Files
stemengine_old/backend/routers_root_backup/system.py
T
2026-03-03 14:52:19 +09:00

47 lines
1.7 KiB
Python
Executable File

"""System information endpoints (GPU hardware)."""
from fastapi import APIRouter
router = APIRouter(prefix="/api/system", tags=["system"])
@router.post("/gpu-reset")
def gpu_reset():
"""
Re-enable GPU acceleration for all services without restarting the container.
After a CUDA OOM or 'unknown error', the lyric transcriber marks the GPU
context as failed and falls back to CPU. The separator may also fail with
a CUDA context error. Calling this endpoint clears the CUDA allocation
cache and resets the Whisper device flag, so both services can retry on GPU.
NOTE: if the CUDA context is truly corrupted (persistent 'unknown error'),
a container restart is still required.
"""
from ..services.lyric_transcriber_service import reset_cuda
return reset_cuda()
@router.get("/gpu-info")
def gpu_info():
"""Return GPU name, VRAM figures, and cuda_ok health flag."""
try:
from ..services.lyric_transcriber_service import _cuda_ok as whisper_ok
except Exception:
whisper_ok = True
try:
import torch # type: ignore
if not torch.cuda.is_available():
return {"cuda_available": False, "cuda_ok": whisper_ok, "name": None, "total_mb": 0, "free_mb": 0}
props = torch.cuda.get_device_properties(0)
free_bytes, total_bytes = torch.cuda.mem_get_info(0)
return {
"cuda_available": True,
"cuda_ok": whisper_ok,
"name": props.name,
"total_mb": total_bytes // (1024 * 1024),
"free_mb": free_bytes // (1024 * 1024),
}
except Exception as exc: # pragma: no cover
return {"cuda_available": False, "cuda_ok": False, "name": None, "total_mb": 0, "free_mb": 0, "error": str(exc)}