vk::Device::getFenceStatus: ErrorDeviceLost

[Najmulk424]

Environment

OS: Windows (fill in exact build) — e.g. Windows 10/11 (build )

LM Studio version:

Model: gemma-3-4b-it-Q4_K_M.gguf (path from log: C:\Users\nazmu.lmstudio\models\lmstudio-community\gemma-3-4b-it-GGUF...)

GPU: Intel(R) Arc(TM) A750 Graphics (driver/version: ). Vulkan backend in use.

VRAM reported free at load: ~7567 MiB (see logs)

Backend: llama.cpp / Vulkan / LM Studio LLM engine

Steps to reproduce

Start LM Studio server (default HTTP endpoints appear).

Load the Gemma 3 4B model (GGUF Q4_K).

Trigger a generation (POST to /v1/chat/completions or use UI) that causes actual token prediction (n_predict > 0).

Observe error during prediction.

What I expected
Model to generate tokens normally without the Vulkan device being lost.

What actually happened
Prediction failed with a Vulkan device error and process logs show vk::Device::getFenceStatus: ErrorDeviceLost. Generation aborted.

Relevant log snippet

[LLM Engine bindings] PredictWorker::Execute - caught exception: vk::Device::getFenceStatus: ErrorDeviceLost
[lmstudio-llama-cpp] Error in predictTokens: vk::Device::getFenceStatus: ErrorDeviceLost
1 Error predicting: _0x340f8e [Error]: vk::Device::getFenceStatus: ErrorDeviceLost
at _0x40d3bf.LLMEngineWrapper.predictTokens (...\LM Studio\resources\app.webpack\lib\llmworker.js:85:29855)
at async _0x339349.predictTokens (...\llmworker.js:121:17859)
at async _0x339349.handleMessage (...\llmworker.js:121:7111) {
title: 'vk::Device::getFenceStatus: ErrorDeviceLost'
}

Full relevant log context (attach if possible)

Server started and endpoints listening on http://localhost:1234

Model load summary: Gemma 3 4b It, GGUF V3 Q4_K, model params ~3.88B

Offloaded 35/35 layers to GPU; Vulkan0 (Intel Arc A750) used

Memory breakdown shows Vulkan0 ~8080 MiB etc.

Final failure occurs shortly after "BeginProcessingPrompt" during predictTokens

Files / attachments
Please see and attach server logs: C:\Users\nazmu.lmstudio\server-logs (all files there).
Attach the full LM Studio output / stack trace (the llmworker.js dump) and model file name.

Additional helpful info (please include when reporting)

Exact LM Studio version (Help → About or executable properties)

GPU driver version and Vulkan runtime version (e.g., Intel driver version, Vulkan SDK/runtime)

Frequency: does this happen every run or intermittently? (I observed it during this run.)

Does reducing Num Offload Layers or switching off Vulkan offload avoid the issue?

Any GPU stress tests / other Vulkan apps succeed on this machine?

Possible guesses / hints

Could be Intel driver or Vulkan runtime issue causing device lost under compute load.

Could relate to timing/timeouts in fence polling when offloading many layers.

Might reproduce with large context / when offloading many layers to GPU.

[karmicnull]

I'm seeing something similar since moving to 0.3.35

System:
CPU: Intel Core i7-14700K
GPU: Intel Arc A770 16GB
RAM: 64GB
Vulkan API : 1.4.328
OS: Windows 11 build 26100.7462
LM Studio version: 0.3.35
Model: Error occurs with all models

Steps to fix: revert back to previous version of LM Studio (I've gone back to 0.3.20 as I've got the installer for it)

Steps to reproduce: reinstall 0.3.35, load a model and attempt to chat with it.

Log:
[2025-12-14 14:54:37.751] [info] App starting...
[2025-12-14 14:54:37.753] [info] [AppUpdater] Update channel set to 'stable'
[2025-12-14 14:54:38.008] [info] Hardware survey for general system resources through 'vulkan' took 243.59ms
[2025-12-14 14:54:38.235] [info] [AppUpdater] Checking for updates... (current state: idle)
[2025-12-14 14:54:38.235] [info] AppUpdater state changed to checking-for-updates-periodic
[2025-12-14 14:54:38.236] [info] [AppUpdater] Fetching version info from https://versions-prod.lmstudio.ai/update/win32/x86/0.3.35
[2025-12-14 14:54:38.559] [info] [AppUpdater] Received version info response
[2025-12-14 14:54:38.559] [info] [AppUpdater] Current version: 0.3.35 (build: 1), new version: 0.3.35 (build: 1)
[2025-12-14 14:54:38.559] [info] No update available: 0.3.35 (build: 1) <= 0.3.35 (build: 1)
[2025-12-14 14:54:38.560] [info] AppUpdater state changed to idle
[2025-12-14 14:54:43.272] [error] [IndexedModelRecommendationProvider] No engine is selected to load 'gguf' model 'lmstudio-community/gemma-3-12b-it-GGUF/gemma-3-12b-it-Q4_K_M.gguf'. Cannot estimate model fit, returning that the model will not fit.
[2025-12-14 14:54:43.273] [error] [IndexedModelRecommendationProvider] No engine is selected to load 'gguf' model 'lmstudio-community/Mistral-Nemo-Instruct-2407-GGUF/Mistral-Nemo-Instruct-2407-Q6_K.gguf'. Cannot estimate model fit, returning that the model will not fit.
[2025-12-14 14:54:43.273] [error] [IndexedModelRecommendationProvider] No engine is selected to load 'gguf' model 'lmstudio-community/gpt-oss-20b-GGUF/gpt-oss-20b-MXFP4.gguf'. Cannot estimate model fit, returning that the model will not fit.
[2025-12-14 14:54:43.439] [error] Tried to get CUDA environment variable overrides for LLM usage, but GPU survey data showed no GPUs!
[2025-12-14 14:54:59.144] [error] (node:27892) [DEP0147] DeprecationWarning: In future versions of Node.js, fs.rmdir(path, { recursive: true }) will be removed. Use fs.rm(path, { recursive: true }) instead
(Use LM Studio --trace-deprecation ... to show where the warning was created)
[2025-12-14 14:55:02.707] [error] [LMSInternal][Client=plugin:builtin:lmstudio/default-prediction-loop-handler][Endpoint=predict] Error in channel handler: Error: received prediction-error
at _0x3765df. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:1196:145385)
at _0x210b9a._0x55a50d (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:119:13335)
at _0x210b9a.emit (node:events:519:28)
at _0x210b9a.onChildMessage (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:107:214342)
at _0x210b9a.onChildMessage (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:124:4046)
at _0x3a24ae. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:107:213331)
at _0x3a24ae.emit (node:events:519:28)
at ForkUtilityProcess. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:667:17569)
at ForkUtilityProcess.emit (node:events:519:28)
at ForkUtilityProcess.a.emit (node:electron/js2c/browser_init:2:75042)

