Flash-LLM: Enabling Cost-Effective and Highly-Efficient Large Generative Model Inference with Unstructured Sparsity

Key idea:

Load-as-Sparse and Compute-as-Dense