KoboldCpp error messages and what they mean
Search for the message, or paste your whole console log into the analyzer: it runs in your browser and finds the real cause, which is often a few lines above the last error. Messages such as Failed to execute script, Could not load text model and Failed to fetch are wrappers: they only say that something failed earlier. Each row gives the meaning, the likely causes in order, and the fix.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
- gha-w25: Windows Server 2025 test machine · 4 vCPU · 16 GB · no GPU
- gha-w22: Windows Server 2022 test machine · 4 vCPU · 16 GB · no GPU
- gha-u24: Ubuntu 24.04 test machine · 4 vCPU · 16 GB · no GPU
- gha-mac15: macOS 15 test VM (Apple Silicon) · 3 vCPU · 7 GB · paravirtual Metal
On this page
Failed to execute script 'koboldcpp' due to unhandled exception!
Wrapper message. It says that something failed, not why. The real cause is printed above it. Why the last line is often not the real error.
This line only says that KoboldCpp's launcher stopped with an unhandled error. It never says why: the real cause is printed above it.
Likely causes, most likely first
- Read the lines just above for the real error (port in use, missing DLL, out of memory, a crash while loading the model)
- An old CPU without AVX2 running a normal build
- Antivirus removed a file that KoboldCpp unpacks at start
What to do
- Scroll up to the first error line above this message and look that one up
- Start KoboldCpp from a terminal so the window stays open
- On a CPU older than about 2013, try the oldpc file or --noavx2
OSError: Cannot start the server: no socket could bind.[PYI-10566:ERROR] Failed to execute script 'koboldcpp' due to unhandled exception!Warning: Port 45327 already appears to be in use by another program.OSError: Cannot start the server: no socket could bind.
KoboldCpp could not open its port because another program is already using it, most often another KoboldCpp window.
Likely causes, most likely first
- A second KoboldCpp is still running (check the system tray, Task Manager or ps)
- Another program uses port 5001
- Your PC blocks the port
What to do
- Close the other KoboldCpp window, or end its process
- Start this one on another port with --port 5002
- Check what holds the port: netstat -ano (Windows) or ss -ltnp (Linux)
File "koboldcpp.py", line 9178, in RunServerMultiThreaded raise OSError("Cannot start the server: no socket could bind.")OSError: Cannot start the server: no socket could bind.OSError: exception: access violation reading 0x…
The model loader crashed while reading memory. It points to a model the build cannot handle, a memory problem, or a CPU instruction the build assumes.
Likely causes, most likely first
- A model architecture that is newer than your KoboldCpp
- A broken or incomplete model file
- Not enough RAM or VRAM for the model plus context
- A CPU without AVX2 using a normal build
What to do
llama_model_load_from_file_impl: failed to load model
Wrapper message. It says that something failed, not why. The real cause is printed above it. Why the last line is often not the real error.
The model could not be loaded. The reason is printed a few lines above this message.
Likely causes, most likely first
- The model file is damaged or the download stopped early
- The model architecture is newer than this KoboldCpp
- A split model was loaded from the wrong part
What to do
- Look at the lines above for tensor, architecture or bounds errors
- Compare the file size with the one on Hugging Face and download again if it differs
- Update KoboldCpp if the log says unknown model architecture
llama_model_load: error loading model: tensor 'output.weight' data is not within the file bounds, model is corrupted or incompletellama_model_load_from_file_impl: failed to load modelgpttype_load_model: error: failed to load model '/home/<user>/lab/results/run1-2026-11-13-local/lab-lnx/_files/truncated.gguf'Error: Could not load text model: <path>
Wrapper message. It says that something failed, not why. The real cause is printed above it. Why the last line is often not the real error.
KoboldCpp gives up after the loader failed. This is the last line of a failed load, not the cause.
