// 39 个修复 · Windows、Linux、Mac 与 iPhone
本地 LLM 错误排查
粘贴你看到的错误信息。大多数本地 AI 错误都归结为一件事——模型比你的硬件所能承载的更大——每篇指南都会告诉你确切的修复方法。
不知从何开始？运行下面的命令并查看 Processor 一列：
ollama ps无法启动，或识别不到 GPU29 个修复
- 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" (insufficient memory)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 won't use the GPU / Metal not engagedggml_metal_init: failed to allocate buffer
- Out of memory at long context (the KV cache, not the weights)llama_new_context_with_model: failed to allocate KV cache
- AMD GPU not detected / ROCm unsupported (Ollama & llama.cpp)Error: no compatible GPUs were discovered
- Out of disk space / failed GGUF download for large modelsError: 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
- Ollama /v1/systemone errors: 400, 404, 413 explained{"error": "request body must not exceed 64 KiB without images"}
- Ollama /v1/systemone image errors: 400, 413, HEIC, rotationHTTP 400 or 413 from POST /v1/systemone as soon as the request has an "images" array, or a 200 whose answers ignore the picture
- CUDA works on Windows but not inside 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.
- Move your models off C:Error: write /Users/you/.ollama/models/blobs/sha256-...: There is not enough space on the disk.
- Can't reach Ollama from another devicecurl: (7) Failed to connect to 192.168.1.42 port 11434 after 2004 ms: Could not connect to server
- SmartScreen and antivirus are blocking your local AI installWindows protected your PC
- nvidia-smi stopped working after a kernel updateNVIDIA-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 is blocking your GPU driver modulemodprobe: ERROR: could not insert 'nvidia': Key was rejected by service
- Your Docker container can't see the GPUdocker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]].
- ROCm doesn't support your AMD cardError: no compatible GPUs were discovered
- "Permission denied" on /dev/kfdUnable to open /dev/kfd read-write: Permission denied
- Ollama on a headless Linux serverError: write /usr/share/ollama/.ollama/models/blobs/sha256-...: no space left on device
- Raising the Metal VRAM ceiling on 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.
- Running local LLMs on an Intel MacLM Studio system requirements: Apple Silicon Mac (M1/M2/M3/M4), macOS 14.0 or newer.
- MLX won't install, or runs on the CPUERROR: Could not find a version that satisfies the requirement mlx (from versions: none)
- The model crashes your iPhone app, or drops to a smaller oneThe app disappears to the home screen with no dialog, no crash report, and no error.
- iPhone model downloads that fail, fill your storage, or stopDownload interrupted. Please try again.
能跑，但很慢7 个修复
- Model loads but runs painfully slowllm_load_tensors: offloaded 0/33 layers to GPU
- Ollama runs on the CPU on Windows despite an NVIDIA GPUNAME ID SIZE PROCESSOR UNTIL
- Windows shared GPU memory makes your model crawl, not crashTask Manager → Performance → GPU
- Your laptop runs the model on the integrated GPU, not the RTX+-----------------------------------------+
- Your Ollama environment variables are ignored on Linux$ export OLLAMA_MODELS=/mnt/data/models
- Your Mac is swapping to disk instead of running the modelActivity Monitor → Memory
- Your MacBook is fast for two minutes, then half as fastminute 1: 42.6 tok/s
能跑，但输出不对3 个修复
- Which GGUF quant should I download? (Q4 vs Q5 vs Q8)Which GGUF quant should I download?
- Local LLM repeating itself or stuck in a loopThe same sentence again. The same sentence again. The same sentence again. The same sentence ag
- Local LLM outputs gibberish? Wrong chat template<|im_start|>assistant<|im_end|> <|im_start|>user I I I the the of of<|endoftext|>assistantassistant