Spec Info:
LM Studio Version: 0.4.4 (Build 1)
Runtime: Vulkan llama.cpp (Windows) 2.4.0
GPU: AMD Radeon RX 6700 XT (12 GB VRAM)
System RAM: 64 GB
OS: Windows
Model: mistralai/ministral-3-14b-reasoning
Parameters: Default parameters (also tested reduced offload and shorter context length; no effect)
What is the bug?
After updating LM Studio to 0.4.4 (Build 1), the model mistralai/ministral-3-14b-reasoning no longer works with the Vulkan llama.cpp runtime. The model previously worked correctly. The issue appears specific to the Vulkan runtime, since the CPU llama.cpp runtime works normally with the same model.
Behaviour:
The model begins generating a reasoning response but outputs only repeated special tokens:
<SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30>...
Generation never progresses beyond these tokens.
When manually stopping generation, the model/runtime crashes with the following error:
The model has crashed without additional information. (Exit code: 18446744072635812000)
Suspected:
Regression or compatibility issue in Vulkan llama.cpp runtime 2.4.0, possibly affecting AMD GPUs (RDNA2 / RX 6000 series), causing incorrect token generation and runtime crash.
Additional findings:
I performed a completely clean reinstall of LM Studio. I uninstalled the application and manually deleted all remaining folders, including .lmstudio, AppData\Local\lmstudio, AppData\Local\lm_updater, and AppData\Roaming\lmstudio. After reinstalling LM Studio 0.4.4 and downloading the model again, the issue occurred immediately with no change.
When switching the runtime to CPU llama.cpp, the model works normally with the same model file and parameters. This indicates the problem is specific to the Vulkan runtime and not the model itself.
I previously had Vulkan runtime 2.3.0 installed, but it showed the error "Couldn't survey hardware." After removing it, older Vulkan runtime versions are no longer available for reinstall in LM Studio.
I also performed a full AMD GPU driver reinstall using the factory reset / clean install option. This did not resolve the issue.
I tested different parameter changes, including reducing GPU offload and lowering the context length (previously 8000, which worked fine before). These changes had no effect and the same SPECIAL_30 token loop and crash still occur.
Also tested with
zai-org/glm-4.6v-flash
Same issues only difference is that it respons with
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""...
instead of
<SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30>...
Spec Info:
LM Studio Version: 0.4.4 (Build 1)
Runtime: Vulkan llama.cpp (Windows) 2.4.0
GPU: AMD Radeon RX 6700 XT (12 GB VRAM)
System RAM: 64 GB
OS: Windows
Model: mistralai/ministral-3-14b-reasoning
Parameters: Default parameters (also tested reduced offload and shorter context length; no effect)
What is the bug?
After updating LM Studio to 0.4.4 (Build 1), the model mistralai/ministral-3-14b-reasoning no longer works with the Vulkan llama.cpp runtime. The model previously worked correctly. The issue appears specific to the Vulkan runtime, since the CPU llama.cpp runtime works normally with the same model.
Behaviour:
The model begins generating a reasoning response but outputs only repeated special tokens:
<SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30>...
Generation never progresses beyond these tokens.
When manually stopping generation, the model/runtime crashes with the following error:
The model has crashed without additional information. (Exit code: 18446744072635812000)
Suspected:
Regression or compatibility issue in Vulkan llama.cpp runtime 2.4.0, possibly affecting AMD GPUs (RDNA2 / RX 6000 series), causing incorrect token generation and runtime crash.
Additional findings:
I performed a completely clean reinstall of LM Studio. I uninstalled the application and manually deleted all remaining folders, including .lmstudio, AppData\Local\lmstudio, AppData\Local\lm_updater, and AppData\Roaming\lmstudio. After reinstalling LM Studio 0.4.4 and downloading the model again, the issue occurred immediately with no change.
When switching the runtime to CPU llama.cpp, the model works normally with the same model file and parameters. This indicates the problem is specific to the Vulkan runtime and not the model itself.
I previously had Vulkan runtime 2.3.0 installed, but it showed the error "Couldn't survey hardware." After removing it, older Vulkan runtime versions are no longer available for reinstall in LM Studio.
I also performed a full AMD GPU driver reinstall using the factory reset / clean install option. This did not resolve the issue.
I tested different parameter changes, including reducing GPU offload and lowering the context length (previously 8000, which worked fine before). These changes had no effect and the same SPECIAL_30 token loop and crash still occur.
Also tested with
zai-org/glm-4.6v-flash
Same issues only difference is that it respons with
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""...
instead of
<SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30><SPECIAL_30>...