Last released Nov 4, 2026
Composable efficiency techniques (speculative decoding, early exit, quantization, pruning, KV-cache compression, ...) for Hugging Face causal LMs.