Likely causes, most likely first
- See the error printed above this line
What to do
- Find the first error above this line and look it up
- If the file is missing or not a GGUF, check the path and the file
Could not load text model: /home/<user>/lab/results/run1-2026-11-13-local/lab-lnx/_files/not-a-model.ggufCannot find text model file: <path>
The path you gave for --model does not point to a file.
Likely causes, most likely first
- A typo or a wrong folder in the path
- A path with spaces that was not quoted
- The file was moved or deleted
What to do
- Check the path and put it in quotes if it has spaces
- Use the launcher's Browse button instead of typing the path
Cannot find text model file: /home/<user>/lab/models/does-not-exist.ggufload_model: error: invalid or unsupported model file '<path>'
The file you chose is not a GGUF model that KoboldCpp can read.
Likely causes, most likely first
- A different format such as safetensors, bin or a zip
- An HTML page saved instead of the model (a failed download)
- A wrong file picked in the launcher
What to do
- Use a .gguf file; find one on the model's Hugging Face Files tab
- Check the file size: a few kilobytes means a saved web page, not a model
Loading Text Model: /home/<user>/lab/results/run1-2026-11-13-local/lab-lnx/_files/not-a-model.ggufload_model: error: invalid or unsupported model file '/home/<user>/lab/results/run1-2026-11-13-local/lab-lnx/_files/not-a-model.gguf'Load Text Model OK: Falsellama_model_load: error loading model: tensor '…' data is not within the file bounds, model is corrupted or incomplete
The model file is cut short. The download stopped before the end.
Likely causes, most likely first
- The download was interrupted
- The disk filled up while saving
- A browser saved only part of the file
What to do
- Download the file again and compare its size with the Hugging Face page
- Free disk space before downloading
Loading Text Model: /home/<user>/lab/results/run1-2026-11-13-local/lab-lnx/_files/truncated.ggufllama_model_load: error loading model: tensor 'output.weight' data is not within the file bounds, model is corrupted or incompletellama_model_load_from_file_impl: failed to load modelunknown model architecture: '<arch>'
The model is built on a design this KoboldCpp does not know yet.
Likely causes, most likely first
- Your KoboldCpp is older than the model
- The model needs a newer or rolling build
What to do
missing tensor 'blk.N.…'
The model file lacks a part this build expects.
Likely causes, most likely first
- A model variant that this build does not support yet
- A broken quantisation
What to do
tensor '…' has wrong shape; expected … got …
A part of the file has a different size than the loader expects.
Likely causes, most likely first
- An old-format or modified model
- The wrong kind of model in a slot, for example an image model in the text slot
What to do
mtmd_init_from_file: error: mismatch between text model (n_embd = N) and mmproj (n_embd = M)
The vision projector (mmproj) belongs to a different model than the text model you loaded.
Likely causes, most likely first
- The mmproj was downloaded for another model
- Two different model sizes of one family were mixed
What to do
- Download the mmproj that sits in the same Hugging Face repo as your model
- Or remove the mmproj if you do not need vision
get_dummy_batch: warmup with image size = 512 x 512mtmd_init_from_file: error: mismatch between text model (n_embd = 896) and mmproj (n_embd = 576)hint: you may be using wrong mmprojkey clip.vision.… not found in file
The projector file is in a format this build does not read.
Likely causes, most likely first
- An mmproj that does not match the model
- An old projector with a new build or the other way round
What to do
Unable to detect VRAM, please set layers manually.
KoboldCpp could not read how much video memory you have, so automatic layer choice is not possible. It continues, usually on CPU.
Likely causes, most likely first
- No GPU, or no working GPU driver
- An integrated or Intel GPU the tools cannot query
- A tool such as nvidia-smi is missing
What to do
- Choose a GPU backend in the launcher and set GPU layers yourself
- Check the model buffer lines to see where the model really went
No GPU or CPU backend was selected. Trying to assign one for you automatically...Unable to detect VRAM.Auto Selected Default Backend (flag=0)Unable to determine GPU Memory
Same situation as 'Unable to detect VRAM': the free video memory was not readable.
