v4.py tests/test_renderer_config.py tests/test_reasoning_boundaries.py: 989 passed.

🤖 Generated with Claude Code

Note

Add DeepSeekV41Renderer for DeepSeek-V4.1-Flash

• Adds a new renderer subclassing DeepSeekV4Renderer in deepseek_v41.py, with thinking enabled by default and a reasoning_effort field accepting low, high, max, or integers 1–100 (named levels map to budgets 50, 75, 100).
• Registers deepseek-v4.1 in the renderer name registry, the lazy renderer map, and the RendererConfig discriminated union in configs.py and base.py.
• Rendering rejects non-empty tool lists and system messages after the first message (NotImplementedError), and splits a leading inline think block in assistant messages into separate reasoning content.
• Tests in test_deepseek_v41.py check output parity with the checkpoint encoder across 32 combinations of conversation, thinking mode, and reasoning effort, plus config validation and rejection cases.
• Risk: bridge_to_next_turn on DeepSeekV41Renderer always returns None, so incremental turn bridging is unavailable for this model.

^{Macroscope summarized c949a28. (Automatic summaries will resume when PR exits draft mode or review begins).}