Skip to content

Sync-Strata 0.1.3

Choose a tag to compare

@hiruhiru2005 hiruhiru2005 released this 20 Nov 14:10
· 1 commit to main since this release

Sync-Strata 0.1.3

日本語

  • 「拒否を減らす(無検閲・試験的)」を Strata の設定に足しました(既定は切)。本家 Strata と同じ制御ベクトルで、モデルが断ることを減らします。安全のための振る舞いも弱まるので、書かせた内容は使う人の責任です。
  • 歯車の設定の一番上に「Strata の設定」を足しました。読み込み中や生成中、スマホからでも開けます(起動の設定の名前も「Strata の設定」にそろえました)。
  • 窓の部品(閉じる小窓・起動中の画面・状態の窓・トレイのメニュー)を、英語と中国語にも対応させました。会話の画面で選んだ言語と同じ言語で出ます(中国語の字体が無い Linux 向けに、字体も同梱しました)。
  • 英語の画面で、GUI が起動せずに終わる不具合を直しました(sync status --json などの出力に翻訳の印が付き、JSON として読めなくなっていたため)。
  • RAM から GPU に渡す専門家の割合を、その PC で前回落ち着いた値から始めるようにしました。起動の直後や短い頼みでも、合った割合で動きます。合わせ方も、2 回の実測から一気に寄せる形にして速くしました。
  • Linux の Radeon で、短い文の decode が遅くなっていた件に対処しました(割合の初期値を 0.1.1 と同じ 0.7 に戻しました。Windows の Radeon は 0.4 のままです)。
  • Radeon で、空けておく VRAM が少なめに決まっていた不具合を直しました(起動の設定に VRAM の量を渡していなかったため)。
  • RAM の置き場の取り方を起動時に測る仕組み(試験的。既定は止め)の、Windows での測り方を直しました。
  • 既知の問題: Windows の RX 6900 XT(RDNA2)で、起動の直後に止まる報告があり、調べています(Issue #2)。
  • ご報告いただいた皆さま、ありがとうございました。
  • エンジン: Strata のフォーク ql/rel-0113(Windows 762a3b05・Linux 829f5c13)。
  • 0.1.2 からの更新: アプリの「更新」か、update.cmd(Windows)・update.sh(Linux)。

English

  • Added "Fewer refusals (uncensored, experimental)" to Strata settings (off by default). The same control vector as upstream Strata makes the model decline less often. Its safety behavior is weakened too, so you are responsible for what you have it write.
  • Added "Strata settings" at the top of the gear settings, so it opens while the model loads or writes, and from a phone (the start settings are now called "Strata settings").
  • The app window parts (the close dialog, the start-up screen, the status window and the tray menu) now also come in English and Chinese, in the language chosen on the chat screen (a font is bundled for Linux without a Chinese font).
  • Fixed the GUI closing right after start with the English UI (a translation mark was added to the output of sync status --json and similar commands, so it was no longer valid JSON).
  • The share of experts sent from RAM to the GPU now starts from the value this PC settled on last time, so the share already fits right after start and on short requests. It also settles faster, moving straight toward the balance point from two measurements.
  • Addressed slower decode of short prompts on Radeon on Linux (the starting share is back to 0.7, as in 0.1.1; Radeon on Windows stays at 0.4).
  • Fixed Radeon keeping too little VRAM free (the amount of VRAM was not passed to the start settings).
  • Fixed how the experimental start-up measurement of the RAM copy allocation (off by default) times copies on Windows.
  • Known issue: an RX 6900 XT (RDNA2) on Windows was reported to stop right after start; this is being investigated (Issue #2).
  • Thank you to everyone who reported problems.
  • Engine: the Strata fork ql/rel-0113 (Windows 762a3b05, Linux 829f5c13).
  • Updating from 0.1.2: "Update" in the app, or update.cmd (Windows) / update.sh (Linux).