1 article
FP8 KV cache halves bytes per cached token. Exact vLLM, SGLang and TensorRT-LLM flags, scale handling, a worked Qwen3-32B example, and an eval gate.
