Why does KoboldCpp close immediately?

The window closes when the program stops, and the error was printed inside it. Start KoboldCpp from a Command Prompt or terminal so the text stays, then read the last lines. In our tests a missing model file, a file that is not a GGUF and a busy port each stopped it within seconds.

Where is the KoboldCpp log?

In the console window. In our Linux test no log file appeared in the working folder, the home folder or the temp folder, so copy the text from the window or start KoboldCpp from a terminal so the window stays open.

Why does Kobold Lite say "Failed to fetch"?

The browser could not reach the server. KoboldCpp stopped, crashed, or a tunnel timed out. Check the console window first: if KoboldCpp is still running, check the address and port.

How do I know if KoboldCpp is using my GPU?

Find the line that says KV buffer size. If it starts with CUDA0, Vulkan0 or MTL0 the context memory is on the GPU. If it starts with CPU it is not, even when another line says all layers were offloaded.

KoboldCpp worked yesterday and not today: what changed?

The usual suspects are a KoboldCpp update, a graphics driver update, a different model file, or a security program blocking the file. Change one thing at a time to find which.

KoboldCpp Not Working? Find the Fix by Symptom

KoboldCpp not working? Find your symptom

What is going wrong?

What is going wrong?

It won't open, or the window closes straight away: Can you read the last lines the window printed?
- No, it closes too fast: Start it from a Command Prompt so the window stays open Open the KoboldCpp folder, type cmd in the address bar and press Enter. Type the file name, for example koboldcpp.exe, and press Enter. Read the last lines when it stops. How to see the error (Users report: F58)
- Yes: Which of these is in the last lines?
  - Illegal instruction, or an access violation on an older PC: Use the old-PC build, or Failsafe mode Download koboldcpp-oldpc (or the nocuda oldpc build), or turn on Failsafe / --noavx2 in the launcher. Old CPU crashes (Tested by us: F03, F15)
  - PyInstallerImportError, DLL load failed, WinError 1114, or Could not find module … koboldcpp_cublas.dll: A DLL could not load: check the file, driver and folder Make sure the download finished and you have the right file for your hardware. Update the graphics driver, then move KoboldCpp to a short folder such as C:\kobold. If it still fails, paste the log into the analyzer. DLL errors (Tested by us: F56, F57)
  - WinError 4551, or Windows says an app was blocked: Windows blocked the program Confirm the file came from the official GitHub release, then check Smart App Control or SmartScreen before changing anything. Do not turn off antivirus as a first step. Smart App Control (Users report: users report: issues #1946, #2590)
  - Only Failed to execute script 'koboldcpp': That line is a wrapper: read the line above it Scroll up to the first error line. Or paste the whole log into the analyzer. Failed to execute script (Tested by us: F49, F56)
  - PrefetchVirtualMemory not found in KERNEL32.dll: Windows 7 is not supported by current builds Use a newer Windows, or an older KoboldCpp release that still supports it. Windows 7 (Users report: users report)
  - Failed to extract … decompression resulted in return code -1: The self-extracting file can't unpack: check the temp folder Free disk space on the drive that holds your Temp folder. Run it again from a short folder path. Temp folder (Users report: users report)
  - no $DISPLAY, or Exception in Tkinter callback (the launcher never appears): The launcher window cannot open here On a server or SSH session, skip the launcher: run with --model and --skiplauncher. On Linux desktops, check that a display is set. Launcher won't open (Users report: users report)
  - A message about the model file: What does the log say about the model?
    - Cannot find text model file: The path is wrong Check the path and spelling, and use the full path or put the file next to KoboldCpp. Quote paths that contain spaces. Cannot find model file (Tested by us: F19)
    - invalid or unsupported model file: That file is not a GGUF Download the GGUF version of the model (the file name ends in .gguf). Do not load safetensors or .bin files. Not a GGUF (Tested by us: F20)