Likely causes, most likely first
- No working GPU driver
- An unsupported GPU for the memory query
What to do
System: Linux #1 SMP PREEMPT_DYNAMIC @0 x86_64 x86_64Unable to determine GPU MemoryDetected Available RAM: 15260 MBNo GPU backend found, or could not automatically determine GPU layers. Please set it manually.
KoboldCpp could not pick a GPU backend or a layer count and falls back to CPU.
Likely causes, most likely first
- A build without GPU support, or no backend chosen
- Missing drivers
What to do
Unable to detect VRAM.No GPU backend found, or could not automatically determine GPU layers. You may prefer to set layers manually.System: Linux #1 SMP PREEMPT_DYNAMIC @0 x86_64 x86_64Auto Selected Default Backend (flag=0)
No GPU backend was chosen, so everything runs on the CPU.
Likely causes, most likely first
- No GPU flag was given and none was detected
- The GPU flag was ignored because the build has no such backend
What to do
- If you expected the GPU, choose CUDA or Vulkan and set GPU layers
- Check the KV buffer line: CPU means CPU
Unable to detect VRAM.Auto Selected Default Backend (flag=0)ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory
The Vulkan driver refused to give KoboldCpp more video memory.
Likely causes, most likely first
- Too many layers or too much context for the card
- Another program uses the video memory
What to do
ggml_vulkan: Device memory allocation of size N failed.
A single large block of video memory could not be reserved on a Vulkan device.
Likely causes, most likely first
- Not enough free video memory
- A driver limit on the size of one allocation
What to do
cudaMalloc failed: out of memory
The NVIDIA card has no room for what KoboldCpp asked for.
Likely causes, most likely first
- Too many layers, too much context, or a large batch size
- Another program uses VRAM
- With two GPUs, an uneven split
What to do
failed to allocate compute (pp) buffers
The working memory for processing the prompt did not fit.
Likely causes, most likely first
- A large context or a large batch size with too little free memory
What to do
llama_init_from_model: failed to initialize the context: failed to allocate buffer for kv cache
The context memory (KV cache) could not be created. The context you asked for is too big for the free memory.
Likely causes, most likely first
- A very large --contextsize
- Little free RAM or VRAM
- Other programs using memory
What to do
- Lower --contextsize; every doubling doubles the KV cache
- Quantise the KV cache with --quantkv
- Check the number with the VRAM calculator
ggml_backend_buft_alloc_buffer_n_default: failed to allocate CPU buffer of size 6445596672llama_init_from_model: failed to initialize the context: failed to allocate buffer for kv cachegpttype_load_model: error: failed to load model '<models>/qwen2.5-0.5b-instruct-q4_k_m.gguf'ggml_cuda_host_malloc: failed to allocate N MiB of pinned memory: out of memory
KoboldCpp could not lock host memory for faster transfers. It usually continues without it.
Likely causes, most likely first
- A system limit on locked memory
- --usemlock with a large model
What to do
ggml_new_tensor_impl: not enough space in the context's memory pool
An old allocation error: the loader ran out of its working space.
Likely causes, most likely first
- An old KoboldCpp version
- A model that is too large for the build
What to do
CUDA error: an illegal memory access was encountered
The GPU code touched memory it should not. This is a bug or an unsupported combination.
Likely causes, most likely first
- ContextShift together with a quantised KV cache
- A backend bug in this version
What to do
ggml-cuda was compiled without support for the current GPU architecture
Your NVIDIA card is older than the GPUs this build supports.
Likely causes, most likely first
- A GPU from before the Pascal/Turing era with the normal CUDA build
What to do
CUDA kernel mul_mat_vec has no device code compatible with CUDA arch 520
The same cause as the previous message: the build has no code for your GPU generation.
Likely causes, most likely first
- An older NVIDIA GPU such as Maxwell
What to do
CUDA error: no kernel image is available for execution on the device
The installed CUDA build has nothing that runs on your GPU.
