How could I go about directly adjusting arbitrary weights of a large language model like Llama or Gemma programmatically—say, adding something like +0.00001 to any given weight? With Stable Diffusion, I could use an interface like Automatic1111's and hook into the merge checkpoints tab Python code to alter weights. But with LLMs, where architectures vary and merging isn’t as clearly defined, what’s the best way to handle this? Bonus if there’s a way to modify quantized model weights while respecting the quantization level.