    - data is not within the file bounds, model is corrupted or incomplete: The file is incomplete: download it again Compare the file size with the size on the model page, then download it again. Incomplete download (Tested by us: F21)
    - unknown model architecture: KoboldCpp is older than the model Update KoboldCpp to the current release and try again. Unknown architecture (Users report: users report: 17 issues)
    - missing tensor, or has wrong shape: This quantised file does not match the architecture Download a different quantisation or another upload of the same model. Missing tensor (Users report: users report)
    - key clip.vision … not found: You loaded a vision projector as the model Load the main text model as the model and give the projector file under the vision (mmproj) setting. Wrong mmproj (Tested by us: F50)
    - must be within [256, 524288]: The context size is out of range Use a value between 256 and 524288 for --contextsize. Context range (Tested by us: F22)
    - My model is several files ending -00001-of-0000N: Load the first part only Keep all the parts in one folder and select the file ending -00001-of-0000N. Split models (Users report: users report)
    - Only failed to load model or Could not load text model: That is a wrapper: scroll up or use the analyzer Read the first error above it, or paste the log into the analyzer. Could not load text model (Tested by us: F21)
    - Out of memory: The model is bigger than your memory Use a smaller quantisation, offload fewer layers, or lower the context. While loading (Tested by us: F54, F55)
  - Out of memory: When does it run out of memory?
    - While loading the model: The model is bigger than your memory
    - After a while, when the chat gets long: The context cache grows with the chat Lower the context size or quantise the KV cache. Context fills (Tested by us: F55)
    - Only when an image, voice or vision model is also loaded: Extra models share the same memory Unload the image, voice or vision model, or put it on the CPU. Extra models (Users report: users report)
    - It warns about pinned memory but keeps running: Usually harmless If it keeps running, ignore the pinned memory warning. Pinned memory (Users report: users report)
    - It's my system RAM, not the GPU: System RAM ran out Close other programs, use a smaller file or quantisation, or add RAM. System RAM (Tested by us: F55)
  - Something else: Paste the log into the analyzer Copy the console text, paste it into the analyzer on this page, and read the verdict. It finds the deepest real error, not the last wrapper line. Paste your log here (Tested by us: F49, F56)
It opens, but the model doesn't load: What does the log say about the model?
- Cannot find text model file: The path is wrong
- invalid or unsupported model file: That file is not a GGUF
- data is not within the file bounds, model is corrupted or incomplete: The file is incomplete: download it again
- unknown model architecture: KoboldCpp is older than the model
- missing tensor, or has wrong shape: This quantised file does not match the architecture
- key clip.vision … not found: You loaded a vision projector as the model
- must be within [256, 524288]: The context size is out of range
- My model is several files ending -00001-of-0000N: Load the first part only
- Only failed to load model or Could not load text model: That is a wrapper: scroll up or use the analyzer
- Out of memory: The model is bigger than your memory
It runs, but my GPU isn't used: What does the log show near the top?
- Auto Selected Default Backend (flag=0), or the launcher says Use CPU: Which file did you download, and which GPU do you have?
  - koboldcpp-nocuda with an NVIDIA GPU: Use Vulkan, or get the CUDA file Choose Vulkan in the launcher, or download koboldcpp.exe for CUDA. Wrong file (Tested by us: F02)
  - koboldcpp.exe with an AMD or Intel GPU: Choose Vulkan Choose Vulkan in the launcher. The smaller koboldcpp-nocuda.exe is enough. Pick a backend (Tested by us: F02)
  - koboldcpp.exe with NVIDIA, but CUDA isn't offered: Update the NVIDIA driver Install the current NVIDIA driver and restart. Old cards may need the oldpc build or Vulkan. Driver (Tested by us: F56)
  - A Mac: Metal is used automatically Look for MTL0 lines in the log. If they are present, the GPU is in use. Mac GPU (Tested by us: F29)
  - I'm not sure: Open the download picker Answer three questions and get the right file. Find the right file (Tested by us: F02)
