# -*- coding: utf-8 -*- """Pemeriksa kesiapan GPU / CUDA untuk aplikasi Vision AI (Windows/Linux). Jalankan: check_cuda.bat (di folder instalasi - memakai Python bawaan aplikasi, jadi torch yang diuji = torch yang dipakai aplikasi) atau: python check_cuda.py (Python apa pun >= 3.8; tanpa torch = cek driver saja) Yang diperiksa: 1. Driver NVIDIA & versi CUDA yang didukung driver (nvidia-smi) 2. Build PyTorch aplikasi (CUDA berapa), torch.cuda.is_available(), uji hitung di GPU 3. GPU: nama, VRAM, compute capability 4. Kesimpulan LULUS/GAGAL + langkah perbaikan + tabel kebutuhan minimum Aturan kecocokan (CUDA "minor version compatibility"): aplikasi yang dibangun dengan CUDA 12.x berjalan pada driver apa pun yang mendukung CUDA >= 12.0. Driver yang mendukung versi CUDA sama/lebih baru dari build aplikasi = dukungan penuh (disarankan). Bukti paling menentukan adalah torch.cuda.is_available() + uji hitung, bukan angka driver. Exit code: 0 = GPU siap, 1 = perlu perbaikan / jalan di CPU, 2 = tidak ada GPU NVIDIA. """ import os import platform import re import shutil import subprocess import sys APP_TORCH_CUDA = "12.6" # build CUDA PyTorch yang dibundel aplikasi (dibaca dari torch bila ada) MIN_CUDA_DRIVER = "12.0" # minimum absolut (minor version compatibility) MIN_DRIVER = { # driver minimum utk CUDA tsb: (Windows, Linux) - tabel resmi NVIDIA "12.0": ("527.41", "525.60"), "12.6": ("560.76", "560.28"), } MIN_CC = (5, 0) # compute capability minimum (Maxwell) utk wheel PyTorch cu12x MIN_VRAM_GB = 4 # inferensi; training model small/medium disarankan >= 6 GB IS_WIN = platform.system() == "Windows" def vtuple(v): return tuple(int(x) for x in re.findall(r"\d+", str(v))[:3]) def run(cmd, timeout=15): try: return subprocess.run(cmd, capture_output=True, text=True, timeout=timeout).stdout except Exception: return "" def section(title): print("\n" + "=" * 66) print(title) print("=" * 66) def main(): print("Pemeriksa CUDA / GPU - " + platform.platform()) print("Python " + sys.version.split()[0] + " (" + sys.executable + ")") problems, notes = [], [] # ---------------------------------------------------------------- 1) driver section("1) Driver NVIDIA") smi = shutil.which("nvidia-smi") or (r"C:\Windows\System32\nvidia-smi.exe" if IS_WIN else "") driver = cuda_drv = "" gpus = [] if smi and os.path.exists(smi): head = run([smi]) m = re.search(r"Driver Version:\s*([\d.]+)", head) driver = m.group(1) if m else "" m = re.search(r"CUDA Version:\s*([\d.]+)", head) cuda_drv = m.group(1) if m else "" q = run([smi, "--query-gpu=name,memory.total,compute_cap", "--format=csv,noheader"]) for line in q.splitlines(): parts = [p.strip() for p in line.split(",")] if len(parts) >= 2: gpus.append({"name": parts[0], "vram": parts[1], "cc": parts[2] if len(parts) > 2 else ""}) if not driver: q2 = run([smi, "--query-gpu=driver_version", "--format=csv,noheader"]) driver = q2.strip().splitlines()[0] if q2.strip() else "" if not driver: print(" nvidia-smi tidak ditemukan / tidak ada GPU NVIDIA yang dikenali driver.") print(" -> Pasang driver NVIDIA terbaru dari https://www.nvidia.com/drivers lalu restart.") else: print(f" Driver : {driver}") print(f" CUDA (driver) : {cuda_drv or '?'} <- versi CUDA tertinggi yang didukung driver ini") for g in gpus: print(f" GPU : {g['name']} | VRAM {g['vram']} | compute capability {g['cc'] or '?'}") cv = os.environ.get("CUDA_VISIBLE_DEVICES") if cv is not None: print(f" CUDA_VISIBLE_DEVICES = {cv!r}") if cv.strip() in ("", "-1"): problems.append("Variabel lingkungan CUDA_VISIBLE_DEVICES menyembunyikan GPU - hapus variabel itu") # ---------------------------------------------------------------- 2) torch section("2) PyTorch aplikasi") torch_cuda = APP_TORCH_CUDA torch_ok = None # None = torch tidak ada di Python ini torch_err = "" try: import torch # noqa torch_cuda = torch.version.cuda or APP_TORCH_CUDA print(f" torch : {torch.__version__} (build CUDA {torch_cuda})") try: torch_ok = bool(torch.cuda.is_available()) except Exception as e: torch_ok, torch_err = False, str(e) print(f" cuda available : {torch_ok}") if torch_ok: for i in range(torch.cuda.device_count()): p = torch.cuda.get_device_properties(i) print(f" device {i} : {p.name} | {p.total_memory / 1e9:.1f} GB | sm_{p.major}{p.minor}") if (p.major, p.minor) < MIN_CC: problems.append(f"GPU {p.name} compute capability {p.major}.{p.minor} < {MIN_CC[0]}.