Some models in Transformers have merged qkv_proj Linear modules, like Phi-3, so our current scheme of adding output activation observers to separate k_proj and v_proj Linear modules will not work.

We should be able to use the Cache and QuantizedCache classes in the HF Transformers library, which has been added to most modeling definitions as the class to manage past_key_values: https://github.com/huggingface/transformers/blob/8820fe8b8c4b9da94cf1e4761876f85c562e0efe/src/transformers/cache_utils.py#L770

This allows us to implement our own QuantizedCache with a quantize and dequantize function, where we can calculate the statistics needed for our kv_cache_quant scheme.

Prototype QuantizedCacheConfig impl #87