- Unable to detect VRAM: Choose the backend and the layers yourself Pick CUDA, Vulkan or Metal in the launcher, set GPU layers by hand, and use the memory calculator to find the number. Unable to detect VRAM (Tested by us: F18)
- CUDA or Vulkan is selected, but it's as slow as CPU: Find the line with KV buffer size. Does it start with CUDA0, Vulkan0 or MTL0, or with CPU?
  - CPU: Nothing is behind the GPU flag Check that a GPU device was found, and that --lowvram is off. Choose a backend yourself. Offloaded line (Tested by us: F17)
  - A GPU name, but offloaded shows only some layers (for example 7/41): Autofit was cautious Lower the context size, or set GPU layers by hand using the memory calculator. Autofit (Tested by us: F18)
  - A GPU name, and all layers: Tune the speed Work through the speed guide: batch size, flash attention and context. Speed (Tested by us: F07)
Out of memory: When does it run out of memory?
- While loading the model: The model is bigger than your memory
- After a while, when the chat gets long: The context cache grows with the chat
- Only when an image, voice or vision model is also loaded: Extra models share the same memory
- It warns about pinned memory but keeps running: Usually harmless
- It's my system RAM, not the GPU: System RAM ran out
It's slow: Which number is low? After each reply the console prints a processing speed and a generation speed.
- Writing (generation) is slow: Are all layers on the GPU? (the KV buffer line starts with your GPU name)
  - No: Fit more layers on the GPU Use the memory calculator to find the number of layers that fit, then set it. GPU layers (Tested by us: F17)
  - Yes: Tune the speed
- Reading the prompt (processing) is slow every turn: Your context is reprocessed each turn Keep context shift on, and avoid editing the start of the chat. Context (Users report: users report)
- Everything got slower after an update: Compare with the previous release Read the release notes, then try the previous release. Releases (Users report: users report)
- I have no GPU: Set the CPU threads Match threads to your physical cores. Speed (Tested by us: F07)
Replies are wrong, repeat or stop early: What's wrong with the replies?
- Gibberish or random symbols: Check the chat template Choose the template the model was trained with. Chat templates (Users report: users report)
- <think> text or endless thinking: Handle thinking tags Use the thinking setting for the model and the matching template. Chat templates (Users report: users report)
- They stop too early: Check the length settings Raise the response length. EOS token triggered is a normal end of reply. Output section (Users report: users report)
- They repeat: Lower the repetition samplers Check repetition penalty, DRY and XTC are not too high. Samplers (Users report: users report)
- It feels dumber after an update: Compare with the previous release Try the previous release with the same model and settings. Releases (Users report: users report)
An app or the browser can't connect: What do you see?
- Failed to fetch or NetworkError: The server stopped or is not reachable Look at the console: is KoboldCpp still running? Restart it and refresh the page. Connection problems (Users report: users report: 5 issues)
- this endpoint does not exist, or 404: Use the right URL for your API type Check the base URL for your app against the table in the connection section. Connection problems (Tested by us: F23)
- 401 Unauthorized: A password is set Send the key as a Bearer token, or start without --password. Connection problems (Tested by us: F25)
- SillyTavern can't connect: Check SillyTavern's connection settings Follow the SillyTavern guide. SillyTavern (Tested by us: F23)
- It works on the PC, but not from my phone: Open it to your network, with a password Follow the remote access guide. Remote access (Users report: users report)
- Port already in use: Another process uses the port Close the other KoboldCpp, or choose another port with --port. Port in use (Tested by us: F49)
A feature (images, voice, vision, web search) doesn't work: Feature problems have short fixes below Open the “A feature doesn't work” section: it has one line and a link for image generation, voice, vision, web search, ContextShift and ROCm. Features that don't work (Upstream says: KoboldCpp 1.124.2 docs)
I have an error message: Paste the log into the analyzer

