2773f98357
run_server is a library/embed entry point (colab.start calls it directly), so the unconditional os.execv for the torch CUDA LD_LIBRARY_PATH fix would replace the live Colab/Jupyter kernel and drop in-memory state. Gate the re-exec behind a new allow_reexec flag defaulting to False; the run.py __main__ path already re-execs before calling run_server, and the unsloth_cli studio/ui entrypoints opt in with allow_reexec=True so the CLI keeps the CUDA LD fix. |
||
|---|---|---|
| .. | ||
| assets | ||
| auth | ||
| core | ||
| loggers | ||
| models | ||
| plugins | ||
| requirements | ||
| routes | ||
| state | ||
| tests | ||
| utils | ||
| __init__.py | ||
| colab.py | ||
| main.py | ||
| run.py | ||