Likely causes, most likely first
- A very old GPU such as Kepler
- A driver too old for the CUDA version
What to do
Illegal instruction (core dumped)
The CPU was asked to run an instruction it does not have.
Likely causes, most likely first
- A CPU without AVX2 (before about 2013) with a normal build
- An ARM device with a build for another CPU
What to do
PrefetchVirtualMemory not found in KERNEL32.dll
The program needs a Windows function that older Windows versions lack.
Likely causes, most likely first
- Windows 7
What to do
PyInstallerImportError: Failed to load dynlib/dll '…'
A file KoboldCpp unpacks into a temporary folder is missing or blocked.
Likely causes, most likely first
- Antivirus removed or quarantined a file
- An old temporary folder left over from a crash
What to do
- Delete the _MEI folders in your temp directory and start again
- Add the folder to your antivirus exclusions, or check the quarantine list
Restore a quarantined file only after checking its hash. Always download KoboldCpp from GitHub only; lookalike sites exist. How to tell.
File "PyInstaller/loader/pyimod03_ctypes.py", line 55, in __init__pyimod03_ctypes.PyInstallerImportError: Failed to load dynlib/dll '/tmp/_MEIpicf7w/koboldcpp_cublas.so'. Most likely this dynlib/dll was not found when the application was frozen.[PYI-4225:ERROR] Failed to execute script 'koboldcpp' due to unhandled exception!OSError: [WinError 1114] A dynamic link library (DLL) initialization routine failed
A DLL started but failed to initialise, often because of the GPU driver.
Likely causes, most likely first
- An outdated or broken NVIDIA driver
- The CUDA build on a PC without a matching GPU
What to do
Restore a quarantined file only after checking its hash. Always download KoboldCpp from GitHub only; lookalike sites exist. How to tell.
FileNotFoundError: Could not find module '…koboldcpp_cublas.dll'
KoboldCpp could not load its CUDA library. The most common reason on a PC without an NVIDIA driver is that the CUDA build was told to use CUDA; antivirus removing the file is the other.
Likely causes, most likely first
- No NVIDIA driver on this PC: the CUDA build was started with --usecuda (tested)
- Antivirus quarantined the library
- The unpacking was interrupted
What to do
- No NVIDIA GPU or driver: use the -nocuda file, or remove --usecuda
- With an NVIDIA GPU: install or update the NVIDIA driver
- Otherwise check your antivirus quarantine and restore the file
Restore a quarantined file only after checking its hash. Always download KoboldCpp from GitHub only; lookalike sites exist. How to tell.
File "ctypes\__init__.py", line 373, in __init__FileNotFoundError: Could not find module 'C:\Users\<user>\AppData\Local\Temp\_MEI22122\koboldcpp_cublas.dll' (or one of its dependencies). Try using the full path with constructor syntax.OSError: [WinError 4551]
Windows blocked the file under its Smart App Control policy.
Likely causes, most likely first
- Smart App Control is on and does not trust the file
What to do
Restore a quarantined file only after checking its hash. Always download KoboldCpp from GitHub only; lookalike sites exist. How to tell.
GGML_ASSERT(…) failed
An internal check failed. The text inside GGML_ASSERT names what was wrong.
Likely causes, most likely first
- A model or setting combination this build cannot handle
What to do
Segmentation fault (core dumped)
The program crashed in native code (Linux and macOS).
Likely causes, most likely first
- --usemlock without mmap
- An unsupported GPU or driver
- A version bug
What to do
double free or corruption
A memory bug inside the program.
Likely causes, most likely first
- A version bug
What to do
Aborted (core dumped)
Wrapper message. It says that something failed, not why. The real cause is printed above it. Why the last line is often not the real error.
The process was stopped after a previous error. The cause is the line above.
Likely causes, most likely first
- A GGML_ASSERT or a CUDA error
What to do
Warning, GUI failed to start: Reason: no display name and no $DISPLAY environment variable
The launcher window needs a screen, and this session has none (a server over SSH, a container or Colab).
