Pull requests / #167

#167 generate: zero session state for native packs before the prompt path

closed · @samuelishida · 0 评论 · 在 GitHub 查看

NVIDIA / CUDAWindows

描述

A native (IQ) pack breaks out of the canonical `put_input` loop before position 0 and runs verify windows instead, so `generate` mode never reached the `session_zero` call that the canonical path uses. On a reused `cudaMalloc` allocation that left GDN, QSA KV/indexer, PLE history, and `R` uninitialized, producing NaN/overflow and degenerate identical-token output.

Zero the session state explicitly when `native_pack` is set.

Also removes an unused `argmax()` helper that emitted a compiler warning, and ignores the local `.pi/` harness directory.

站内延伸阅读

链到安装、模型与版本说明,便于 SEO/GEO,非官方 issue 正文。