aeba18dc6d
Round 35 P1: _raise_if_helper_advisor_busy published a new public pending marker without first checking public_load_pending(). Two public workloads (e.g. training + diffusion) could both pass their idle helper-busy snapshot concurrently, then both run through destructive owner teardown before either flipped its own visibility flag (is_training_active, current_checkpoint, loading_model_identifier, diffusion is_loading). Add the missing self-check under _HELPER_ADVISOR_START_LOCK so the second public workload sees the first's pending marker and gets a 503 retry instead of racing for VRAM. Helper / advisor already checked public_load_pending() on its side via _gpu_workload_busy_for_helper; this closes the symmetric public -> public window. 86 backend tests pass + smoke test confirms second public load is refused with 503 while first is pending, and the next public load is permitted once the first clears. |
||
|---|---|---|
| .. | ||
| backend | ||
| frontend | ||
| src-tauri | ||
| __init__.py | ||
| install_llama_prebuilt.py | ||
| install_python_stack.py | ||
| LICENSE.AGPL-3.0 | ||
| package-lock.json | ||
| package.json | ||
| setup.bat | ||
| setup.ps1 | ||
| setup.sh | ||
| Unsloth_Studio_Colab.ipynb | ||