The tree has no tracking: your answers stay in the page address (after the #), so you can bookmark or share a path.

Have an error message? Paste it here

Paste the error line, or the whole log

The result appears here. Nothing you paste leaves this page.

The analyzer runs in your browser. Nothing is uploaded, and it hides personal details such as your user name.

Fix guides

Five checks that solve most problems

Five checks 0 of 5 done

You can see the console output. Start KoboldCpp from a terminal so the window stays open.
You have the right file for your hardware. The CUDA file on a PC without NVIDIA runs, but silently on the CPU.
Your version is current. The first console line reads Welcome to KoboldCpp - Version followed by the version.
The model fits your memory. Check it with the memory calculator.
You changed one thing at a time, and noted what you changed.

See the output. How to keep the window open.
Right file. The download picker chooses it. In our tests the CUDA file on a machine without NVIDIA started and ran on the CPU, 2 to 4 times slower F02. The old-PC build on a modern CPU was about 40% slower F03.
Current version. Latest release we checked: 1.124.2 (see the releases page for newer ones). See how to update.
Fit. Use the memory calculator before you load a large model.

Read the KoboldCpp console in 60 seconds

The console tells you where the model really is. This excerpt is from a real start on a machine with no GPU. The notes are the lines that matter.

Real start, CPU onlyR-run1-lnx-nocuda-base-default-args · lab-lnx · KoboldCpp 1.124.2

Welcome to KoboldCpp - Version 1.124.2¹No GPU or CPU backend was selected. Trying to assign one for you automatically...²Unable to detect VRAM.³Auto Selected Default Backend (flag=0)⁴Initializing dynamic library: koboldcpp_default.so⁵…load_tensors: offloaded 0/25 layers to GPU⁶llama_context: n_ctx                 = 16640⁷llama_kv_cache:        CPU KV buffer size =   195.00 MiB⁸

Welcome to KoboldCpp - Version 1.124.2: Your version. Compare it with the latest release.
No GPU or CPU backend was selected. Trying to assign one for you autom: No backend was chosen by you.
Unable to detect VRAM.: KoboldCpp cannot size GPU layers automatically (F18).
Auto Selected Default Backend (flag=0): The default (CPU) engine was picked.
Initializing dynamic library: koboldcpp_default.so: Which engine library loaded. A name with noavx2 means the old-CPU path (F03).
load_tensors: offloaded 0/25 layers to GPU: Layers on the GPU. Do not trust this line alone: in F17 it said 25/25 while everything ran on the CPU.
llama_context: n_ctx = 16640: The context you asked for plus 256 (F06).
llama_kv_cache: CPU KV buffer size = 195.00 MiB: Where the context memory lives. This is the line to trust (F17).

The same lines on a machine with a Vulkan device look different. The device name replaces CPU in the buffer lines. This one comes from a software Vulkan device on our test machine (no physical graphics card yet), which prints the same line formats a real GPU does.

Vulkan device found, 12 of 25 layers offloadedR-lab2-3-u24-vk-gl12-alloc-q4-ctx4096 · u24-vk-gl12 · KoboldCpp 1.124.2

Initializing dynamic library: koboldcpp_vulkan.so¹ggml_vulkan: Found 1 Vulkan devices:²load_tensors: offloaded 12/25 layers to GPU³load_tensors:      Vulkan0 model buffer size =   106.39 MiB⁴llama_kv_cache:        CPU KV buffer size =    27.62 MiBllama_kv_cache:    Vulkan0 KV buffer size =    23.38 MiB⁵

Initializing dynamic library: koboldcpp_vulkan.so: The Vulkan engine loaded.
ggml_vulkan: Found 1 Vulkan devices:: A device was found. On a real PC this line names your graphics card.
load_tensors: offloaded 12/25 layers to GPU: Only some layers: we set 12 by hand.
load_tensors: Vulkan0 model buffer size = 106.39 MiB: Model weights on the device.
llama_kv_cache: Vulkan0 KV buffer size = 23.38 MiB: Context memory split between CPU and device.

If your buffer lines all say CPU, the work is on the CPU, whatever else the log says. See not using the GPU.

Connection problems

“Failed to fetch” or “NetworkError” in Kobold Lite

The browser could not reach KoboldCpp: the server stopped or crashed, or a tunnel timed out. Look at the console window first. If it is gone, start KoboldCpp again from a terminal. Users report this on 5 issues; the message is explained under Failed to fetch.

