Lossless Inference

How to make LLM serving faster without touching the model. Exact kernels, speculative decoding, lossless compression, KV reuse and scheduling, with the math and how to verify it with logit parity.