Prompt caching for RL training: 7.5x speedup on long-prompt, short-response workloads

prompt caching, but for rl training - 7.5x speedup on long-prompt/short-response workloads

The author proposes prompt caching for RL training. On Qwen3.5-4B, it reports a 7.5x speedup with 16k-token prompts and 64-token outputs, and the G=8 example with 1000-token prompts and 100-token responses reduces 8800 processed tokens to 1800 unique tokens.

Why it matters: HKR-H/K/R all pass: the angle is novel, and the post gives 16k/64 plus G=8 token-dedup numbers. Kept at 78 because this is a single Reddit post without independent replication or a paper/code artifact disclosed.