Repository navigation
Sync-Strata 0.1.2
日本語
- 「Strata 準備中」のまま返答が来ない不具合を直しました。
- Windows で、最初のメッセージの直後にエンジンが止まる(「layer 1 でタイムアウト」)PC がありました。GPU とのやり取りの完了を知らせる方式が原因で、Windows では NVIDIA・Radeon とも別の方式を既定にしました(RTX 3080 では速さは変わりません)。
- エンジンが止まったときは、次のメッセージで自動的に起こし直します(続けて止まるときは、2 回でエラーを表示します)。
- 終了に約 2 分かかることがある不具合を直しました。
- GPU が 2 枚以上ある PC で、選んだ GPU ではなく 1 枚目の GPU で動いていた不具合を直しました。
- Radeon で出力が崩れる不具合を直しました(同梱の llama.cpp の部分で、1 バイトごとの引き算が飽和していたため)。
- Radeon の改善(試験的):
- Windows の Radeon で、起動できない・最初のメッセージで固まる件に対処しました(GPU の仮想メモリーを使わない設定などを既定にしました)。
- Radeon の decode を速くするため、RAM から GPU に渡す割合の初期値を下げました。RAM の置き場の取り方を起動時に測って選ぶ仕組みも入れました(試験的。既定は今までどおりで、
--env STRATA_HIP_ARENA=autoで試せます)。
- 更新の安全性を上げました。配布元に届かないとき、ダウンロードのフォルダーの zip を自動では使いません(ダウンロードした zip を使うときは
update.cmd --source <zip>)。また、GitHub に届かないことがあった証明書の確認を直しました。 - AVX-512 に対応していない CPU の案内を、実測に合わせて「1 割ほど遅くなる」に直しました。
- アプリ本体とモデルの保存先の変え方を docs/STORAGE.ja.md にまとめました。
- 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-1118(Windows 9fa44244・Linux 2d6076ad)。Radeon 用のエンジンだけ、上の直しを加えて組み直しました(Windows d278210e・Linux 7d53462d)。
- 0.1.1 からの更新: アプリの「更新」か、
update.cmd(Windows)・update.sh(Linux)。「新しい版の置き場に届きませんでした」と出るときは、リリースの zip をダウンロードしてupdate.cmd --source <zip>で更新してください。
English
- Fixed replies never coming while the page stays on "Strata getting ready".
- On some Windows PCs the engine stopped right after the first message ("timed out at layer 1"). The way GPU transfers report completion caused it; Windows now uses another way by default on both NVIDIA and Radeon (no speed change on an RTX 3080).
- When the engine stops, the next message starts it again (if it keeps stopping, an error is shown after two starts).
- Fixed quitting sometimes taking about two minutes.
- Fixed PCs with two or more GPUs running on the first GPU instead of the chosen one.
- Fixed broken output on Radeon (in the bundled llama.cpp part, a per-byte subtraction saturated).
- Radeon improvements (experimental):
- On Windows, addressed Radeon failing to start or freezing at the first message (GPU virtual memory is now off by default, among others).
- For faster decode on Radeon, the starting share sent from RAM to the GPU is lower. A way to measure and choose how the RAM copy is allocated at start was added (experimental; off by default, try it with
--env STRATA_HIP_ARENA=auto).
- Safer updates: when the update source cannot be reached, a zip in the Downloads folder is no longer used automatically (to use a downloaded zip:
update.cmd --source <zip>). Also fixed a certificate check that sometimes could not reach GitHub. - The note for CPUs without AVX-512 now says about 10% slower, as measured.
- How to move the app and the models: docs/STORAGE.md.
- Thank you to everyone who reported problems.
- Engine: the Strata fork ql/rel-1118 (Windows 9fa44244, Linux 2d6076ad). Only the Radeon engines were rebuilt, with the fixes above (Windows d278210e, Linux 7d53462d).
- Updating from 0.1.1: "Update" in the app, or
update.cmd(Windows) /update.sh(Linux). If it says the update source could not be reached, download the release zip and runupdate.cmd --source <zip>.