Because of a problem testing the Gemma 3 1B model using llama.cpp I've made a change, the model has an off-by-one mismatch between its config.json (which says vocab_size: 262144) and its actual tokenizer (which has 262145 tokens). This caused llama.cpp to reject the converted GGUF file with the error: check_tensor_dims: tensor 'token_embd.weight' has wrong shape; expected 1152, 262145, got 1152, 262144
I've switched the model Gemma 3 to DeepSeek-R1-Distill-Qwen-1.5B, I really hope the judges understand the situation, I was using other tools to run the model, but today I realized of this error coming from Gemma. But the training script and dataset are the same I just swicthed the model.
Log in or sign up for Devpost to join the conversation.