Split-KV introduces an exact context parallelism (CP) technique tailored to the training regime of diffusion drafters. Diffusion drafters speed up speculative decoding by denoising multiple draft tokens in a single pass; during training they employ anchor sampling to build compact query sequences and apply a block‑sparse mask for cross‑attention over a much longer context. Empirically, with a fixed query budget the acceptance length stays stable as context grows, while increasing the query budget yields only marginal quality gains, confirming that practical queries are far shorter than the context. Existing CP methods cannot simultaneously exploit this query‑context asymmetry and satisfy the nonlinear, sample‑dependent load‑balancing imposed by anchor sampling. Split‑KV keeps key/value shards local and restricts per‑layer attention communication to tensors sized like the queries, making communication independent of context length. The global query view also enables a cost‑aware longest‑processing‑time‑first (LPT) scheduler that balances context key/value chunks according to their realized attention cost. At a parallel scale of 64 GPUs, Split‑KV retains 85%–91% weak‑scaling efficiency and achieves up to 3.67× speedup over the fastest CP baseline, with LPT further improving resource utilization.\n\nReview