技能 2026-09-01 12:03 UTC
lecture-to-notesv0.7.1
版本【0.7.1】由提交人 drpwchen 提交,主要更新类型为:常规更新与维护。提交说明:v0.7.1: a GPU that is too small is not a GPU that is busy。对项目做了常规维护与小幅调整。
Commit提交信息原文
v0.7.1: a GPU that is too small is not a GPU that is busy
gpu_check.py compared free VRAM against --min-free-mb and told a 2 GB card
"blocked — wait for other tasks"; nothing was ever going to free up. Check
total VRAM first: a card that can never pass gets GPU_TOO_SMALL, exit 0, and
the CPU path runs (--require-gpu keeps the hard block for batch hosts).
transcribe_video.py gains --device {auto,cuda,cpu}; load-time CUDA OOM is an
explicit fallback case; LECTURE_ASR_DEVICE=cpu stops the collapse-retry child
from OOM-ing again. README says a 2-4 GB card counts as no GPU and which
flags to use.
Also carries two upstream changes in the same sync: gpu_lease_ctx (no-op off
the author's machine) and build_L1 deck-text lines.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KzbN7Jt8pSVvJ1y3xi2Xcp