// 37 napraw · Windows, Linux, Mac i iPhone
Rozwiązywanie błędów lokalnych modeli LLM
Wklej błąd, który widzisz. Większość błędów lokalnej AI sprowadza się do jednego — model jest większy niż Twój sprzęt — a każdy przewodnik podpowiada dokładnie, jak to naprawić.
Nie wiesz, od czego zacząć? Uruchom to i sprawdź kolumnę Processor:
ollama psNie startuje albo nie widzi GPU27 napraw
- CUDA out of memorytorch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 8.00 GiB total capacity; 7.21 GiB already allocated)
- Ollama: „model requires more system memory than is available”Error: model requires more system memory (9.2 GiB) than is available (7.6 GiB)
- LM Studio: „Failed to load model” (za mało pamięci)Failed to load model. Error loading model: unable to allocate backend buffer
- „No CUDA-capable device is detected”RuntimeError: No CUDA-capable device is detected
- Apple Silicon: model nie chce korzystać z GPU / Metal nieaktywnyggml_metal_init: failed to allocate buffer
- Brak pamięci przy długim kontekście (winny jest cache KV, nie wagi)llama_new_context_with_model: failed to allocate KV cache
- GPU AMD niewykrywane / ROCm nieobsługiwany (Ollama i llama.cpp)Error: no compatible GPUs were discovered
- Brak miejsca na dysku / nieudane pobieranie GGUF przy dużych modelachError: write /root/.ollama/models/blobs/sha256-...: no space left on device
- Ollama „connection refused”curl: (7) Failed to connect to localhost port 11434: Connection refused
- CUDA działa na Windowsie, ale nie w WSL2NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
- „ollama is not recognized as an internal or external command”'ollama' is not recognized as an internal or external command, operable program or batch file.
- Przenieś modele poza dysk C:Error: write /Users/you/.ollama/models/blobs/sha256-...: There is not enough space on the disk.
- Nie mogę połączyć się z Ollamą z innego urządzeniacurl: (7) Failed to connect to 192.168.1.42 port 11434 after 2004 ms: Could not connect to server
- SmartScreen i antywirus blokują instalację lokalnej AIWindows protected your PC
- nvidia-smi przestało działać po aktualizacji jądraNVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
- „Failed to initialize NVML: Driver/library version mismatch”Failed to initialize NVML: Driver/library version mismatch
- Secure Boot blokuje moduł sterownika GPUmodprobe: ERROR: could not insert 'nvidia': Key was rejected by service
- Twój kontener Dockera nie widzi GPUdocker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]].
- ROCm nie obsługuje twojej karty AMDError: no compatible GPUs were discovered
- „Permission denied” na /dev/kfdUnable to open /dev/kfd read-write: Permission denied
- Ollama na serwerze linuksowym bez monitoraError: write /usr/share/ollama/.ollama/models/blobs/sha256-...: no space left on device
- Podnoszenie limitu VRAM Metala na Apple Siliconggml_metal_init: recommendedMaxWorkingSetSize = 98304.00 MB
- „App is damaged and can't be opened”"LM Studio" is damaged and can't be opened. You should move it to the Bin.
- Lokalne modele LLM na Macu z IntelemLM Studio system requirements: Apple Silicon Mac (M1/M2/M3/M4), macOS 14.0 or newer.
- MLX nie chce się zainstalować albo liczy na CPUERROR: Could not find a version that satisfies the requirement mlx (from versions: none)
- Model wywala aplikację na iPhonie albo zrzuca ją do mniejszegoThe app disappears to the home screen with no dialog, no crash report, and no error.
- Pobrania modeli na iPhonie, które zawodzą, zapychają pamięć albo się zatrzymująDownload interrupted. Please try again.
Działa, ale źle7 napraw
- Model się ładuje, ale działa boleśnie wolnollm_load_tensors: offloaded 0/33 layers to GPU
- Ollama liczy na CPU na Windowsie mimo karty NVIDIANAME ID SIZE PROCESSOR UNTIL
- Współdzielona pamięć GPU na Windowsie każe modelowi pełznąć, zamiast go zatrzymaćTask Manager → Performance → GPU
- Laptop liczy model na karcie zintegrowanej, a nie na RTX-ie+-----------------------------------------+
- Twoje zmienne środowiskowe Ollamy są na Linuksie ignorowane$ export OLLAMA_MODELS=/mnt/data/models
- Twój Mac swapuje na dysk, zamiast uruchomić modelActivity Monitor → Memory
- Twój MacBook jest szybki przez dwie minuty, potem o połowę wolniejszyminute 1: 42.6 tok/s
Działa, ale wynik jest zły3 napraw
- Którą kwantyzację GGUF pobrać? (Q4, Q5 czy Q8)Which GGUF quant should I download?
- Lokalny LLM powtarza się albo utknął w pętliThe same sentence again. The same sentence again. The same sentence again. The same sentence ag
- Lokalny LLM wypisuje bełkot? Zły szablon czatu, nie brak VRAM-u<|im_start|>assistant<|im_end|> <|im_start|>user I I I the the of of<|endoftext|>assistantassistant