Likely causes, most likely first
- No desktop session
- A headless machine
What to do
***Warning, GUI failed to start: Reason: no display name and no $DISPLAY environment variableFile selection GUI unsupported.Exception in Tkinter callback
The launcher window hit an error while handling a click.
Likely causes, most likely first
- A launcher bug in this version
What to do
exo-open: symbol lookup error: … libharfbuzz.so.0: undefined symbol: FT_Get_Transform
KoboldCpp tried to open your browser and a library from the bundle clashed with the system's. It is harmless.
Likely causes, most likely first
- Auto-open of the browser on some Linux desktops
What to do
[PYI-…:ERROR] Failed to extract koboldcpp-launcher: decompression resulted in return code -1!
The single-file program could not unpack itself.
Likely causes, most likely first
- A full or dirty temp folder
- A damaged download
What to do
Error: KoboldCpp HTTP Server is running, but this endpoint does not exist. Please check the URL.
The server works, but the app asked for a URL or method it does not serve. A missing /v1 is usually not the cause.
Likely causes, most likely first
- The wrong method, for example GET on /v1/chat/completions
- A path under /api that does not exist
What to do
Error while submitting prompt: TypeError: Failed to fetch
The browser lost contact with the server while Lite was sending your prompt.
Likely causes, most likely first
- KoboldCpp stopped or crashed
- A tunnel link expired
- An https page calling an http server
What to do
Whisper Transcribe Generating...error: failed to open WAV file from stdin
The speech-to-text input was not a WAV file the program could read.
Likely causes, most likely first
- An unsupported audio format
What to do
Error: KCPP SD Failed to create context!
The image model could not start, most often because companion files are missing.
Likely causes, most likely first
- Flux or SD3 without its text encoders and VAE
- Not enough memory
What to do
Warning: OpenBLAS library file not found. Non-BLAS library will be used.
An informational note. It does not cause a crash.
Likely causes, most likely first
- A build that does not ship OpenBLAS
What to do
load_tensors: offloaded N/N layers to GPU
This line counts layers assigned to the GPU path. On a machine with no GPU it can still say all layers were offloaded.
Likely causes, most likely first
- The flag was accepted but no GPU backend exists
What to do
load_tensors: offloading 0 repeating layers to GPUload_tensors: offloaded 0/25 layers to GPUload_tensors: CPU model buffer size = 434.91 MiBerror: argument --contextsize/--ctx-size/-c: must be within [256, 524288]
The context size is outside the allowed range.
Likely causes, most likely first
- A typo in the number
What to do
[model_param] [port_param]koboldcpp-linux-x64-nocuda: error: argument --contextsize/--ctx-size/-c: must be within [256, 524288]OSError: libcuda.so.1: cannot open shared object file: No such file or directory
The NVIDIA driver library is not installed, so the CUDA build cannot start with CUDA.
Likely causes, most likely first
- No NVIDIA driver on this Linux machine
- A container without the NVIDIA runtime
- Using --usecuda on a machine without an NVIDIA GPU
What to do
- Install the NVIDIA driver, or in Docker use the NVIDIA container runtime
- Without an NVIDIA GPU use the -nocuda file or drop --usecuda
File "ctypes/__init__.py", line 373, in __init__OSError: libcuda.so.1: cannot open shared object file: No such file or directoryFileNotFoundError: Could not find module '…\koboldcpp_common.dll' (or one of its dependencies)
KoboldCpp's library needs a system library that is missing. With --usevulkan this is the Vulkan runtime that the graphics driver installs.
Likely causes, most likely first
- No graphics driver, or a broken one
- --usevulkan on a machine without Vulkan
What to do
File "ctypes\__init__.py", line 373, in __init__FileNotFoundError: Could not find module 'C:\Users\<user>\AppData\Local\Temp\_MEI28282\koboldcpp_common.dll' (or one of its dependencies). Try using the full path with constructor syntax.EOS token triggered!
