KVSharer: Efficient Inference via Layer-Wise Dissimilar KV Cache Sharing

Recommended software environment

• python == 3.10
• torch == 2.1.2
• transformers >= 4.38.0
• scikit-learn >= 1.0
• tqdm >= 4.49.0
• numpy >= 1.20.2

Description

• The process of KVSharer is described in two runnable .ipynb files test_llama.ipynb and test_internlm.ipynb, detailing how to conduct strategy search and how to integrate KVSharer for inference.
• The main implementation of sharing during inference can be found in the llama_real_share/cache_utils.py, internlm2_real_share/cache_utils.py where we introduce a new class called DynamicDictCache to store only a portion of the layers' KV cache.
• We add the kv_cache_share_layers_map parameter in the LlamaForCausalLM and InternLM2ForCausalLM to set the sharing strategy. The implementation can be found in llama_real_share/modeling_llama_kvsharer.py, internlm2_real_share/modeling_internlm2_kvsharer.py.
• We provide a wiki_demo.txt file in ./data folder for test.