404: “this endpoint does not exist”

The exact text is KoboldCpp HTTP Server is running, but this endpoint does not exist. Please check the URL and METHOD. F23. It means the path or the method is wrong for that API type. One server answers the KoboldAI, OpenAI, Ollama and A1111 families, so check the base URL your app needs on the API page and try it in the API tester. See also the 404 row.

401 Unauthorized

A password is set with --password. Generation endpoints return 401 without the key and answer 200 with it sent as a Bearer token. The web page and the model name stay reachable without it F25.

Port 5001 already in use

A second KoboldCpp on the same port prints a normal start-up banner, then stops with OSError: Cannot start the server: no socket could bind. followed by the wrapper line Failed to execute script 'koboldcpp'; the first one keeps running F49. Close the other one, or choose another port with . See the row.

My phone or another PC can’t reach it

By default it is available on the PC only. Follow remote access.

SillyTavern won’t connect

Check the API type and the address. Use the SillyTavern guide and try the same address in the API tester.

Bad, repeating or cut-off replies

Work through the three causes in this order:

The chat template is wrong. Pick the one the model was made for: chat templates.
A sampler is too strong. Repetition penalty, DRY or XTC set too high can cause loops: samplers.
A backend or version problem. Try --noflashattention, try another backend such as CUDA or Vulkan, or try the previous release from the releases.

`<think>` text shows, or it never stops thinking

A reasoning model prints its thinking first. Use the matching template and thinking setting: chat templates.

Replies stop early

EOS token triggered! is a normal end-of-reply note (row). If replies are too short, raise the response length in Kobold Lite and check the samplers.

It feels dumber after an update

Try the previous release with the same model and settings, and compare. Releases are listed on the releases page.

A feature doesn’t work

Image generation: image generation, and the context error.
Voice (speech, Whisper): voice, and the WAV error.
Vision: vision, and the missing-key error.
Web search: embeddings and search.
ContextShift “not working”: expected with a quantised KV cache or sliding-window models, see context size and the row.
ROCm not working: AMD; the Windows ROCm build is a third-party fork.

How to change the model

Close KoboldCpp and start it again with another model file, using the launcher’s model field or the model argument. To switch without restarting, start it with . See the launcher and flags.

Still stuck? Ask for help the right way

Include:

the first lines of the console (the version line), your operating system and your GPU,
the name of the KoboldCpp file you downloaded and the model file name,
the whole log, redacted with the analyzer’s copy button,
what you already tried.

Search the existing GitHub issues first. Then ask on the KoboldAI Discord (koboldai.org/discord) or r/KoboldAI. koboldcpp.dev is independent and cannot fix KoboldCpp bugs: report bugs upstream. Our methods are on how we test.

KoboldCpp troubleshooting: start from what you see

KoboldCpp not working? Find your symptom

What is going wrong?

Have an error message? Paste it here

Fix guides

Five checks that solve most problems

Read the KoboldCpp console in 60 seconds

Connection problems

“Failed to fetch” or “NetworkError” in Kobold Lite

404: “this endpoint does not exist”

401 Unauthorized

Port 5001 already in use

My phone or another PC can’t reach it

SillyTavern won’t connect

Bad, repeating or cut-off replies

`<think>` text shows, or it never stops thinking

Replies stop early

It feels dumber after an update

A feature doesn’t work

How to change the model

Still stuck? Ask for help the right way

Frequently asked questions

KoboldCpp not working? Find your symptom

What is going wrong?

Have an error message? Paste it here

Fix guides

Five checks that solve most problems

Read the KoboldCpp console in 60 seconds

Connection problems

“Failed to fetch” or “NetworkError” in Kobold Lite

404: “this endpoint does not exist”

401 Unauthorized

Port 5001 already in use

My phone or another PC can’t reach it

SillyTavern won’t connect

Bad, repeating or cut-off replies

<think> text shows, or it never stops thinking

Replies stop early

It feels dumber after an update

A feature doesn’t work

How to change the model

Still stuck? Ask for help the right way

Frequently asked questions

Related

`<think>` text shows, or it never stops thinking