The model wrote its end-of-text token, so the reply ended normally. It is not an error.
Likely causes, most likely first
- A normal end of a reply
- If replies end too early, a chat template that does not match the model
What to do
Note that using SWA Mode cannot be used with Context Shifting
The model uses sliding-window attention, and ContextShift is not used with it. Long chats then reprocess more of the prompt.
Likely causes, most likely first
- A sliding-window model such as the Gemma family
What to do
device vulkan0 does not support async, host buffers or events
An informational line: the Vulkan device lacks some optional transfer features, so KoboldCpp does without them.
Likely causes, most likely first
- The Vulkan driver's feature set
What to do
Could not find suitable download software, or all download methods failed.
KoboldCpp tried to fetch a model from a web address and every method it knows failed.
Likely causes, most likely first
- No network or a blocked connection
- A link that needs a login
- A missing downloader tool on the system
What to do
Generate Image: The response could not be sent, maybe connection was terminated?
The app that asked for the image closed the connection before the image was ready.
Likely causes, most likely first
- A client time-out shorter than the generation time
- The app was closed
What to do
IPv6 Socket Failed to Bind. IPv6 will be unavailable.
KoboldCpp could not listen on IPv6, so it uses IPv4 only. Local use is not affected.
Likely causes, most likely first
- IPv6 is turned off or unavailable on this machine
What to do
Why the last line is often not the real error
KoboldCpp ends a failed start with general messages. They are true, but they do not say why. Our own log of a truncated model file shows the pattern: the cause comes first, and every line after it is a wrapper.
llama_model_load: error loading model: tensor 'output.weight' data is not within the file bounds, model is corrupted or incomplete1llama_model_load_from_file_impl: failed to load model2gpttype_load_model: error: failed to load model '<path>'3Load Text Model OK: FalseCould not load text model: <path>4- llama_model_load: error loading model: tensor 'output.weight' data is : This is the cause: the file is cut short. Read upward from the last error until you reach the first one.
- llama_model_load_from_file_impl: failed to load model: A generic wrapper from the loader.
- gpttype_load_model: error: failed to load model '<path>': KoboldCpp reports the loader failed.
- Could not load text model: <path>: The final wrapper. It only says the load failed.
The wrappers to recognise are Failed to execute script, llama_model_load_from_file_impl: failed to load model, Could not load text model, Aborted (core dumped) and, in the Lite browser app, Failed to fetch. Rows marked “Wrapper message” in the list above point back here. The log analyzer does this reading for you and shows the earliest real cause.
Where to find the KoboldCpp error log
The console window is the log. In our test KoboldCpp wrote no log file by default F58, so if the window closes, the text is gone. Start KoboldCpp from a terminal so it stays open (how, in Will not start). Then select the text and copy it with your terminal’s copy command, or choose the whole window with select all.
To save the output to a text file, use your shell’s redirect to a file when you start KoboldCpp, and keep the window open to read it too. Set Extra diagnostic output: -1 quiet (used for Horde), 0 normal, 1 verbose, higher for more detail.Default 0Launcher Hardware tab: Debug ModeSince v1.15All details when you are asked for more detail. Before you post a log anywhere, redact it: the log analyzer does that and gives you a copy that is safe to share.
Exit codes we observed
When KoboldCpp stops at start-up, the exit code separates a bad argument from a bad file. These came from our own runs, identical on Windows, Linux and macOS where we tested both.