{MIN_CC[1]}") try: x = torch.ones(64, 64, device="cuda") @ torch.ones(64, 64, device="cuda") torch.cuda.synchronize() print(f" uji hitung GPU : OK (matmul 64x64 = {float(x[0, 0]):.0f})") except Exception as e: torch_ok, torch_err = False, str(e) print(f" uji hitung GPU : GAGAL: {str(e)[:160]}") else: # gali alasannya: driver terlalu lama / DLL CUDA tidak termuat / tidak ada device try: torch.cuda.init() except Exception as e: torch_err = torch_err or str(e) if torch_err: print(f" alasan torch : {torch_err[:220]}") except ImportError: print(f" torch tidak ada di Python ini -> memakai asumsi build aplikasi: CUDA {APP_TORCH_CUDA}") print(" (jalankan lewat check_cuda.bat di folder instalasi untuk menguji torch yang sebenarnya)") # ---------------------------------------------------------------- 3) kecocokan section("3) Kecocokan driver vs build CUDA aplikasi") need_min = MIN_DRIVER["12.0"][0 if IS_WIN else 1] need_full = MIN_DRIVER["12.6"][0 if IS_WIN else 1] if driver: if cuda_drv and vtuple(cuda_drv) < vtuple(MIN_CUDA_DRIVER): print(f" GAGAL: driver hanya mendukung CUDA {cuda_drv}; aplikasi (CUDA {torch_cuda}) butuh driver dengan CUDA >= {MIN_CUDA_DRIVER}.") print(f" Minimal driver {need_min}, disarankan >= {need_full}. Terpasang: {driver}.") problems.append(f"Driver {driver} (CUDA {cuda_drv}) terlalu lama: butuh CUDA >= {MIN_CUDA_DRIVER} (driver >= {need_min})") elif cuda_drv and vtuple(cuda_drv) < vtuple(torch_cuda): print(f" OK (kompatibilitas minor): driver mendukung CUDA {cuda_drv}, aplikasi dibangun dengan CUDA {torch_cuda}.") print(f" Berjalan lewat CUDA minor-version compatibility. Untuk dukungan penuh update driver ke >= {need_full}.") notes.append(f"Driver {driver} (CUDA {cuda_drv}) masih di bawah build aplikasi CUDA {torch_cuda}; disarankan update ke >= {need_full}") else: print(f" OK: driver {driver} mendukung CUDA {cuda_drv or '?'} >= build aplikasi {torch_cuda}.") for g in gpus: m = re.search(r"(\d+)", g["vram"]) if m and int(m.group(1)) < MIN_VRAM_GB * 1024: notes.append(f"VRAM {g['vram']} < {MIN_VRAM_GB} GB: inferensi mungkin jalan, training model besar akan gagal") if g["cc"]: try: if vtuple(g["cc"]) < MIN_CC: problems.append(f"Compute capability {g['cc']} < {MIN_CC[0]}.{MIN_CC[1]} (GPU terlalu lama untuk PyTorch cu12)") except Exception: pass if driver and torch_ok is False and not any("terlalu lama" in p for p in problems): hint = ("torch.cuda.is_available() = False padahal driver cukup. Penyebab umum: (1) aplikasi/servis " "dimulai sebelum driver siap -> restart PC lalu jalankan lagi; (2) DLL CUDA aplikasi gagal dimuat " "-> pasang 'Microsoft Visual C++ Redistributable 2015-2022 x64'; (3) GPU dinonaktifkan di Device Manager; " "(4) driver rusak -> pasang ulang driver dengan opsi Clean Install") if torch_err: hint += f" | pesan torch: {torch_err[:160]}" problems.append(hint) # ---------------------------------------------------------------- 4) kesimpulan section("4) Kesimpulan") if not driver: code = 2 print(" TIDAK ADA GPU NVIDIA -> aplikasi jalan di CPU (lebih lambat, tetap berfungsi).") elif problems: code = 1 print(" GPU BELUM BISA DIPAKAI aplikasi. Masalah:") for p in problems: print(" - " + p) print(" Langkah umum: update driver NVIDIA (Game Ready / Studio) ke versi terbaru dari") print(" https://www.nvidia.com/drivers , restart PC, jalankan pemeriksaan ini lagi.") elif torch_ok: code = 0 print(" LULUS: GPU siap dan TERBUKTI dipakai PyTorch aplikasi (inferensi & training di CUDA).") else: code = 0 print(" LULUS (driver): driver & GPU memenuhi syarat. Uji torch belum dijalankan - jalankan") print(" check_cuda.bat dari folder instalasi untuk pembuktian penuh.") for n in notes: print(" Catatan: " + n) section("Kebutuhan minimum aplikasi (GPU opsional; tanpa GPU = CPU)") print(f" OS : Windows 10/11 64-bit (build 19045+) atau Ubuntu 20.04+/JetPack 6") print(f" GPU : NVIDIA compute capability >= {MIN_CC[0]}.{MIN_CC[1]} (GTX 900 series / Maxwell ke atas)") print(f" VRAM : >= {MIN_VRAM_GB} GB inferensi, >= 6 GB training (8 GB+ disarankan)") print(f" Driver NVIDIA : minimum {MIN_DRIVER['12.0'][0]} Win / {MIN_DRIVER['12.0'][1]} Linux (CUDA >= {MIN_CUDA_DRIVER});") print(f" disarankan >= {MIN_DRIVER['12.6'][0]} Win / {MIN_DRIVER['12.6'][1]} Linux (CUDA {torch_cuda}, dukungan penuh)") print(f" CUDA toolkit : TIDAK perlu dipasang terpisah - runtime CUDA {torch_cuda} sudah dibundel aplikasi") print(f" RAM : >= 16 GB (8 GB minimum, hanya inferensi)") print(f" Disk : >= 12 GB bebas untuk instalasi + ruang data project") print(f" CPU tanpa GPU : tetap jalan; inferensi ~5-20x lebih lambat, training tidak disarankan") return code if __name__ == "__main__": try: sys.exit(main()) except KeyboardInterrupt: sys.exit(1)