From c7e5e0ae381c94c6320707ff2b87d20b26945984 Mon Sep 17 00:00:00 2001 From: hallerite Date: Fri, 29 May 2026 00:33:33 +0530 Subject: [PATCH] chore(deps): bump renderers (offset-tokenizer config-build fallback for Laguna) Bumps the deps/renderers submodule 89ab3f0 (v0.1.8.dev35) -> 35c2407 (v0.1.8.dev37), pulling in PrimeIntellect-ai/renderers#75: apply the config-build fallback to ``_get_offset_tokenizer`` too. #72 only wired the fallback into ``load_tokenizer``; ``_get_offset_tokenizer`` still called ``AutoTokenizer.from_pretrained`` directly to avoid the fastokens shim and so kept hitting HF's RoPE validator on Laguna-family models. Every rollout through a hand-coded renderer (LagunaXS2, Qwen35, ...) crashed with the same KeyError #72 was supposed to fix. Verified end-to-end against poolside/Laguna-XS.2 with reverse-text: two RL steps complete cleanly (reward 0.36 / 0.31, ~100 tok/sample); the run then hits an unrelated single-GPU FSDP OOM on the 256-expert trainer, which is a config-sizing issue, not the renderers bug. Also pulls in renderers#74 (``message_tool_names`` field for per-message tool attribution). Co-Authored-By: Claude Opus 4.7 (1M context) --- deps/renderers | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/deps/renderers b/deps/renderers index 89ab3f0bed..35c2407e92 160000 --- a/deps/renderers +++ b/deps/renderers @@ -1 +1 @@ -Subproject commit 89ab3f0bedd1ec24cd1aec1b6e1bed31bc87424e +Subproject commit 35c2407e9219e22d4da7e68fdb206d17b771120a