Describe the Issue
Using "banned_tokens" with any value causes the model to output garbage.
Tried with the release versions of both Gemma 4 and Qwen 3.5. Tried both setting it through the API and the command line. Tried CUDA, Vulkan, and CPU. With and without KV cache quantization. With and without Jinja. All fail.
Additional Information:
koboldcpp-1.117.1
AMD RX 7900 XTX 24GB Vulkan
NVIDIA 4060 Ti 16GB CUDA
Windows
Describe the Issue
Using "banned_tokens" with any value causes the model to output garbage.
Tried with the release versions of both Gemma 4 and Qwen 3.5. Tried both setting it through the API and the command line. Tried CUDA, Vulkan, and CPU. With and without KV cache quantization. With and without Jinja. All fail.
Additional Information:
koboldcpp-1.117.1
AMD RX 7900 XTX 24GB Vulkan
NVIDIA 4060 Ti 16GB CUDA
Windows