af874bf1e5
On AMD/ROCm a clean launch stalls at "Requested to load LTXAV" while system RAM fills and spills to swap. This is host-side pinned-memory exhaustion, not VRAM pressure. The static pin_memory path ignores ensure_pin_registerable()'s result and does not check ensure_pin_budget() before cudaHostRegister. It can therefore keep registering offloaded weights after current host-RAM pressure says to stop. Honor both existing helpers before registering. If either budget cannot be met, leave the tensor in pageable RAM. Add focused coverage for RAM-budget rejection, registration-budget rejection, successful accounting, and the --high-ram contract. |
||
|---|---|---|
| .. | ||
| folder_path_test.py | ||
| model_detection_test.py | ||
| model_management_test.py | ||
| seedvr_vae_forward_test.py | ||
| test_seedvr2_dtype.py | ||
| test_seedvr2_internals.py | ||
| test_seedvr2_model.py | ||
| test_seedvr2_vae_decode.py | ||
| test_seedvr2_vae_tiled.py | ||