Tags: PrimeIntellect-ai/renderers
Tags
renderers v0.1.11: feature release — two new renderers, optional tran… …sformers feat: add DeepSeek V4 Flash 0731 and Qwen3.8 Flash Next renderers. feat: make transformers an optional dependency; structural Tokenizer / ChatTemplate protocols let text-only consumers bring their own tokenizer, including offsetless BYO tokenizers. change: renderer pool APIs are removed. test: unify renderer template parity coverage under one declarative model catalog and shared scenario corpus.
renderers v0.1.10: feature release — six new renderers feat: add Laguna M.1, Laguna-S-2.1, Nemotron 3.5 (Lightning), Gemma 4, Inkling / Inkling-Small, and Qwen3.8 27B renderers. change: chat-template kwargs are now validated against explicit per-renderer allowlists; unknown kwargs raise TypeError. ci: publish workflows bump pypa/gh-action-pypi-publish to v1.14.2 so distributions with packaging metadata 2.5 upload.
renderers v0.1.9: patch release — client and parsing fixes fix(client): update vLLM token serving import; reject malformed sampled-token evidence. fix(parsing): validate GLM tool names against the provided tools list. fix(qwen): keep empty think wrapper on historical turns when thinking is disabled. fix(renderers): bridge merge no longer mutates prior multi_modal_data. Deps: setuptools 83; dev-deps torch 2.13, pillow 12.3.
renderers v0.1.8: first stable since 0.1.7 — new model renderers + SF… …T/analytics surface New renderers: Tencent Hunyuan Hy3, Laguna-XS-2.1, DeepSeek-V3/R1 (split), Nemotron-3 (Nano/Super/Ultra), Llama-3, PrimeIntellect Qwen3 templates. SFT/RL: multimodal payload through build_training_sample, opt-in final stop token, thinking_retention (replaces preserve_* bools), kept_tokens sampling mask parsing, per-token is_content mask, per-message/per-role analytics, message_tool_names. Typed renderer configs. fastokens removed.
PreviousNext