| 1 | Start-up crash: second server on a busy port | OSError: Cannot start the server: no socket could bind. | F49 |
| 1 | CUDA build with --usecuda and no NVIDIA driver (Linux, Windows) | libcuda.so.1: cannot open shared object file / Could not find module '…koboldcpp_cublas.dll' | F56 |
| 1 | --usevulkan on Windows without a graphics driver | Could not find module '…koboldcpp_common.dll' (or one of its dependencies) | F57 |
| 2 | Model file missing | Cannot find text model file: <path> | F19 |
| 2 | Context size out of range | argument --contextsize/--ctx-size/-c: must be within [256, 524288] | F22 |
| 2 | Launcher with no display | GUI failed to start: Reason: no display name and no $DISPLAY environment variable | F51 |
| 3 | File is not a GGUF | invalid or unsupported model file '<path>' | F20 |
| 3 | Incomplete download | tensor 'output.weight' data is not within the file bounds, model is corrupted or incomplete | F21 |
| 3 | Projector does not match the model | mismatch between text model (n_embd = 896) and mmproj (n_embd = 576) | F50 |
| 3 | KV cache cannot be allocated in RAM | failed to allocate buffer for kv cache | F55 |
Messages you can ignore
These lines look alarming and are not the cause of a problem. Rows are in the list above under the “harmless” filter.
Warning: OpenBLAS library file not found. Non-BLAS library will be used.is informational on older builds.ggml_cuda_host_malloc: failed to allocate … pinned memoryis usually a harmless fallback unless you use Locks the model in RAM so the OS cannot swap it out.Default falseLauncher Hardware tab: Use mlockSince v1.18All details.IPv6 Socket Failed to Bind. IPv6 will be unavailable.means local use through IPv4 is unaffected.device vulkan0 does not support async, host buffers or eventsis informational.Note that using SWA Mode cannot be used with Context Shiftingis expected for sliding-window models.EOS token triggered!is a normal end of a reply.exo-open: symbol lookup error … libharfbuzzis a clash when KoboldCpp tries to open your browser; open the address yourself.Unable to detect VRAMon a machine that really has no GPU is correct F18.
“Internal server error” and API errors in apps
Apps such as SillyTavern or a browser page show their own short message, and the real one is in KoboldCpp’s console. Check the console first.
- A 404 with
KoboldCpp HTTP Server is running, but this endpoint does not existmeans the wrong method or a path under/api/that does not exist. KoboldCpp accepted/v1/v1/models,/modelsand a missing/v1in our probes, so a missing/v1is rarely the cause F52. - A 401 means the server has a password and the request has no key; the web page and model name stay reachable without it F25.
- If a reply stops early,
finish_reason: lengthmeans the token limit was reached; in our test a prompt larger than the server’s context was cut from the start without any error F53.
The API tester shows which URL and method work on your server, and the OpenAI API page lists the endpoints.
Your error isn’t listed?
Paste the whole log into the analyzer. Search the GitHub issues with the stable part of the message only: drop paths, numbers and hex addresses. When you report it, follow what to include. If a message should be on this page, send the redacted line to hello@koboldcpp.dev through the contact page and we will add it.
Frequently asked questions
What does "Failed to execute script 'koboldcpp'" mean?
It is a wrapper from the launcher: something stopped KoboldCpp with an unhandled error, and this line does not say what. The reason is printed above it. In our tests it appeared after a busy port, a missing CUDA library and other start-up failures. Read upward to the first error line, or paste the whole log into the analyzer.
Where does KoboldCpp save its log?
It does not save one by default. In our Linux test no log file appeared in the working folder, the home folder or the temp folder. The console window is the log, so copy its text or start KoboldCpp from a terminal so the window stays open.
What does "EOS token triggered" mean?
The model wrote its end-of-text token, so the reply ended normally. Users report it as a harmless line. If replies end too early, check that the chat template matches the model.
Is "Unable to detect VRAM" a problem?
Only if you expected to use a GPU. KoboldCpp could not read your video memory, so automatic layer choice is off and it usually continues on the CPU. On a machine with no GPU, our logs show this line on a normal start.
Why does my app say "internal server error" when KoboldCpp is running?
The app shows a generic message and the real one is in KoboldCpp's console. Check the console window first. A 404 from KoboldCpp usually means the wrong method or a path that does not exist, not a missing /v1.