• Caused By: Error: vk::Device::getFenceStatus: ErrorDeviceLost
  at _0x77d4c1..predictTokens (C:\Program Files\LM Studio\resources\app.webpack\lib\llmworker.js:85:30145)
  at async _0x3fe65d.predictTokens (C:\Program Files\LM Studio\resources\app.webpack\lib\llmworker.js:124:17803)
  at async _0x3fe65d.handleMessage (C:\Program Files\LM Studio\resources\app.webpack\lib\llmworker.js:124:7013)
  [2025-12-14 14:55:02.709] [error] [LMSInternal][Client=LM Studio][Endpoint=sendMessage] Error in RPC handler: Error: Rehydrated error
  tFenceStatus: ErrorDeviceLost
  at _0x3765df. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:1196:145385)
  at _0x576842.subscriber (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:375:1530)
  at _0x576842.notifier (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:1024:205738)
• Caused By: Error: Channel Error
  at handle (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:480:4507)
• Caused By: Error: received prediction-error
  at _0x3765df. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:1196:145385)
  at _0x210b9a._0x55a50d (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:119:13335)
  at _0x210b9a.emit (node:events:519:28)
  at _0x210b9a.onChildMessage (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:107:214342)
  at _0x210b9a.onChildMessage (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:124:4046)
  at _0x3a24ae. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:107:213331)
  at _0x3a24ae.emit (node:events:519:28)
  at ForkUtilityProcess. (C:\Program Files\LM Studio\resources\app.webpack\main\index.js:667:17569)
  at ForkUtilityProcess.emit (node:events:519:28)
  at ForkUtilityProcess.a.emit (node:electron/js2c/browser_init:2:75042)
• Caused By: Error: vk::Device::getFenceStatus: ErrorDeviceLost
  at _0x77d4c1..predictTokens (C:\Program Files\LM Studio\resources\app.webpack\lib\llmworker.js:85:30145)
  at async _0x3fe65d.predictTokens (C:\Program Files\LM Studio\resources\app.webpack\lib\llmworker.js:124:17803)
  at async _0x3fe65d.handleMessage (C:\Program Files\LM Studio\resources\app.webpack\lib\llmworker.js:124:7013)

[orionmachine]

I too am having the same issues. Same as above (Error: vk::Device::getFenceStatus: ErrorDeviceLost) I have noticed that the only version of Vulkan that is working with the A770 16GB is v1.57.1, so there's something about that version that must be missing or changed going forward that makes it incompatible with A770.

System:
CPU: Intel Core i5-12400F
GPU: Intel Arc A770 16GB
RAM: 32GB
Vulkan API : 1.4.328
OS: Windows 10 build 19045.6466
LM Studio version: 0.3.35
Model: Error occurs with all models

[emerzon]

Are there any safe links for previous versions? I have the same and confirm it was working up to v1.57.1

[AmineKehh]

I need also a download link of old versions.

[savvadesogle]

That is strange. GPU 0 and 1 are used during PP, although I select only 1 of the two GPUs in the settings.

its for 1.58 llama.cpp Vulkan Windows

The latest 1.66 is not working

[f-lombardo]

Same error here with version 0.3.36 on Ubuntu 24.04.3 on LENOVO_MT_21L3_BU_Think_FM_ThinkPad L16 Gen 1

[amanturisaev]

Same error with version 0.3.36 on Windows 11 Pro on Intel Arc B580 with 12GB of VRAM factory overclocked. Decreasing the GPU offload to 0/32 works on CPU Intel Core i5-14600KF, but it is slow. The error appears when trying to use the loaded model meta-llama-3.1-8b-instruct. In my case it worked without errors up until I cut off the Internet connectivity using basic firewall rules. I have no iGPU.