-
Reload on the CPU and retry once when the GPU runs out of memory (#1472)
frostbyte_neo 发布于
2026-09-04 03:49:35 +00:00 Reload on the CPU and retry once when the GPU runs out of memory
sona v0.6.3 reports a failed GPU allocation as gpu_out_of_memory, on the response, on the
streaming error event, and on model load; when the allocation aborts the process instead the
only evidence is the abort line in its stderr. The transcribe queue reads both, reloads the
model with noGpu, puts the file back for one more try, and stays on the CPU for the rest of
the run. The verdict is kept per model so the next launch starts on the CPU too; batch and
dictation read it at load. One analytics event says how often, and whether the code or the
process death signalled it.Closes #1471
Claude-Session: https://claude.ai/code/session_01TgmkaiiuPu65pSqXptMAna
Co-authored-by: mikemikimike 13286568797@163.com下载附件